cuCIM API Reference#
Clara Submodules#
 class cucim.clara.CuImage#
 Attributes:
associated_images
Returns a set of associated image names.
channel_names
A channel name list.
coord_sys
Coordinate frame in which the direction cosines are measured.
device
A device type.
dims
A string containing a list of dimensions being requested.
direction
Direction cosines (size is always 3x3).
dtype
The data type of the image.
is_loaded
True if image data is loaded & available.
metadata
A metadata object as dict.
ndim
The number of dimensions.
origin
Physical location of (0, 0, 0) (size is always 3).
path
Underlying file path for this object.
raw_metadata
A raw metadata string.
resolutions
Returns a dict that includes resolution information.
shape
A tuple of dimension sizes (in the order of dims)
typestr
The data type of the image in string format.
Methods
associated_image
(self[, name, device])Returns an associated image for the given name, as a CuImage object.
cache
([type])Get cache object.
close
(self)Closes the file handle.
profiler
(**kwargs)Get profiler object.
read_region
(self[, location, size, level, ...])Returns a subresolution image.
save
(self, arg0)Saves image data to the file path.
size
(self[, dim_order])Returns size as a tuple for the given dimension order.
spacing
(self[, dim_order])Returns physical size in tuple.
spacing_units
(self[, dim_order])Units for each spacing element (size is same with ndim).
 associated_image(self: cucim.clara._cucim.CuImage, name: str = '', device: cucim.clara._cucim.io.Device = cpu) object #
Returns an associated image for the given name, as a CuImage object.
 property associated_images#
Returns a set of associated image names.
Digital Pathology image usually has a label/thumbnail or a macro image(lowpower snapshot of the entire glass slide). Names of those images (such as ‘macro’ and ‘label’) are in associated_images.
 static cache(type: object = None, **kwargs) cucim.clara._cucim.cache.ImageCache #
Get cache object.
 property channel_names#
A channel name list.
 close(self: cucim.clara._cucim.CuImage) None #
Closes the file handle.
Once the file handle is closed, the image object (if loaded before) still exists but cannot read additional images from the file.
 property coord_sys#
Coordinate frame in which the direction cosines are measured.
Available Coordinate frame names are not finalized yet.
 property device#
A device type.
By default t is cpu (It will be changed since v0.19.0).
 property dims#
A string containing a list of dimensions being requested.
 The default is to return the six standard dims (‘STCZYX’) unless it is a DP multiresolution image.
[sites, time, channel(or wavelength), z, y, x]. S  Sites or multiposition locations.
NOTE: in OMETIFF’s metadata, dimension order would be specified as ‘XYZCTS’ (first one is fastiterating dimension).
 property direction#
Direction cosines (size is always 3x3).
 property dtype#
The data type of the image.
 property is_loaded#
True if image data is loaded & available.
 is_trace_enabled = False#
 property metadata#
A metadata object as dict.
It would be a dictionary(keyvalue pair) in general but can be a complex object (e.g., OMETIFF metadata).
 property ndim#
The number of dimensions.
 property origin#
Physical location of (0, 0, 0) (size is always 3).
 property path#
Underlying file path for this object.
 static profiler(**kwargs) cucim.clara._cucim.profiler.Profiler #
Get profiler object.
 property raw_metadata#
A raw metadata string.
 read_region(self: cucim.clara._cucim.CuImage, location: Iterable = (), size: List[int] = (), level: int = 0, num_workers: int = 0, batch_size: int = 1, drop_last: bool = False, prefetch_factor: int = 2, shuffle: bool = False, seed: int = 0, device: cucim.clara._cucim.io.Device = cpu, buf: object = None, shm_name: str = '', **kwargs) object #
Returns a subresolution image.
location and size’s dimension order is reverse of image’s dimension order.
Need to specify (X,Y) and (Width, Height) instead of (Y,X) and (Height, Width).
If location is not specified, location would be (0, 0) if Z=0. Otherwise, location would be (0, 0, 0)
Like OpenSlide, location is level0 based coordinates (using the level0 reference frame)
If size is not specified, size would be (width, height) of the image at the specified level.
<not supported yet> Additional parameters (S,T,C,Z) are similar to <https://allencellmodeling.github.io/aicsimageio/aicsimageio.html#aicsimageio.aics_image.AICSImage.get_image_data>
Do not yet support indices/ranges for (S,T,C,Z).
Default value for level, S, T, Z are zero.
Default value for C is 1 (whole channels)
<not supported yet> device could be one of the following strings or Device object: e.g., ‘cpu’, ‘cuda’, ‘cuda:0’ (use index 0), cucim.clara.io.Device(cucim.clara.io.CUDA,0).
<not supported yet> If buf is specified (buf’s type can be either numpy object that implements __array_interface__, or cupycompatible object that implements __cuda_array_interface__), the read image would be saved into buf object without creating CPU/GPU memory.
<not supported yet> If shm_name is specified, shared memory would be created and data would be read in the shared memory.
 property resolutions#
Returns a dict that includes resolution information.
level_count: The number of levels
level_dimensions: A tuple of dimension tuples (width, height)
level_downsamples: A tuple of downsample factors
level_tile_sizes: A tuple of tile size tuple (tile width, tile_height)
 save(self: cucim.clara._cucim.CuImage, arg0: str) None #
Saves image data to the file path.
Currently it supports only .ppm file format that can be viewed by eog command in Ubuntu.
 property shape#
A tuple of dimension sizes (in the order of dims)
 size(self: cucim.clara._cucim.CuImage, dim_order: str = '') List[int] #
Returns size as a tuple for the given dimension order.
 spacing(self: cucim.clara._cucim.CuImage, dim_order: str = '') List[float] #
Returns physical size in tuple.
If dim_order is specified, it returns phisical size for the dimensions. If a dimension given by the dim_order doesn’t exist, it returns 1.0 by default for the missing dimension.
 Args:
dim_order: A dimension string (e.g., ‘XYZ’)
 Returns:
A tuple with physical size for each dimension
 spacing_units(self: cucim.clara._cucim.CuImage, dim_order: str = '') List[str] #
Units for each spacing element (size is same with ndim).
 property typestr#
The data type of the image in string format.
The value can be converted to NumPy’s dtype using numpy.dtype(). (e.g., numpy.dtype(img.typestr)).
 class cucim.clara.DLDataType#
 Attributes:
 property bits#
Number of bits, common choices are 8, 16, 32.
 property code#
Type code of base types.
 property lanes#
Number of lanes in the type, used for vector types.
 class cucim.clara.DLDataTypeCode#
Members:
DLInt
DLUInt
DLFloat
DLBfloat
 Attributes:
name
name(self: handle) > str
 value
 DLBfloat = <DLDataTypeCode.DLBfloat: 4>#
 DLFloat = <DLDataTypeCode.DLFloat: 2>#
 DLInt = <DLDataTypeCode.DLInt: 0>#
 DLUInt = <DLDataTypeCode.DLUInt: 1>#
 property name#
 property value#
cache#
 class cucim.clara.cache.CacheType#
Members:
NoCache
PerProcess
SharedMemory
 Attributes:
name
name(self: handle) > str
 value
 NoCache = <CacheType.NoCache: 0>#
 PerProcess = <CacheType.PerProcess: 1>#
 property name#
 property value#
 class cucim.clara.cache.ImageCache#
 Attributes:
capacity
A capacity of list/hashmap.
config
Returns the dictionary of configuration.
free_memory
A cache memory size available in the cache memory.
hit_count
A cache hit count.
memory_capacity
A capacity of cache memory.
memory_size
A size of cache memory used.
miss_count
A cache miss count.
size
A size of list/hashmap.
type
A Cache type.
Methods
record
(self[, value])Records the cache statistics.
reserve
(self, memory_capacity, **kwargs)Reserves more memory if possible.
 property capacity#
A capacity of list/hashmap.
 property config#
Returns the dictionary of configuration.
 property free_memory#
A cache memory size available in the cache memory.
 property hit_count#
A cache hit count.
 property memory_capacity#
A capacity of cache memory.
 property memory_size#
A size of cache memory used.
 property miss_count#
A cache miss count.
 record(self: cucim.clara._cucim.cache.ImageCache, value: object = None) bool #
Records the cache statistics.
 reserve(self: cucim.clara._cucim.cache.ImageCache, memory_capacity: int, **kwargs) None #
Reserves more memory if possible.
 property size#
A size of list/hashmap.
 property type#
A Cache type.
 cucim.clara.cache.preferred_memory_capacity(img: object = None, image_size: Optional[List[int]] = None, tile_size: Optional[List[int]] = None, patch_size: Optional[List[int]] = None, bytes_per_pixel: int = 3) int #
Returns a good cache memory capacity value in MiB for the given conditions.
Please see how the value is calculated: https://godbolt.org/z/8vxnPfKM5
 Args:
img: A CuImage object that can provide image_size, tile_size, bytes_per_pixel information. If this argument is provided, only patch_size from the arguments is used for the calculation. image_size: A list of values that presents the image size (width, height). tile_size: A list of values that presents the tile size (width, height). The default value is (256, 256). patch_size: A list of values that presents the patch size (width, height). The default value is (256, 256). bytes_per_pixel: The number of bytes that each pixel in the 2D image takes place. The default value is 3.
 Returns:
int: The suggested memory capacity in MiB.
filesystem#
 class cucim.clara.filesystem.CuFileDriver#
Methods
close
(self)Closes opened file if not closed.
pread
(self, buf, count, file_offset[, ...])Reads up to count bytes from the file driver at offset file_offset (from the start of the file) into the buffer buf starting at offset buf_offset.
pwrite
(self, buf, count, file_offset[, ...])Reads up to count bytes from the file driver at offset file_offset (from the start of the file) into the buffer buf starting at offset buf_offset.
 close(self: cucim.clara._cucim.filesystem.CuFileDriver) bool #
Closes opened file if not closed.
 pread(self: cucim.clara._cucim.filesystem.CuFileDriver, buf: object, count: int, file_offset: int, buf_offset: int = 0) int #
Reads up to count bytes from the file driver at offset file_offset (from the start of the file) into the buffer buf starting at offset buf_offset. The file offset is not changed.
 Args:
buf: A buffer where read bytes are stored. Buffer can be either in CPU memory or (CUDA) GPU memory. count: The number of bytes to read. file_offset: An offset from the start of the file. buf_offset: An offset from the start of the buffer. Default value is 0.
 Returns:
The number of bytes read if succeed, 1 otherwise.
 pwrite(self: cucim.clara._cucim.filesystem.CuFileDriver, buf: object, count: int, file_offset: int, buf_offset: int = 0) int #
Reads up to count bytes from the file driver at offset file_offset (from the start of the file) into the buffer buf starting at offset buf_offset. The file offset is not changed.
 Args:
buf: A buffer where write bytes come from. Buffer can be either in CPU memory or (CUDA) GPU memory. count: The number of bytes to write. file_offset: An offset from the start of the file. buf_offset: An offset from the start of the buffer. Default value is 0.
 Returns:
The number of bytes written if succeed, 1 otherwise.
 cucim.clara.filesystem.close(arg0: cucim.clara._cucim.filesystem.CuFileDriver) bool #
Closes the given file driver.
 Args:
fd: An CuFileDriver object.
 Returns:
True if succeed, False otherwise.
 cucim.clara.filesystem.discard_page_cache(file_path: str) bool #
Discards a system (page) cache for the given file path.
 Args:
file_path: A file path to drop system cache.
 Returns:
True if succeed, False otherwise.
 cucim.clara.filesystem.open(file_path: str, flags: str, mode: int = 420) cucim.clara._cucim.filesystem.CuFileDriver #
Open file with specific flags and mode.
‘flags’ can be one of the following flag string:
“r”: os.O_RDONLY
“r+”: os.O_RDWR
“w”: os.O_RDWR  os.O_CREAT  os.O_TRUNC
“a”: os.O_RDWR  os.O_CREAT
In addition to above flags, the method append os.O_CLOEXEC and os.O_DIRECT by default.
The following is optional flags that can be added to above string:
‘p’: Use POSIX APIs only (first try to open with O_DIRECT). It does not use GDS.
‘n’: Do not add O_DIRECT flag.
‘m’: Use memorymapped file. This flag is supported only for the readonly file descriptor.
When ‘m’ is used, PROT_READ and MAP_SHARED are used for the parameter of mmap() function.
 Args:
file_path: A file path to open. flags: File flags in string. Default value is “r”. mode: A file mode. Default value is ‘0o644’.
 Returns:
An object of CuFileDriver.
 cucim.clara.filesystem.pread(fd: cucim.clara._cucim.filesystem.CuFileDriver, buf: object, count: int, file_offset: int, buf_offset: int = 0) int #
Reads up to count bytes from file driver fd at offset offset (from the start of the file) into the buffer buf starting at offset buf_offset. The file offset is not changed.
 Args:
fd: An object of CuFileDriver. buf: A buffer where read bytes are stored. Buffer can be either in CPU memory or (CUDA) GPU memory. count: The number of bytes to read. file_offset: An offset from the start of the file. buf_offset: An offset from the start of the buffer. Default value is 0.
 Returns:
The number of bytes read if succeed, 1 otherwise.
 cucim.clara.filesystem.pwrite(fd: cucim.clara._cucim.filesystem.CuFileDriver, buf: object, count: int, file_offset: int, buf_offset: int = 0) int #
Write up to count bytes from the buffer buf starting at offset buf_offset to the file driver fd at offset offset (from the start of the file). The file offset is not changed.
 Args:
fd: An object of CuFileDriver. buf: A buffer where write bytes come from. Buffer can be either in CPU memory or (CUDA) GPU memory. count: The number of bytes to write. file_offset: An offset from the start of the file. buf_offset: An offset from the start of the buffer. Default value is 0.
 Returns:
The number of bytes written if succeed, 1 otherwise.
io#
 class cucim.clara.io.Device#

Methods
parse_type
(arg0)Create DeviceType object from the device name string.
 property index#
Device index.
 static parse_type(arg0: str) cucim.clara._cucim.io.DeviceType #
Create DeviceType object from the device name string.
 property type#
Device type.
 class cucim.clara.io.DeviceType#
Members:
CPU
CUDA
CUDAHost
CUDAManaged
CPUShared
CUDAShared
 Attributes:
name
name(self: handle) > str
 value
 CPU = <DeviceType.CPU: 1>#
 CUDA = <DeviceType.CUDA: 2>#
 CUDAHost = <DeviceType.CUDAHost: 3>#
 CUDAManaged = <DeviceType.CUDAManaged: 13>#
 property name#
 property value#
core Submodules#
color#
 cucim.core.operations.color.color_jitter(img: Any, brightness=0, contrast=0, saturation=0, hue=0)#
Applies color jitter by random sequential application of 4 operations (brightness, contrast, saturation, hue).
 Parameters:
 imgchannel first, cupy.ndarray or numpy.ndarray
Input data of shape (C, H, W). Can also batch process input of shape (N, C, H, W). Can be a numpy.ndarray or cupy.ndarray.
 brightnessfloat or 2tuple of float, optional
Nonnegative factor to jitter the brightness by. When brightness is a scalar, scaling will be by a random value in range
[max(0, 1  brightness), (1 + brightness)]
. brightness can also be a 2tuple specifying the range for the random scaling factor. A value of 0 or (1, 1) will result in no change. contrastfloat or 2tuple of float, optional
Nonnegative factor to jitter the contrast by. When contrast is a scalar, scaling will be by a random value between
[max(0, 1  contrast), (1 + contrast)]
. contrast can also be a 2tuple specifying the range for the random scaling factor. A value of 0 or (1, 1) will result in no change. saturationfloat or 2tuple of float, optional
Nonnegative factor to jitter the saturation by. When saturation is a scalar, scaling will be by a random value between
[max(0, 1  saturation), (1 + saturation)]
. saturation can also be a 2tuple specifying the range for the random scaling factor. A value of 0 or (1, 1) will result in no change. huefloat or 2tuple of float, optional
Factor between [0.5, 0.5] to jitter hue by. When hue is a scalar, scaling will be by a random value between in the range
[hue, hue]
. hue can also be a 2tuple specifying the range. A value of 0 or (0, 0) will result in no change.
 Returns:
 outcupy.ndarray or numpy.ndarray
Output data. Same dimensions and type as input.
 Raises:
 ValueError
If ‘brightness’,’contrast’,’saturation’ or ‘hue’ is outside of allowed range
 TypeError
If input ‘img’ is not cupy.ndarray or numpy.ndarray
Examples
>>> import cucim.core.operations.color as ccl >>> # input is channel first 3d array >>> output_array = ccl.color_jitter(input_arr,.25,.75,.25,.04)
 cucim.core.operations.color.image_to_absorbance(image, source_intensity=255.0, dtype=<class 'numpy.float32'>)#
Convert an image to units of absorbance (optical density).
 Parameters:
 imagendarray
The image to convert to absorbance. Can be single or multichannel.
 source_intensityfloat, optional
Reference intensity for image.
 dtypenumpy.dtype, optional
The floating point precision at which to compute the absorbance.
 Returns:
 absorbancendarray
The absorbance computed from image.
Notes
If image has an integer dtype it will be clipped to range
[1, source_intensity]
, while float image inputs are clipped to range[source_intensity/255, source_intensity]
. The minimum is to avoid log(0). Absorbance is then given byabsorbance = log(image / source_intensity)
.
 cucim.core.operations.color.normalize_colors_pca(image, source_intensity: float = 240.0, alpha: float = 1.0, beta: float = 0.345, ref_stain_coeff: Union[tuple, ndarray] = ((0.5626, 0.2159), (0.7201, 0.8012), (0.4062, 0.5581)), ref_max_conc: Union[tuple, ndarray] = (1.9705, 1.0308), image_type: str = 'intensity', channel_axis: int = 0)#
Extract the matrix of stain coefficient from the image.
 Parameters:
 imagenp.ndarray
RGB image to determine concentrations for. Intensities should typically be within unsigned 8bit integer intensity range ([0, 255]) when
image_type == "intensity"
. source_intensityfloat, optional
Transmitted light intensity. The algorithm will clip image intensities above the specified source_intensity and then normalize by source_intensity so that image intensities are <= 1.0. Only used when image_type==”intensity”.
 alphafloat, optional
Algorithm parameter controlling the
[alpha, 100  alpha]
percentile range used as a robust [min, max] estimate. betafloat, optional
Absorbance (optical density) threshold below which to consider pixels as transparent. Transparent pixels are excluded from the estimation.
 ref_stain_coeffarraylike
Reference stain coefficients as determined by the output of stain_extraction_pca for a reference image.
 ref_max_conctuple or cp.ndarray
The reference maximum concentrations.
 image_type{“intensity”, “absorbance”}, optional
With the default image_type of “intensity”, the image will be transformed to an absorbance using
image_to_absorbance
. If the input image is already an absorbance image, then image_type should be set to “absorbance” instead. channel_axisint, optional
The axis corresponding to color channels (default is the last axis).
 Returns:
 stain_coeffnp.ndarray
Stain attenuation coefficient matrix derived from the image, where the first column corresponds to H, the second column is E and the rows are RGB values.
Notes
The default beta of 0.345 is equivalent to the use of 0.15 in [1]. The difference is due to our use of the natural log instead of a decadic log (log10) when computing the absorbance.
References
[1]M. Macenko et al., “A method for normalizing histology slides for quantitative analysis,” 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2009, pp. 11071110, doi: 10.1109/ISBI.2009.5193250. http://wwwx.cs.unc.edu/~mn/sites/default/files/macenko2009.pdf
 cucim.core.operations.color.stain_extraction_pca(image, source_intensity=240, alpha=1, beta=0.345, *, channel_axis=0, image_type='intensity')#
Extract the matrix of H & E stain coefficient from an image.
Uses a method that selects stain vectors based on the angle distribution within a bestfit plane determined by principle component analysis (PCA) [1].
 Parameters:
 imagecp.ndarray
RGB image to perform stain extraction on. Intensities should typically be within unsigned 8bit integer intensity range ([0, 255]) when
image_type == "intensity"
. source_intensityfloat, optional
Transmitted light intensity. The algorithm will clip image intensities above the specified source_intensity and then normalize by source_intensity so that image intensities are <= 1.0. Only used when image_type==”intensity”.
 alphafloat, optional
Algorithm parameter controlling the
[alpha, 100  alpha]
percentile range used as a robust [min, max] estimate. betafloat, optional
Absorbance (optical density) threshold below which to consider pixels as transparent. Transparent pixels are excluded from the estimation.
 Returns:
 stain_coeffcp.ndarray
Stain attenuation coefficient matrix derived from the image, where the first column corresponds to H, the second column is E and the rows are RGB values.
Notes
The default beta of 0.345 is equivalent to the use of 0.15 in [1]. The difference is due to our use of the natural log instead of a decadic log (log10) when computing the absorbance.
References
[1] (1,2)M. Macenko et al., “A method for normalizing histology slides for quantitative analysis,” 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2009, pp. 11071110, doi: 10.1109/ISBI.2009.5193250. http://wwwx.cs.unc.edu/~mn/sites/default/files/macenko2009.pdf
expose#
intensity#
 cucim.core.operations.intensity.normalize_data(img: Any, norm_constant: float, min_value: float, max_value: float, type: str = 'range') Any #
Apply intensity normalization to the input array. Normalize intensities to the range of [0, norm_constant].
 Parameters:
 imgchannel first, cupy.ndarray or numpy.ndarray
Input data of shape (C, H, W). Can also batch process input of shape (N, C, H, W). Can be a numpy.ndarray or cupy.ndarray.
 norm_constant: float
Normalization range of the input data. [0, norm_constant]
 min_valuefloat
Minimum intensity value in input data.
 max_valuefloat
Maximum intensity value in input data.
 type{‘range’, ‘atan’}
Type of normalization.
 Returns:
 outcupy.ndarray or numpy.ndarray
Output data. Same dimensions and type as input.
 Raises:
 TypeError
If input ‘img’ is not cupy.ndarray or numpy.ndarray
 ValueError
If input original intensity min and max are same
 ValueError
If incorrect normalization type is invoked
Examples
>>> import cucim.core.operations.intensity as its >>> # input is channel first 3d array >>> output_array = its.normalize_data(input_arr, 10, 0 , 255)
 cucim.core.operations.intensity.rand_zoom(img: Any, min_zoom: Union[Sequence[float], float] = 0.9, max_zoom: Union[Sequence[float], float] = 1.1, prob: float = 0.1, whole_batch: bool = False)#
Randomly Calls zoom with random zoom factor
 Parameters:
 imgchannel first, cupy.ndarray or numpy.ndarray
Input data of shape (C, H, W). Can also batch process input of shape (N, C, H, W). Can be a numpy.ndarray or cupy.ndarray.
 min_zoom: Min zoom factor. Can be float or sequence same size as image.
If a float, select a random factor from [min_zoom, max_zoom] then apply to all spatial dims to keep the original spatial shape ratio. If a sequence, min_zoom should contain one value for each spatial axis. If 2 values provided for 3D data, use the first value for both H & W dims to keep the same zoom ratio.
 max_zoom: Max zoom factor. Can be float or sequence same size as image.
If a float, select a random factor from [min_zoom, max_zoom] then apply to all spatial dims to keep the original spatial shape ratio. If a sequence, max_zoom should contain one value for each spatial axis. If 2 values provided for 3D data, use the first value for both H & W dims to keep the same zoom ratio.
 prob: Probability of zooming.
 whole_batch: Flag to apply transform on whole batch.
If False, each image in the batch is randomly transformed It True, entire batch is transformed randomly.
 Returns:
 outcupy.ndarray or numpy.ndarray
Output data. Same dimensions and type as input.
 Raises:
 TypeError
If input ‘img’ is not cupy.ndarray or numpy.ndarray
Examples
>>> import cucim.core.operations.intensity as its >>> # input is channel first 3d array >>> output_array = its.rand_zoom(input_arr)
 cucim.core.operations.intensity.scale_intensity_range(img: Any, b_max: float, b_min: float, a_max: float, a_min: float, clip: bool = False) Any #
Apply intensity scaling to the input array. Scaling from [a_min, a_max] to [b_min, b_max] with clip option.
 Parameters:
 imgchannel first, cupy.ndarray or numpy.ndarray
Input data of shape (C, H, W). Can also batch process input of shape (N, C, H, W). Can be a numpy.ndarray or cupy.ndarray.
 b_minfloat
intensity target range min.
 b_maxfloat
intensity target range max.
 a_minfloat
intensity original range min.
 a_maxfloat
intensity original range max.
 clipfloat
whether to perform clip after scaling.
 Returns:
 outcupy.ndarray or numpy.ndarray
Output data. Same dimensions and type as input.
 Raises:
 TypeError
If input ‘img’ is not cupy.ndarray or numpy.ndarray
 ValueError
If input original intensity min and max are same
Examples
>>> import cucim.core.operations.intensity as its >>> # input is channel first 3d array >>> output_array = its.scale_intensity_range(input_arr, 0.0, 255.0, 1.0, 1.0, False)
 cucim.core.operations.intensity.zoom(img: Any, zoom_factor: Sequence[float])#
Zooms an ND image
 Parameters:
 imgchannel first, cupy.ndarray or numpy.ndarray
Input data of shape (C, H, W). Can also batch process input of shape (N, C, H, W). Can be a numpy.ndarray or cupy.ndarray.
 zoom_factor: Sequence[float]
The zoom factor along the spatial axes. Zoom factor should contain one value for each spatial axis.
 Returns
 ——
 outcupy.ndarray or numpy.ndarray
Output data. Same dimensions and type as input.
 Raises:
 TypeError
If input ‘img’ is not cupy.ndarray or numpy.ndarray
Examples
>>> import cucim.core.operations.intensity as its >>> # input is channel first 3d array >>> output_array = its.zoom(input_arr,[1.1,1.1])
morphology#
 cucim.core.operations.morphology.distance_transform_edt(image, sampling=None, return_distances=True, return_indices=False, distances=None, indices=None, *, block_params=None, float64_distances=False)#
Exact Euclidean distance transform.
This function calculates the distance transform of the input, by replacing each foreground (nonzero) element, with its shortest distance to the background (any zerovalued element).
In addition to the distance transform, the feature transform can be calculated. In this case the index of the closest background element to each foreground element is returned in a separate array.
 Parameters:
 imagearray_like
Input data to transform. Can be any type but will be converted into binary: 1 wherever image equates to True, 0 elsewhere.
 samplingfloat, or sequence of float, optional
Spacing of elements along each dimension. If a sequence, must be of length equal to the image rank; if a single number, this is used for all axes. If not specified, a grid spacing of unity is implied.
 return_distancesbool, optional
Whether to calculate the distance transform.
 return_indicesbool, optional
Whether to calculate the feature transform.
 distancescupy.ndarray, optional
An output array to store the calculated distance transform, instead of returning it. return_distances must be
True
. It must be the same shape as image. Should have dtypecp.float32
if float64_distances isFalse
, otherwise it should becp.float64
. indicescupy.ndarray, optional
An output array to store the calculated feature transform, instead of returning it. return_indicies must be
True
. Its shape must be(image.ndim,) + image.shape
. Its dtype must be a signed or unsigned integer type of at least 16bits in 2D or 32bits in 3D.
 Returns:
 distancescupy.ndarray, optional
The calculated distance transform. Returned only when return_distances is
True
and distances is not supplied. It will have the same shape as image. Will have dtype cp.float64 if float64_distances isTrue
, otherwise it will have dtypecp.float32
. indicesndarray, optional
The calculated feature transform. It has an imageshaped array for each dimension of the image. See example below. Returned only when return_indices is
True
and indices is not supplied.
 Other Parameters:
 block_params3tuple of int
The m1, m2, m3 algorithm parameters as described in [2]. If None, suitable defaults will be chosen. Note: This parameter is specific to cuCIM and does not exist in SciPy.
 float64_distancesbool, optional
If True, use double precision in the distance computation (to match SciPy behavior). Otherwise, single precision will be used for efficiency. Note: This parameter is specific to cuCIM and does not exist in SciPy.
Notes
The Euclidean distance transform gives values of the Euclidean distance.
\[y_i = \sqrt{\sum_{i}^{n} (x[i]  b[i])^2}\]where \(b[i]\) is the background point (value 0) with the smallest Euclidean distance to input points \(x[i]\), and \(n\) is the number of dimensions.
Note that the indices output may differ from the one given by
scipy.ndimage.distance_transform_edt()
in the case of input pixels that are equidistant from multiple background points.The parallel banding algorithm implemented here was originally described in [1]. The kernels used here correspond to the revised PBA+ implementation that is described on the author’s website [2]. The source code of the author’s PBA+ implementation is available at [3].
References
[1]ThanhTung Cao, Ke Tang, Anis Mohamed, and TiowSeng Tan. 2010. Parallel Banding Algorithm to compute exact distance transform with the GPU. In Proceedings of the 2010 ACM SIGGRAPH symposium on Interactive 3D Graphics and Games (I3D ’10). Association for Computing Machinery, New York, NY, USA, 83–90. DOI:https://doi.org/10.1145/1730804.1730818
Examples
>>> import cupy as cp >>> from cucim.core.operations import morphology >>> a = cp.array(([0,1,1,1,1], ... [0,0,1,1,1], ... [0,1,1,1,1], ... [0,1,1,1,0], ... [0,1,1,0,0])) >>> morphology.distance_transform_edt(a) array([[ 0. , 1. , 1.4142, 2.2361, 3. ], [ 0. , 0. , 1. , 2. , 2. ], [ 0. , 1. , 1.4142, 1.4142, 1. ], [ 0. , 1. , 1.4142, 1. , 0. ], [ 0. , 1. , 1. , 0. , 0. ]])
With a sampling of 2 units along x, 1 along y:
>>> morphology.distance_transform_edt(a, sampling=[2,1]) array([[ 0. , 1. , 2. , 2.8284, 3.6056], [ 0. , 0. , 1. , 2. , 3. ], [ 0. , 1. , 2. , 2.2361, 2. ], [ 0. , 1. , 2. , 1. , 0. ], [ 0. , 1. , 1. , 0. , 0. ]])
Asking for indices as well:
>>> edt, inds = morphology.distance_transform_edt(a, return_indices=True) >>> inds array([[[0, 0, 1, 1, 3], [1, 1, 1, 1, 3], [2, 2, 1, 3, 3], [3, 3, 4, 4, 3], [4, 4, 4, 4, 4]], [[0, 0, 1, 1, 4], [0, 1, 1, 1, 4], [0, 0, 1, 4, 4], [0, 0, 3, 3, 4], [0, 0, 3, 3, 4]]])
spatial#
 cucim.core.operations.spatial.image_flip(img: Any, spatial_axis: ()) Any #
Shape preserving order reversal of elements in input array along the given spatial axis
 Parameters:
 imgcupy.ndarray or numpy.ndarray
Input data. Can be numpy.ndarray or cupy.ndarray
 spatial_axistuple
spatial axis along which to flip over the input array
 Returns:
 outcupy.ndarray or numpy.ndarray
Output data. Same dimensions and type as input.
 Raises:
 TypeError
If input ‘img’ is not cupy.ndarray or numpy.ndarray
Examples
>>> import cucim.core.operations.spatial as spt >>> # input is channel first 3d array >>> output_array = spt.image_flip(input_arr, (1, 2))
 cucim.core.operations.spatial.image_rotate_90(img: Any, k: int, spatial_axis: ()) Any #
Rotate input array by 90 degress along the given axis
 Parameters:
 imgcupy.ndarray or numpy.ndarray
Input data. Can be numpy.ndarray or cupy.ndarray
 kint
number of times to rotate
 spatial_axistuple
spatial axis along which to rotate the input array by 90 degrees
 Returns:
 outcupy.ndarray or numpy.ndarray
Output data. Same dimensions and type as input.
 Raises:
 TypeError
If input ‘img’ is not cupy.ndarray or numpy.ndarray
Examples
>>> import cucim.core.operations.spatial as spt >>> # input is channel first 3d array >>> output_array = spt.image_rotate_90(input_arr,1,(1,2))
 cucim.core.operations.spatial.rand_image_flip(img: Any, spatial_axis: (), prob: float = 0.1, whole_batch: bool = False) Any #
Randomly flips the image along axis.
 Parameters:
 imgcupy.ndarray or numpy.ndarray
Input data. Can be numpy.ndarray or cupy.ndarray
 prob: Probability of flipping.
 spatial_axistuple
spatial axis along which to flip over the input array
 whole_batch: Flag to apply transform on whole batch.
If False, each image in the batch is randomly transformed It True, entire batch is transformed randomly.
 Returns:
 outcupy.ndarray or numpy.ndarray
Output data. Same dimensions and type as input.
 Raises:
 TypeError
If input ‘img’ is not cupy.ndarray or numpy.ndarray
Examples
>>> import cucim.core.operations.spatial as spt >>> # input is channel first 3d array >>> output_array = spt.rand_image_flip(input_arr,spatial_axis=(1,2))
 cucim.core.operations.spatial.rand_image_rotate_90(img: Any, spatial_axis: (), prob: float = 0.1, max_k: int = 3, whole_batch: bool = False) Any #
With probability prob, input arrays are rotated by 90 degrees in the plane specified by spatial_axis.
 Parameters:
 imgcupy.ndarray or numpy.ndarray
Input data. Can be numpy.ndarray or cupy.ndarray
 prob: probability of rotating.
(Default 0.1, with 10% probability it returns a rotated array)
 max_k: number of rotations
will be sampled from np.random.randint(max_k) + 1, (Default 3).
 spatial_axistuple
spatial axis along which to rotate the input array by 90 degrees
 whole_batch: Flag to apply transform on whole batch.
If False, each image in the batch is randomly transformed It True, entire batch is transformed randomly.
 Returns:
 outcupy.ndarray or numpy.ndarray
Output data. Same dimensions and type as input.
 Raises:
 TypeError
If input ‘img’ is not cupy.ndarray or numpy.ndarray
Examples
>>> import cucim.core.operations.spatial as spt >>> # input is channel first 3d array >>> output_array = spt.rand_image_rotate_90(input_arr, spatial_axis=(1, 2))
skimage Submodules#
color#
 cucim.skimage.color.combine_stains(stains, conv_matrix, *, channel_axis=1)#
Stain to RGB color space conversion.
 Parameters:
 stains(…, 3, …) array_like
The image in stain color space. By default, the final dimension denotes channels.
 conv_matrix: ndarray
The stain separation matrix as described by G. Landini [1].
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If stains is not at least 2D with shape (…, 3, …).
Notes
Stain combination matrices available in the
color
module and their respective colorspace:rgb_from_hed
: Hematoxylin + Eosin + DABrgb_from_hdx
: Hematoxylin + DABrgb_from_fgx
: Feulgen + Light Greenrgb_from_bex
: Giemsa stain : Methyl Blue + Eosinrgb_from_rbd
: FastRed + FastBlue + DABrgb_from_gdx
: Methyl Green + DABrgb_from_hax
: Hematoxylin + AECrgb_from_bro
: Blue matrix Anilline Blue + Red matrix Azocarmine + Orange matrix OrangeGrgb_from_bpx
: Methyl Blue + Ponceau Fuchsinrgb_from_ahx
: Alcian Blue + Hematoxylinrgb_from_hpx
: Hematoxylin + PAS
References
[1][2]A. C. Ruifrok and D. A. Johnston, “Quantification of histochemical staining by color deconvolution,” Anal. Quant. Cytol. Histol., vol. 23, no. 4, pp. 291–299, Aug. 2001.
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage.color import (separate_stains, combine_stains, ... hdx_from_rgb, rgb_from_hdx) >>> ihc = cp.array(data.immunohistochemistry()) >>> ihc_hdx = separate_stains(ihc, hdx_from_rgb) >>> ihc_rgb = combine_stains(ihc_hdx, rgb_from_hdx)
 cucim.skimage.color.convert_colorspace(arr, fromspace, tospace, *, channel_axis=1)#
Convert an image array to a new color space.
 Valid color spaces are:
‘RGB’, ‘HSV’, ‘RGB CIE’, ‘XYZ’, ‘YUV’, ‘YIQ’, ‘YPbPr’, ‘YCbCr’, ‘YDbDr’
 Parameters:
 arr(…, 3, …) array_like
The image to convert. By default, the final dimension denotes channels.
 fromspacestr
The color space to convert from. Can be specified in lower case.
 tospacestr
The color space to convert to. Can be specified in lower case.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The converted image. Same dimensions as input.
 Raises:
 ValueError
If fromspace is not a valid color space
 ValueError
If tospace is not a valid color space
Notes
Conversion is performed through the “central” RGB color space, i.e. conversion from XYZ to HSV is implemented as
XYZ > RGB > HSV
instead of directly.Examples
>>> import cupy as cp >>> from skimage import data >>> img = cp.array(data.astronaut()) >>> img_hsv = convert_colorspace(img, 'RGB', 'HSV')
 cucim.skimage.color.deltaE_cie76(lab1, lab2, channel_axis=1)#
Euclidean distance between two points in Lab color space
 Parameters:
 lab1array_like
reference color (Lab colorspace)
 lab2array_like
comparison color (Lab colorspace)
 channel_axisint, optional
This parameter indicates which axis of the arrays corresponds to channels.
 Returns:
 dEarray_like
distance between colors lab1 and lab2
References
[2]A. R. Robertson, “The CIE 1976 colordifference formulae,” Color Res. Appl. 2, 711 (1977).
 cucim.skimage.color.deltaE_ciede2000(lab1, lab2, kL=1, kC=1, kH=1, *, channel_axis=1)#
Color difference as given by the CIEDE 2000 standard.
CIEDE 2000 is a major revision of CIDE94. The perceptual calibration is largely based on experience with automotive paint on smooth surfaces.
 Parameters:
 lab1array_like
reference color (Lab colorspace)
 lab2array_like
comparison color (Lab colorspace)
 kLfloat (range), optional
lightness scale factor, 1 for “acceptably close”; 2 for “imperceptible” see deltaE_cmc
 kCfloat (range), optional
chroma scale factor, usually 1
 kHfloat (range), optional
hue scale factor, usually 1
 channel_axisint, optional
This parameter indicates which axis of the arrays corresponds to channels.
 Returns:
 deltaEarray_like
The distance between lab1 and lab2
Notes
CIEDE 2000 assumes parametric weighting factors for the lightness, chroma, and hue (kL, kC, kH respectively). These default to 1.
References
[3]M. Melgosa, J. Quesada, and E. Hita, “Uniformity of some recent color metrics tested with an accurate colordifference tolerance dataset,” Appl. Opt. 33, 80698077 (1994).
 cucim.skimage.color.deltaE_ciede94(lab1, lab2, kH=1, kC=1, kL=1, k1=0.045, k2=0.015, *, channel_axis=1)#
Color difference according to CIEDE 94 standard
Accommodates perceptual nonuniformities through the use of application specific scale factors (kH, kC, kL, k1, and k2).
 Parameters:
 lab1array_like
reference color (Lab colorspace)
 lab2array_like
comparison color (Lab colorspace)
 kHfloat, optional
Hue scale
 kCfloat, optional
Chroma scale
 kLfloat, optional
Lightness scale
 k1float, optional
first scale parameter
 k2float, optional
second scale parameter
 channel_axisint, optional
This parameter indicates which axis of the arrays corresponds to channels.
 Returns:
 dEarray_like
color difference between lab1 and lab2
Notes
deltaE_ciede94 is not symmetric with respect to lab1 and lab2. CIEDE94 defines the scales for the lightness, hue, and chroma in terms of the first color. Consequently, the first color should be regarded as the “reference” color.
kL, k1, k2 depend on the application and default to the values suggested for graphic arts
Parameter
Graphic Arts
Textiles
kL
1.000
2.000
k1
0.045
0.048
k2
0.015
0.014
References
 cucim.skimage.color.deltaE_cmc(lab1, lab2, kL=1, kC=1, *, channel_axis=1)#
Color difference from the CMC l:c standard.
This color difference was developed by the Colour Measurement Committee (CMC) of the Society of Dyers and Colourists (United Kingdom). It is intended for use in the textile industry.
The scale factors kL, kC set the weight given to differences in lightness and chroma relative to differences in hue. The usual values are
kL=2
,kC=1
for “acceptability” andkL=1
,kC=1
for “imperceptibility”. Colors withdE > 1
are “different” for the given scale factors. Parameters:
 lab1array_like
reference color (Lab colorspace)
 lab2array_like
comparison color (Lab colorspace)
 channel_axisint, optional
This parameter indicates which axis of the arrays corresponds to channels.
 Returns:
 dEarray_like
distance between colors lab1 and lab2
Notes
deltaE_cmc the defines the scales for the lightness, hue, and chroma in terms of the first color. Consequently
deltaE_cmc(lab1, lab2) != deltaE_cmc(lab2, lab1)
References
[3]F. J. J. Clarke, R. McDonald, and B. Rigg, “Modification to the JPC79 colourdifference formula,” J. Soc. Dyers Colour. 100, 128132 (1984).
 cucim.skimage.color.gray2rgb(image, *, channel_axis=1)#
Create an RGB representation of a graylevel image.
 Parameters:
 imagearray_like
Input image.
 channel_axisint, optional
This parameter indicates which axis of the output array will correspond to channels.
 Returns:
 rgb(…, 3, …) ndarray
RGB image. A new dimension of length 3 is added to input image.
Notes
If the input is a 1dimensional image of shape
(M, )
, the output will be shape(M, 3)
.
 cucim.skimage.color.gray2rgba(image, alpha=None, *, channel_axis=1)#
Create a RGBA representation of a graylevel image.
 Parameters:
 imagearray_like
Input image.
 alphaarray_like, optional
Alpha channel of the output image. It may be a scalar or an array that can be broadcast to
image
. If not specified it is set to the maximum limit corresponding to theimage
dtype. channel_axisint, optional
This parameter indicates which axis of the output array will correspond to channels.
 Returns:
 rgbandarray
RGBA image. A new dimension of length 4 is added to input image shape.
 cucim.skimage.color.hed2rgb(hed, *, channel_axis=1)#
HaematoxylinEosinDAB (HED) to RGB color space conversion.
 Parameters:
 hed(…, 3, …) array_like
The image in the HED color space. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB. Same dimensions as input.
 Raises:
 ValueError
If hed is not at least 2D with shape (…, 3, …).
References
[1]A. C. Ruifrok and D. A. Johnston, “Quantification of histochemical staining by color deconvolution.,” Analytical and quantitative cytology and histology / the International Academy of Cytology [and] American Society of Cytology, vol. 23, no. 4, pp. 2919, Aug. 2001.
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage.color import rgb2hed, hed2rgb >>> ihc = cp.array(data.immunohistochemistry()) >>> ihc_hed = rgb2hed(ihc) >>> ihc_rgb = hed2rgb(ihc_hed)
 cucim.skimage.color.hsv2rgb(hsv, *, channel_axis=1)#
HSV to RGB color space conversion.
 Parameters:
 hsv(…, 3, …) array_like
The image in HSV format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If hsv is not at least 2D with shape (…, 3, …).
Notes
Conversion between RGB and HSV color spaces results in some loss of precision, due to integer arithmetic and rounding [1].
References
Examples
>>> import cupy as cp >>> from skimage import data >>> img = cp.array(data.astronaut()) >>> img_hsv = rgb2hsv(img) >>> img_rgb = hsv2rgb(img_hsv)
 cucim.skimage.color.lab2lch(lab, *, channel_axis=1)#
CIELAB to CIELCH color space conversion.
LCH is the cylindrical representation of the LAB (Cartesian) colorspace
 Parameters:
 lab(…, 3, …) array_like
The ND image in CIELAB format. The last (
N+1
th) dimension must have at least 3 elements, corresponding to theL
,a
, andb
color channels. Subsequent elements are copied. channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in LCH format, in a ND array with same shape as input lab.
 Raises:
 ValueError
If lch does not have at least 3 color channels (i.e. l, a, b).
Notes
The Hue is expressed as an angle between
(0, 2*pi)
Examples
>>> from skimage import data >>> from cucim.skimage.color import rgb2lab, lab2lch >>> img = cp.array(data.astronaut()) >>> img_lab = rgb2lab(img) >>> img_lch = lab2lch(img_lab)
 cucim.skimage.color.lab2rgb(lab, illuminant='D65', observer='2', *, channel_axis=1)#
Convert image in CIELAB to sRGB color space.
 Parameters:
 lab(…, 3, …) array_like
The input image in CIELAB color space. Unless channel_axis is set, the final dimension denotes the CIELAB channels. The L* values range from 0 to 100; the a* and b* values range from 128 to 127.
 illuminant{“A”, “B”, “C”, “D50”, “D55”, “D65”, “D75”, “E”}, optional
The name of the illuminant (the function is NOT case sensitive).
 observer{“2”, “10”, “R”}, optional
The aperture angle of the observer.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in sRGB color space, of same shape as input.
 Raises:
 ValueError
If lab is not at least 2D with shape (…, 3, …).
See also
Notes
This function uses
lab2xyz()
andxyz2rgb()
. The CIE XYZ tristimulus values are x_ref = 95.047, y_ref = 100., and z_ref = 108.883. See functionxyz_tristimulus_values()
for a list of supported illuminants.References
 cucim.skimage.color.lab2xyz(lab, illuminant='D65', observer='2', *, channel_axis=1)#
Convert image in CIELAB to XYZ color space.
 Parameters:
 lab(…, 3, …) array_like
The input image in CIELAB color space. Unless channel_axis is set, the final dimension denotes the CIELAB channels. The L* values range from 0 to 100; the a* and b* values range from 128 to 127.
 illuminant{“A”, “B”, “C”, “D50”, “D55”, “D65”, “D75”, “E”}, optional
The name of the illuminant (the function is NOT case sensitive).
 observer{“2”, “10”, “R”}, optional
The aperture angle of the observer.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in XYZ color space, of same shape as input.
 Raises:
 ValueError
If lab is not at least 2D with shape (…, 3, …).
 ValueError
If either the illuminant or the observer angle are not supported or unknown.
 UserWarning
If any of the pixels are invalid (Z < 0).
See also
Notes
The CIE XYZ tristimulus values are x_ref = 95.047, y_ref = 100., and z_ref = 108.883. See function
xyz_tristimulus_values()
for a list of supported illuminants.References
 cucim.skimage.color.label2rgb(label, image=None, colors=None, alpha=0.3, bg_label=0, bg_color=(0, 0, 0), image_alpha=1, kind='overlay', *, saturation=0, channel_axis=1)#
Return an RGB image where colorcoded labels are painted over the image.
 Parameters:
 labelndarray
Integer array of labels with the same shape as image.
 imagendarray, optional
Image used as underlay for labels. It should have the same shape as labels, optionally with an additional RGB (channels) axis. If image is an RGB image, it is converted to grayscale before coloring.
 colorslist, optional
List of colors. If the number of labels exceeds the number of colors, then the colors are cycled.
 alphafloat [0, 1], optional
Opacity of colorized labels. Ignored if image is None.
 bg_labelint, optional
Label that’s treated as the background. If bg_label is specified, bg_color is None, and kind is overlay, background is not painted by any colors.
 bg_colorstr or array, optional
Background color. Must be a name in color_dict or RGB float values between [0, 1].
 image_alphafloat [0, 1], optional
Opacity of the image.
 kindstring, one of {‘overlay’, ‘avg’}
The kind of color image desired. ‘overlay’ cycles over defined colors and overlays the colored labels over the original image. ‘avg’ replaces each labeled segment with its average color, for a stainedclass or pastel painting appearance.
 saturationfloat [0, 1], optional
Parameter to control the saturation applied to the original image between fully saturated (original RGB, saturation=1) and fully unsaturated (grayscale, saturation=0). Only applies when kind=’overlay’.
 channel_axisint, optional
This parameter indicates which axis of the output array will correspond to channels. If image is provided, this must also match the axis of image that corresponds to channels.
 Returns:
 resultarray of float, shape (M, N, 3)
The result of blending a cycling colormap (colors) for each distinct value in label with the image, at a certain alpha value.
 cucim.skimage.color.lch2lab(lch, *, channel_axis=1)#
CIELCH to CIELAB color space conversion.
LCH is the cylindrical representation of the LAB (Cartesian) colorspace
 Parameters:
 lch(…, 3, …) array_like
The ND image in CIELCH format. The last (
N+1
th) dimension must have at least 3 elements, corresponding to theL
,a
, andb
color channels. Subsequent elements are copied. channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in LAB format, with same shape as input lch.
 Raises:
 ValueError
If lch does not have at least 3 color channels (i.e. l, c, h).
Examples
>>> from skimage import data >>> from cucim.skimage.color import rgb2lab, lch2lab >>> img = cp.array(data.astronaut()) >>> img_lab = rgb2lab(img) >>> img_lch = lab2lch(img_lab) >>> img_lab2 = lch2lab(img_lch)
 cucim.skimage.color.luv2rgb(luv, *, channel_axis=1)#
Luv to RGB color space conversion.
 Parameters:
 luv(…, 3, …) array_like
The image in CIE Luv format. By default, the final dimension denotes channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If luv is not at least 2D with shape (…, 3, …).
Notes
This function uses luv2xyz and xyz2rgb.
 cucim.skimage.color.luv2xyz(luv, illuminant='D65', observer='2', *, channel_axis=1)#
CIELuv to XYZ color space conversion.
 Parameters:
 luv(…, 3, …) array_like
The image in CIELuv format. By default, the final dimension denotes channels.
 illuminant{“A”, “B”, “C”, “D50”, “D55”, “D65”, “D75”, “E”}, optional
The name of the illuminant (the function is NOT case sensitive).
 observer{“2”, “10”, “R”}, optional
The aperture angle of the observer.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in XYZ format. Same dimensions as input.
 Raises:
 ValueError
If luv is not at least 2D with shape (…, 3, …).
 ValueError
If either the illuminant or the observer angle are not supported or unknown.
Notes
XYZ conversion weights use observer=2A. Reference whitepoint for D65 Illuminant, with XYZ tristimulus values of
(95.047, 100., 108.883)
. See functionxyz_tristimulus_values()
for a list of supported illuminants.References
 cucim.skimage.color.rgb2gray(rgb, *, channel_axis=1)#
Compute luminance of an RGB image.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 Returns:
 outndarray
The luminance image  an array which is the same size as the input array, but with the channel dimension removed.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
Notes
The weights used in this conversion are calibrated for contemporary CRT phosphors:
Y = 0.2125 R + 0.7154 G + 0.0721 B
If there is an alpha channel present, it is ignored.
References
Examples
>>> import cupy as cp >>> from cucim.skimage.color import rgb2gray >>> from skimage import data >>> img = cp.array(data.astronaut()) >>> img_gray = rgb2gray(img)
 cucim.skimage.color.rgb2hed(rgb, *, channel_axis=1)#
RGB to HaematoxylinEosinDAB (HED) color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in HED format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
References
[1]A. C. Ruifrok and D. A. Johnston, “Quantification of histochemical staining by color deconvolution.,” Analytical and quantitative cytology and histology / the International Academy of Cytology [and] American Society of Cytology, vol. 23, no. 4, pp. 2919, Aug. 2001.
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage.color import rgb2hed >>> ihc = cp.array(data.immunohistochemistry()) >>> ihc_hed = rgb2hed(ihc)
 cucim.skimage.color.rgb2hsv(rgb, *, channel_axis=1)#
RGB to HSV color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in HSV format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
Notes
Conversion between RGB and HSV color spaces results in some loss of precision, due to integer arithmetic and rounding [1].
References
Examples
>>> import cupy as cp >>> from cucim.skimage import color >>> from skimage import data >>> img = cp.array(data.astronaut()) >>> img_hsv = color.rgb2hsv(img)
 cucim.skimage.color.rgb2lab(rgb, illuminant='D65', observer='2', *, channel_axis=1)#
Conversion from the sRGB color space (IEC 6196621:1999) to the CIE Lab colorspace under the given illuminant and observer.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 illuminant{“A”, “B”, “C”, “D50”, “D55”, “D65”, “D75”, “E”}, optional
The name of the illuminant (the function is NOT case sensitive).
 observer{“2”, “10”, “R”}, optional
The aperture angle of the observer.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in Lab format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
Notes
RGB is a devicedependent color space so, if you use this function, be sure that the image you are analyzing has been mapped to the sRGB color space.
This function uses rgb2xyz and xyz2lab. By default Observer=”2”, Illuminant=”D65”. CIE XYZ tristimulus values x_ref=95.047, y_ref=100., z_ref=108.883. See function
xyz_tristimulus_values()
for a list of supported illuminants.References
 cucim.skimage.color.rgb2luv(rgb, *, channel_axis=1)#
RGB to CIELuv color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in CIE Luv format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
Notes
This function uses rgb2xyz and xyz2luv.
References
 cucim.skimage.color.rgb2rgbcie(rgb, *, channel_axis=1)#
RGB to RGB CIE color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB CIE format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
References
Examples
>>> from skimage import data >>> from cucim.skimage.color import rgb2rgbcie >>> img = cp.array(data.astronaut()) >>> img_rgbcie = rgb2rgbcie(img)
 cucim.skimage.color.rgb2xyz(rgb, *, channel_axis=1)#
RGB to XYZ color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in XYZ format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
Notes
The CIE XYZ color space is derived from the CIE RGB color space. Note however that this function converts from sRGB.
References
Examples
>>> import cupy as cp >>> from skimage import data >>> img = cp.array(data.astronaut()) >>> img_xyz = rgb2xyz(img)
 cucim.skimage.color.rgb2ycbcr(rgb, *, channel_axis=1)#
RGB to YCbCr color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in YCbCr format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
Notes
Y is between 16 and 235. This is the color space commonly used by video codecs; it is sometimes incorrectly called “YUV”.
References
 cucim.skimage.color.rgb2ydbdr(rgb, *, channel_axis=1)#
RGB to YDbDr color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in YDbDr format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
Notes
This is the color space commonly used by video codecs. It is also the reversible color transform in JPEG2000.
References
 cucim.skimage.color.rgb2yiq(rgb, *, channel_axis=1)#
RGB to YIQ color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in YIQ format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
 cucim.skimage.color.rgb2ypbpr(rgb, *, channel_axis=1)#
RGB to YPbPr color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in YPbPr format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
References
 cucim.skimage.color.rgb2yuv(rgb, *, channel_axis=1)#
RGB to YUV color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in YUV format. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
Notes
Y is between 0 and 1. Use YCbCr instead of YUV for the color space commonly used by video codecs, where Y ranges from 16 to 235.
References
 cucim.skimage.color.rgba2rgb(rgba, background=(1, 1, 1), *, channel_axis=1)#
RGBA to RGB conversion using alpha blending [1].
 Parameters:
 rgba(…, 4, …) array_like
The image in RGBA format. By default, the final dimension denotes channels.
 backgroundarray_like
The color of the background to blend the image with (3 floats between 0 to 1  the RGB value of the background).
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If rgba is not at least 2D with shape (…, 4, …).
References
Examples
>>> import cupy as cp >>> from cucim.skimage import color >>> from skimage import data >>> img_rgba = cp.array(data.logo()) >>> img_rgb = color.rgba2rgb(img_rgba)
 cucim.skimage.color.rgbcie2rgb(rgbcie, *, channel_axis=1)#
RGB CIE to RGB color space conversion.
 Parameters:
 rgbcie(…, 3, …) array_like
The image in RGB CIE format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If rgbcie is not at least 2D with shape (…, 3, …).
References
Examples
>>> from skimage import data >>> from cucim.skimage.color import rgb2rgbcie, rgbcie2rgb >>> img = cp.array(data.astronaut()) >>> img_rgbcie = rgb2rgbcie(img) >>> img_rgb = rgbcie2rgb(img_rgbcie)
 cucim.skimage.color.separate_stains(rgb, conv_matrix, *, channel_axis=1)#
RGB to stain color space conversion.
 Parameters:
 rgb(…, 3, …) array_like
The image in RGB format. By default, the final dimension denotes channels.
 conv_matrix: ndarray
The stain separation matrix as described by G. Landini [1].
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in stain color space. Same dimensions as input.
 Raises:
 ValueError
If rgb is not at least 2D with shape (…, 3, …).
Notes
Stain separation matrices available in the
color
module and their respective colorspace:hed_from_rgb
: Hematoxylin + Eosin + DABhdx_from_rgb
: Hematoxylin + DABfgx_from_rgb
: Feulgen + Light Greenbex_from_rgb
: Giemsa stain : Methyl Blue + Eosinrbd_from_rgb
: FastRed + FastBlue + DABgdx_from_rgb
: Methyl Green + DABhax_from_rgb
: Hematoxylin + AECbro_from_rgb
: Blue matrix Anilline Blue + Red matrix Azocarmine + Orange matrix OrangeGbpx_from_rgb
: Methyl Blue + Ponceau Fuchsinahx_from_rgb
: Alcian Blue + Hematoxylinhpx_from_rgb
: Hematoxylin + PAS
This implementation borrows some ideas from DIPlib [2], e.g. the compensation using a small value to avoid log artifacts when calculating the BeerLambert law.
References
[1][3]A. C. Ruifrok and D. A. Johnston, “Quantification of histochemical staining by color deconvolution,” Anal. Quant. Cytol. Histol., vol. 23, no. 4, pp. 291–299, Aug. 2001.
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage.color import separate_stains, hdx_from_rgb >>> ihc = cp.array(data.immunohistochemistry()) >>> ihc_hdx = separate_stains(ihc, hdx_from_rgb)
 cucim.skimage.color.xyz2lab(xyz, illuminant='D65', observer='2', *, channel_axis=1)#
XYZ to CIELAB color space conversion.
 Parameters:
 xyz(…, 3, …) array_like
The image in XYZ format. By default, the final dimension denotes channels.
 illuminant{“A”, “B”, “C”, “D50”, “D55”, “D65”, “D75”, “E”}, optional
The name of the illuminant (the function is NOT case sensitive).
 observer{“2”, “10”, “R”}, optional
One of: 2degree observer, 10degree observer, or ‘R’ observer as in R function grDevices::convertColor.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in CIELAB format. Same dimensions as input.
 Raises:
 ValueError
If xyz is not at least 2D with shape (…, 3, …).
 ValueError
If either the illuminant or the observer angle is unsupported or unknown.
Notes
By default Observer=”2”, Illuminant=”D65”. CIE XYZ tristimulus values x_ref=95.047, y_ref=100., z_ref=108.883. See function
xyz_tristimulus_values()
for a list of supported illuminants.References
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage.color import rgb2xyz, xyz2lab >>> img = cp.array(data.astronaut()) >>> img_xyz = rgb2xyz(img) >>> img_lab = xyz2lab(img_xyz)
 cucim.skimage.color.xyz2luv(xyz, illuminant='D65', observer='2', *, channel_axis=1)#
XYZ to CIELuv color space conversion.
 Parameters:
 xyz(…, 3, …) array_like
The image in XYZ format. By default, the final dimension denotes channels.
 illuminant{“A”, “B”, “C”, “D50”, “D55”, “D65”, “D75”, “E”}, optional
The name of the illuminant (the function is NOT case sensitive).
 observer{“2”, “10”, “R”}, optional
The aperture angle of the observer.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in CIELuv format. Same dimensions as input.
 Raises:
 ValueError
If xyz is not at least 2D with shape (…, 3, …).
 ValueError
If either the illuminant or the observer angle are not supported or unknown.
Notes
By default XYZ conversion weights use observer=2A. Reference whitepoint for D65 Illuminant, with XYZ tristimulus values of
(95.047, 100., 108.883)
. See functionxyz_tristimulus_values()
for a list of supported illuminants.References
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage.color import rgb2xyz, xyz2luv >>> img = cp.array(data.astronaut()) >>> img_xyz = rgb2xyz(img) >>> img_luv = xyz2luv(img_xyz)
 cucim.skimage.color.xyz2rgb(xyz, *, channel_axis=1)#
XYZ to RGB color space conversion.
 Parameters:
 xyz(…, 3, …) array_like
The image in XYZ format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If xyz is not at least 2D with shape (…, 3, …).
Notes
The CIE XYZ color space is derived from the CIE RGB color space. Note however that this function converts to sRGB.
References
Examples
>>> from skimage import data >>> from cucim.skimage.color import rgb2xyz, xyz2rgb >>> img = cp.array(data.astronaut()) >>> img_xyz = rgb2xyz(img) >>> img_rgb = xyz2rgb(img_xyz)
 cucim.skimage.color.xyz_tristimulus_values(*, illuminant, observer, dtype=None)#
Get the CIE XYZ tristimulus values.
Given an illuminant and observer, this function returns the CIE XYZ tristimulus values [2] scaled such that \(Y = 1\).
 Parameters:
 illuminant{“A”, “B”, “C”, “D50”, “D55”, “D65”, “D75”, “E”}
The name of the illuminant (the function is NOT case sensitive).
 observer{“2”, “10”, “R”}
One of: 2degree observer, 10degree observer, or ‘R’ observer as in R function
grDevices::convertColor
[3]. dtypenp.dtype, optional
This argument is ignored in the cuCIM implementation of xyz_tristimulus_values since an array is not returned. The output is always a 3tuple of float.
 Returns:
 values3tuple of float
Three elements \(X, Y, Z\) containing the CIE XYZ tristimulus values of the given illuminant.
 Raises:
 ValueError
If either the illuminant or the observer angle are not supported or unknown.
Notes
The return type of this function differs from the one in scikitimage as it always returns a 3tuple of float rather than an array with a userspecified dtype.
The CIE XYZ tristimulus values are calculated from \(x, y\) [1], using the formula
\[X = x / y\]\[Y = 1\]\[Z = (1  x  y) / y\]The only exception is the illuminant “D65” with aperture angle 2° for backwardcompatibility reasons.
References
Examples
Get the CIE XYZ tristimulus values for a “D65” illuminant for a 10 degree field of view
>>> xyz_tristimulus_values(illuminant="D65", observer="10") array([0.94809668, 1. , 1.07305136])
 cucim.skimage.color.ycbcr2rgb(ycbcr, *, channel_axis=1)#
YCbCr to RGB color space conversion.
 Parameters:
 ycbcr(…, 3, …) array_like
The image in YCbCr format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If ycbcr is not at least 2D with shape (…, 3, …).
Notes
Y is between 16 and 235. This is the color space commonly used by video codecs; it is sometimes incorrectly called “YUV”.
References
 cucim.skimage.color.ydbdr2rgb(ydbdr, *, channel_axis=1)#
YDbDr to RGB color space conversion.
 Parameters:
 ydbdr(…, 3, …) array_like
The image in YDbDr format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If ydbdr is not at least 2D with shape (…, 3, …).
Notes
This is the color space commonly used by video codecs, also called the reversible color transform in JPEG2000.
References
 cucim.skimage.color.yiq2rgb(yiq, *, channel_axis=1)#
YIQ to RGB color space conversion.
 Parameters:
 yiq(…, 3, …) array_like
The image in YIQ format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If yiq is not at least 2D with shape (…, 3, …).
 cucim.skimage.color.ypbpr2rgb(ypbpr, *, channel_axis=1)#
YPbPr to RGB color space conversion.
 Parameters:
 ypbpr(…, 3, …) array_like
The image in YPbPr format. By default, the final dimension denotes channels.
 channel_axisint, optional
This parameter indicates which axis of the array corresponds to channels.
 Returns:
 out(…, 3) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If ypbpr is not at least 2D with shape (…, 3).
References
 cucim.skimage.color.yuv2rgb(yuv, *, channel_axis=1)#
YUV to RGB color space conversion.
 Parameters:
 yuv(…, 3, …) array_like
The image in YUV format. By default, the final dimension denotes channels.
 Returns:
 out(…, 3, …) ndarray
The image in RGB format. Same dimensions as input.
 Raises:
 ValueError
If yuv is not at least 2D with shape (…, 3, …).
References
data#
 cucim.skimage.data.binary_blobs(length=512, blob_size_fraction=0.1, n_dim=2, volume_fraction=0.5, seed=None)#
Generate synthetic binary image with several rounded bloblike objects.
 Parameters:
 lengthint, optional
Linear size of output image.
 blob_size_fractionfloat, optional
Typical linear size of blob, as a fraction of
length
, should be smaller than 1. n_dimint, optional
Number of dimensions of output image.
 volume_fractionfloat, default 0.5
Fraction of image pixels covered by the blobs (where the output is 1). Should be in [0, 1].
 seed{None, int, cupy.random.Generator}, optional
If seed is None the cupy.random.Generator singleton is used. If seed is an int, a new
Generator
instance is used, seeded with seed. If seed is already aGenerator
instance then that instance is used.
 Returns:
 blobsndarray of bools
Output binary image
Notes
Warning: CuPy does not give identical randomly generated numbers as NumPy, so using a specific seed here will not give an identical pattern to the scikitimage implementation.
The behavior for a given random seed may also change across CuPy major versions. See: https://docs.cupy.dev/en/stable/reference/random.html
Examples
>>> from cucim.skimage import data >>> # tiny size (5, 5) >>> blobs = data.binary_blobs(length=5, blob_size_fraction=0.2, seed=1) >>> # larger size >>> blobs = data.binary_blobs(length=256, blob_size_fraction=0.1) >>> # Finer structures >>> blobs = data.binary_blobs(length=256, blob_size_fraction=0.05) >>> # Blobs cover a smaller volume fraction of the image >>> blobs = data.binary_blobs(length=256, volume_fraction=0.3)
exposure#
 cucim.skimage.exposure.adjust_gamma(image, gamma=1, gain=1)#
Performs Gamma Correction on the input image.
Also known as Power Law Transform. This function transforms the input image pixelwise according to the equation
O = I**gamma
after scaling each pixel to the range 0 to 1. Parameters:
 imagendarray
Input image.
 gammafloat, optional
Non negative real number. Default value is 1.
 gainfloat, optional
The constant multiplier. Default value is 1.
 Returns:
 outndarray
Gamma corrected output image.
See also
Notes
For gamma greater than 1, the histogram will shift towards left and the output image will be darker than the input image.
For gamma less than 1, the histogram will shift towards right and the output image will be brighter than the input image.
References
Examples
>>> from skimage import data >>> from cucim.skimage import exposure, img_as_float >>> image = img_as_float(cp.array(data.moon())) >>> gamma_corrected = exposure.adjust_gamma(image, 2) >>> # Output is darker for gamma > 1 >>> image.mean() > gamma_corrected.mean() array(True)
 cucim.skimage.exposure.adjust_log(image, gain=1, inv=False)#
Performs Logarithmic correction on the input image.
This function transforms the input image pixelwise according to the equation
O = gain*log(1 + I)
after scaling each pixel to the range 0 to 1.For inverse logarithmic correction, the equation is
O = gain*(2**I  1)
. Parameters:
 imagendarray
Input image.
 gainfloat, optional
The constant multiplier. Default value is 1.
 invfloat, optional
If True, it performs inverse logarithmic correction, else correction will be logarithmic. Defaults to False.
 Returns:
 outndarray
Logarithm corrected output image.
See also
References
 cucim.skimage.exposure.adjust_sigmoid(image, cutoff=0.5, gain=10, inv=False)#
Performs Sigmoid Correction on the input image.
Also known as Contrast Adjustment. This function transforms the input image pixelwise according to the equation
O = 1/(1 + exp*(gain*(cutoff  I)))
after scaling each pixel to the range 0 to 1. Parameters:
 imagendarray
Input image.
 cutofffloat, optional
Cutoff of the sigmoid function that shifts the characteristic curve in horizontal direction. Default value is 0.5.
 gainfloat, optional
The constant multiplier in exponential’s power of sigmoid function. Default value is 10.
 invbool, optional
If True, returns the negative sigmoid correction. Defaults to False.
 Returns:
 outndarray
Sigmoid corrected output image.
See also
References
[1]Gustav J. Braun, “Image Lightness Rescaling Using Sigmoidal Contrast Enhancement Functions”, http://markfairchild.org/PDFs/PAP07.pdf
 cucim.skimage.exposure.cumulative_distribution(image, nbins=256)#
Return cumulative distribution function (cdf) for the given image.
 Parameters:
 imagearray
Image array.
 nbinsint, optional
Number of bins for image histogram.
 Returns:
 img_cdfarray
Values of cumulative distribution function.
 bin_centersarray
Centers of bins.
See also
References
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage import exposure, img_as_float >>> image = img_as_float(cp.array(data.camera())) >>> hi = exposure.histogram(image) >>> cdf = exposure.cumulative_distribution(image) >>> cp.alltrue(cdf[0] == cp.cumsum(hi[0])/float(image.size)) array(True)
 cucim.skimage.exposure.equalize_adapthist(image, kernel_size=None, clip_limit=0.01, nbins=256)#
Contrast Limited Adaptive Histogram Equalization (CLAHE).
An algorithm for local contrast enhancement, that uses histograms computed over different tile regions of the image. Local details can therefore be enhanced even in regions that are darker or lighter than most of the image.
 Parameters:
 image(N1, …,NN[, C]) ndarray
Input image.
 kernel_sizeint or array_like, optional
Defines the shape of contextual regions used in the algorithm. If iterable is passed, it must have the same number of elements as
image.ndim
(without color channel). If integer, it is broadcasted to each image dimension. By default,kernel_size
is 1/8 ofimage
height by 1/8 of its width. clip_limitfloat, optional
Clipping limit, normalized between 0 and 1 (higher values give more contrast).
 nbinsint, optional
Number of gray bins for histogram (“data range”).
 Returns:
 out(N1, …,NN[, C]) ndarray
Equalized image with float64 dtype.
See also
Notes
 For color images, the following steps are performed:
The image is converted to HSV color space
The CLAHE algorithm is run on the V (Value) channel
The image is converted back to RGB space and returned
For RGBA images, the original alpha channel is removed.
Changed in version 0.17: The values returned by this function are slightly shifted upwards because of an internal change in rounding behavior.
References
 cucim.skimage.exposure.equalize_hist(image, nbins=256, mask=None)#
Return image after histogram equalization.
 Parameters:
 imagearray
Image array.
 nbinsint, optional
Number of bins for image histogram. Note: this argument is ignored for integer images, for which each integer is its own bin.
 mask: ndarray of bools or 0s and 1s, optional
Array of same shape as image. Only points at which mask == True are used for the equalization, which is applied to the whole image.
 Returns:
 outfloat array
Image array after histogram equalization.
Notes
This function is adapted from [1] with the author’s permission.
References
 cucim.skimage.exposure.histogram(image, nbins=256, source_range='image', normalize=False, *, channel_axis=None)#
Return histogram of image.
Unlike numpy.histogram, this function returns the centers of bins and does not rebin integer arrays. For integer arrays, each integer value has its own bin, which improves speed and intensityresolution.
If channel_axis is not set, the histogram is computed on the flattened image. For color or multichannel images, set
channel_axis
to use a common binning for all channels. Alternatively, one may apply the function separately on each channel to obtain a histogram for each color channel with separate binning. Parameters:
 imagearray
Input image.
 nbinsint, optional
Number of bins used to calculate histogram. This value is ignored for integer arrays.
 source_rangestring, optional
‘image’ (default) determines the range from the input image. ‘dtype’ determines the range from the expected range of the images of that data type.
 normalizebool, optional
If True, normalize the histogram by the sum of its values.
 channel_axisint or None, optional
If None, the image is assumed to be a grayscale (single channel) image. Otherwise, this parameter indicates which axis of the array corresponds to channels.
 Returns:
 histarray
The values of the histogram. When
channel_axis
is not None, hist will be a 2D array where the first axis corresponds to channels. bin_centersarray
The values at the center of the bins.
See also
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage import exposure, img_as_float >>> image = img_as_float(cp.array(data.camera())) >>> cp.histogram(image, bins=2) (array([ 93585, 168559]), array([0. , 0.5, 1. ])) >>> exposure.histogram(image, nbins=2) (array([ 93585, 168559]), array([0.25, 0.75]))
 cucim.skimage.exposure.is_low_contrast(image, fraction_threshold=0.05, lower_percentile=1, upper_percentile=99, method='linear')#
Determine if an image is low contrast.
 Parameters:
 imagearraylike
The image under test.
 fraction_thresholdfloat, optional
The low contrast fraction threshold. An image is considered low contrast when its range of brightness spans less than this fraction of its data type’s full range. [1]
 lower_percentilefloat, optional
Disregard values below this percentile when computing image contrast.
 upper_percentilefloat, optional
Disregard values above this percentile when computing image contrast.
 methodstr, optional
The contrast determination method. Right now the only available option is “linear”.
 Returns:
 outbool
True when the image is determined to be low contrast.
Notes
For boolean images, this function returns False only if all values are the same (the method, threshold, and percentile arguments are ignored).
References
Examples
>>> import cupy as cp >>> image = cp.linspace(0, 0.04, 100) >>> is_low_contrast(image) array(True) >>> image[1] = 1 >>> is_low_contrast(image) array(True) >>> is_low_contrast(image, upper_percentile=100) array(False)
 cucim.skimage.exposure.match_histograms(image, reference, *, channel_axis=None)#
Adjust an image so that its cumulative histogram matches that of another.
The adjustment is applied separately for each channel.
 Parameters:
 imagendarray
Input image. Can be grayscale or in color.
 referencendarray
Image to match histogram of. Must have the same number of channels as image.
 channel_axisint or None, optional
If None, the image is assumed to be a grayscale (single channel) image. Otherwise, this parameter indicates which axis of the array corresponds to channels.
 Returns:
 matchedndarray
Transformed input image.
 Raises:
 ValueError
Thrown when the number of channels in the input image and the reference differ.
References
 cucim.skimage.exposure.rescale_intensity(image, in_range='image', out_range='dtype')#
Return image after stretching or shrinking its intensity levels.
The desired intensity range of the input and output, in_range and out_range respectively, are used to stretch or shrink the intensity range of the input image. See examples below.
 Parameters:
 imagearray
Image array.
 in_range, out_rangestr or 2tuple, optional
Min and max intensity values of input and output image. The possible values for this parameter are enumerated below.
 ‘image’
Use image min/max as the intensity range.
 ‘dtype’
Use min/max of the image’s dtype as the intensity range.
 dtypename
Use intensity range based on desired dtype. Must be valid key in DTYPE_RANGE.
 2tuple
Use range_values as explicit min/max intensities.
 Returns:
 outarray
Image array after rescaling its intensity. This image is the same dtype as the input image.
See also
Notes
Changed in version 0.17: The dtype of the output array has changed to match the output dtype, or float if the output range is specified by a pair of values.
Examples
By default, the min/max intensities of the input image are stretched to the limits allowed by the image’s dtype, since in_range defaults to ‘image’ and out_range defaults to ‘dtype’:
>>> image = cp.array([51, 102, 153], dtype=np.uint8) >>> rescale_intensity(image) array([ 0, 127, 255], dtype=uint8)
It’s easy to accidentally convert an image dtype from uint8 to float:
>>> 1.0 * image array([ 51., 102., 153.])
Use rescale_intensity to rescale to the proper range for float dtypes:
>>> image_float = 1.0 * image >>> rescale_intensity(image_float) array([0. , 0.5, 1. ])
To maintain the low contrast of the original, use the in_range parameter:
>>> rescale_intensity(image_float, in_range=(0, 255)) array([0.2, 0.4, 0.6])
If the min/max value of in_range is more/less than the min/max image intensity, then the intensity levels are clipped:
>>> rescale_intensity(image_float, in_range=(0, 102)) array([0.5, 1. , 1. ])
If you have an image with signed integers but want to rescale the image to just the positive range, use the out_range parameter. In that case, the output dtype will be float:
>>> image = cp.asarray([10, 0, 10], dtype=np.int8) >>> rescale_intensity(image, out_range=(0, 127)) array([ 0. , 63.5, 127. ])
To get the desired range with a specific dtype, use
.astype()
:>>> rescale_intensity(image, out_range=(0, 127)).astype(np.int8) array([ 0, 63, 127], dtype=int8)
If the input image is constant, the output will be clipped directly to the output range: >>> image = cp.asarray([130, 130, 130], dtype=np.int32) >>> rescale_intensity(image, out_range=(0, 127)).astype(np.int32) array([127, 127, 127], dtype=int32)
feature#
 cucim.skimage.feature.blob_dog(image, min_sigma=1, max_sigma=50, sigma_ratio=1.6, threshold=0.5, overlap=0.5, *, threshold_rel=None, exclude_border=False)#
Finds blobs in the given grayscale image. Blobs are found using the Difference of Gaussian (DoG) method [1], [2]. For each blob found, the method returns its coordinates and the standard deviation of the Gaussian kernel that detected the blob. Parameters ——— image : ndarray
Input grayscale image, blobs are assumed to be light on dark background (white on black).
 min_sigmascalar or sequence of scalars, optional
The minimum standard deviation for Gaussian kernel. Keep this low to detect smaller blobs. The standard deviations of the Gaussian filter are given for each axis as a sequence, or as a single number, in which case it is equal for all axes.
 max_sigmascalar or sequence of scalars, optional
The maximum standard deviation for Gaussian kernel. Keep this high to detect larger blobs. The standard deviations of the Gaussian filter are given for each axis as a sequence, or as a single number, in which case it is equal for all axes.
 sigma_ratiofloat, optional
The ratio between the standard deviation of Gaussian Kernels used for computing the Difference of Gaussians
 thresholdfloat or None, optional
The absolute lower bound for scale space maxima. Local maxima smaller than threshold are ignored. Reduce this to detect blobs with lower intensities. If threshold_rel is also specified, whichever threshold is larger will be used. If None, threshold_rel is used instead.
 overlapfloat, optional
A value between 0 and 1. If the area of two blobs overlaps by a fraction greater than threshold, the smaller blob is eliminated.
 threshold_relfloat or None, optional
Minimum intensity of peaks, calculated as
max(dog_space) * threshold_rel
, wheredog_space
refers to the stack of DifferenceofGaussian (DoG) images computed internally. This should have a value between 0 and 1. If None, threshold is used instead. exclude_bordertuple of ints, int, or False, optional
If tuple of ints, the length of the tuple must match the input array’s dimensionality. Each element of the tuple will exclude peaks from within exclude_borderpixels of the border of the image along that dimension. If nonzero int, exclude_border excludes peaks from within exclude_borderpixels of the border of the image. If zero or False, peaks are identified regardless of their distance from the border.
 Returns:
 A(n, image.ndim + sigma) ndarray
A 2d array with each row representing 2 coordinate values for a 2D image, or 3 coordinate values for a 3D image, plus the sigma(s) used. When a single sigma is passed, outputs are:
(r, c, sigma)
or(p, r, c, sigma)
where(r, c)
or(p, r, c)
are coordinates of the blob andsigma
is the standard deviation of the Gaussian kernel which detected the blob. When an anisotropic gaussian is used (sigmas per dimension), the detected sigma is returned for each dimension.
Notes
The radius of each blob is approximately \(\sqrt{2}\sigma\) for a 2D image and \(\sqrt{3}\sigma\) for a 3D image.
References
[2]Lowe, D. G. “Distinctive Image Features from ScaleInvariant Keypoints.” International Journal of Computer Vision 60, 91–110 (2004). https://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf DOI:10.1023/B:VISI.0000029664.99615.94
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage import feature >>> coins = cp.array(data.coins()) >>> feature.blob_dog(coins, threshold=.05, min_sigma=10, max_sigma=40) array([[128., 155., 10.], [198., 155., 10.], [124., 338., 10.], [127., 102., 10.], [193., 281., 10.], [126., 208., 10.], [267., 115., 10.], [197., 102., 10.], [198., 215., 10.], [123., 279., 10.], [126., 46., 10.], [259., 247., 10.], [196., 43., 10.], [ 54., 276., 10.], [267., 358., 10.], [ 58., 100., 10.], [259., 305., 10.], [185., 347., 16.], [261., 174., 16.], [ 46., 336., 16.], [ 54., 217., 10.], [ 55., 157., 10.], [ 57., 41., 10.], [260., 47., 16.]])
 cucim.skimage.feature.blob_doh(image, min_sigma=1, max_sigma=30, num_sigma=10, threshold=0.01, overlap=0.5, log_scale=False, *, threshold_rel=None)#
Finds blobs in the given grayscale image.
Blobs are found using the Determinant of Hessian method [1]. For each blob found, the method returns its coordinates and the standard deviation of the Gaussian Kernel used for the Hessian matrix whose determinant detected the blob. Determinant of Hessians is approximated using [2].
 Parameters:
 image2D ndarray
Input grayscale image.Blobs can either be light on dark or vice versa.
 min_sigmafloat, optional
The minimum standard deviation for Gaussian Kernel used to compute Hessian matrix. Keep this low to detect smaller blobs.
 max_sigmafloat, optional
The maximum standard deviation for Gaussian Kernel used to compute Hessian matrix. Keep this high to detect larger blobs.
 num_sigmaint, optional
The number of intermediate values of standard deviations to consider between min_sigma and max_sigma.
 thresholdfloat or None, optional
The absolute lower bound for scale space maxima. Local maxima smaller than threshold are ignored. Reduce this to detect blobs with lower intensities. If threshold_rel is also specified, whichever threshold is larger will be used. If None, threshold_rel is used instead.
 overlapfloat, optional
A value between 0 and 1. If the area of two blobs overlaps by a fraction greater than threshold, the smaller blob is eliminated.
 log_scalebool, optional
If set intermediate values of standard deviations are interpolated using a logarithmic scale to the base 10. If not, linear interpolation is used.
 threshold_relfloat or None, optional
Minimum intensity of peaks, calculated as
max(doh_space) * threshold_rel
, wheredoh_space
refers to the stack of DeterminantofHessian (DoH) images computed internally. This should have a value between 0 and 1. If None, threshold is used instead.
 Returns:
 A(n, 3) ndarray
A 2d array with each row representing 3 values,
(y,x,sigma)
where(y,x)
are coordinates of the blob andsigma
is the standard deviation of the Gaussian kernel of the Hessian Matrix whose determinant detected the blob.
Notes
The radius of each blob is approximately sigma. Computation of Determinant of Hessians is independent of the standard deviation. Therefore detecting larger blobs won’t take more time. In methods line
blob_dog()
andblob_log()
the computation of Gaussians for larger sigma takes more time. The downside is that this method can’t be used for detecting blobs of radius less than 3px due to the box filters used in the approximation of Hessian Determinant.References
[2]Herbert Bay, Andreas Ess, Tinne Tuytelaars, Luc Van Gool, “SURF: Speeded Up Robust Features” ftp://ftp.vision.ee.ethz.ch/publications/articles/eth_biwi_00517.pdf
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage import feature >>> img = cp.array(data.coins()) >>> feature.blob_doh(img) array([[197. , 153. , 20.333334], [124. , 336. , 20.333334], [126. , 153. , 20.333334], [195. , 100. , 23.555555], [192. , 212. , 23.555555], [121. , 271. , 30. ], [126. , 101. , 20.333334], [193. , 275. , 23.555555], [123. , 205. , 20.333334], [270. , 363. , 30. ], [265. , 113. , 23.555555], [262. , 243. , 23.555555], [185. , 348. , 30. ], [156. , 302. , 30. ], [123. , 44. , 23.555555], [260. , 173. , 30. ], [197. , 44. , 20.333334]], dtype=float32)
 cucim.skimage.feature.blob_log(image, min_sigma=1, max_sigma=50, num_sigma=10, threshold=0.2, overlap=0.5, log_scale=False, *, threshold_rel=None, exclude_border=False)#
Finds blobs in the given grayscale image. Blobs are found using the Laplacian of Gaussian (LoG) method [1]. For each blob found, the method returns its coordinates and the standard deviation of the Gaussian kernel that detected the blob. Parameters ——— image : ndarray
Input grayscale image, blobs are assumed to be light on dark background (white on black).
 min_sigmascalar or sequence of scalars, optional
the minimum standard deviation for Gaussian kernel. Keep this low to detect smaller blobs. The standard deviations of the Gaussian filter are given for each axis as a sequence, or as a single number, in which case it is equal for all axes.
 max_sigmascalar or sequence of scalars, optional
The maximum standard deviation for Gaussian kernel. Keep this high to detect larger blobs. The standard deviations of the Gaussian filter are given for each axis as a sequence, or as a single number, in which case it is equal for all axes.
 num_sigmaint, optional
The number of intermediate values of standard deviations to consider between min_sigma and max_sigma.
 thresholdfloat or None, optional
The absolute lower bound for scale space maxima. Local maxima smaller than threshold are ignored. Reduce this to detect blobs with lower intensities. If threshold_rel is also specified, whichever threshold is larger will be used. If None, threshold_rel is used instead.
 overlapfloat, optional
A value between 0 and 1. If the area of two blobs overlaps by a fraction greater than threshold, the smaller blob is eliminated.
 log_scalebool, optional
If set intermediate values of standard deviations are interpolated using a logarithmic scale to the base 10. If not, linear interpolation is used.
 threshold_relfloat or None, optional
Minimum intensity of peaks, calculated as
max(log_space) * threshold_rel
, wherelog_space
refers to the stack of LaplacianofGaussian (LoG) images computed internally. This should have a value between 0 and 1. If None, threshold is used instead. exclude_bordertuple of ints, int, or False, optional
If tuple of ints, the length of the tuple must match the input array’s dimensionality. Each element of the tuple will exclude peaks from within exclude_borderpixels of the border of the image along that dimension. If nonzero int, exclude_border excludes peaks from within exclude_borderpixels of the border of the image. If zero or False, peaks are identified regardless of their distance from the border.
 Returns:
 A(n, image.ndim + sigma) ndarray
A 2d array with each row representing 2 coordinate values for a 2D image, or 3 coordinate values for a 3D image, plus the sigma(s) used. When a single sigma is passed, outputs are:
(r, c, sigma)
or(p, r, c, sigma)
where(r, c)
or(p, r, c)
are coordinates of the blob andsigma
is the standard deviation of the Gaussian kernel which detected the blob. When an anisotropic gaussian is used (sigmas per dimension), the detected sigma is returned for each dimension.
References
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage import feature, exposure >>> img = cp.array(data.coins()) >>> img = exposure.equalize_hist(img) # improves detection >>> feature.blob_log(img, threshold = .3) array([[124. , 336. , 11.88888889], [198. , 155. , 11.88888889], [194. , 213. , 17.33333333], [121. , 272. , 17.33333333], [263. , 244. , 17.33333333], [194. , 276. , 17.33333333], [266. , 115. , 11.88888889], [128. , 154. , 11.88888889], [260. , 174. , 17.33333333], [198. , 103. , 11.88888889], [126. , 208. , 11.88888889], [127. , 102. , 11.88888889], [263. , 302. , 17.33333333], [197. , 44. , 11.88888889], [185. , 344. , 17.33333333], [126. , 46. , 11.88888889], [113. , 323. , 1. ]]) Notes  The radius of each blob is approximately :math:`\sqrt{2}\sigma` for a 2D image and :math:`\sqrt{3}\sigma` for a 3D image.
 cucim.skimage.feature.canny(image, sigma=1.0, low_threshold=None, high_threshold=None, mask=None, use_quantiles=False, *, mode='constant', cval=0.0)#
Edge filter an image using the Canny algorithm.
 Parameters:
 image2D array
Grayscale input image to detect edges on; can be of any dtype.
 sigmafloat, optional
Standard deviation of the Gaussian filter.
 low_thresholdfloat, optional
Lower bound for hysteresis thresholding (linking edges). If None, low_threshold is set to 10% of dtype’s max.
 high_thresholdfloat, optional
Upper bound for hysteresis thresholding (linking edges). If None, high_threshold is set to 20% of dtype’s max.
 maskarray, dtype=bool, optional
Mask to limit the application of Canny to a certain area.
 use_quantilesbool, optional
If
True
then treat low_threshold and high_threshold as quantiles of the edge magnitude image, rather than absolute edge magnitude values. IfTrue
then the thresholds must be in the range [0, 1]. modestr, {‘reflect’, ‘constant’, ‘nearest’, ‘mirror’, ‘wrap’}
The
mode
parameter determines how the array borders are handled during Gaussian filtering, wherecval
is the value when mode is equal to ‘constant’. cvalfloat, optional
Value to fill past edges of input if mode is ‘constant’.
 Returns:
 output2D array (image)
The binary edge map.
See also
Notes
The steps of the algorithm are as follows:
Smooth the image using a Gaussian with
sigma
width.Apply the horizontal and vertical Sobel operators to get the gradients within the image. The edge strength is the norm of the gradient.
Thin potential edges to 1pixel wide curves. First, find the normal to the edge at each point. This is done by looking at the signs and the relative magnitude of the XSobel and YSobel to sort the points into 4 categories: horizontal, vertical, diagonal and antidiagonal. Then look in the normal and reverse directions to see if the values in either of those directions are greater than the point in question. Use interpolation to get a mix of points instead of picking the one that’s the closest to the normal.
Perform a hysteresis thresholding: first label all points above the high threshold as edges. Then recursively label any point above the low threshold that is 8connected to a labeled point as an edge.
References
[1]Canny, J., A Computational Approach To Edge Detection, IEEE Trans. Pattern Analysis and Machine Intelligence, 8:679714, 1986 DOI:10.1109/TPAMI.1986.4767851
[2]William Green’s Canny tutorial https://en.wikipedia.org/wiki/Canny_edge_detector
Examples
>>> import cupy as cp >>> from cucim.skimage import feature >>> # Generate noisy image of a square >>> im = cp.zeros((256, 256)) >>> im[64:64, 64:64] = 1 >>> im += 0.2 * cp.random.rand(*im.shape) >>> # First trial with the Canny filter, with the default smoothing >>> edges1 = feature.canny(im) >>> # Increase the smoothing for better results >>> edges2 = feature.canny(im, sigma=3)
 cucim.skimage.feature.corner_foerstner(image, sigma=1)#
Compute Foerstner corner measure response image.
This corner detector uses information from the autocorrelation matrix A:
A = [(imx**2) (imx*imy)] = [Axx Axy] [(imx*imy) (imy**2)] [Axy Ayy]
Where imx and imy are first derivatives, averaged with a gaussian filter. The corner measure is then defined as:
w = det(A) / trace(A) (size of error ellipse) q = 4 * det(A) / trace(A)**2 (roundness of error ellipse)
 Parameters:
 image(M, N) ndarray
Input image.
 sigmafloat, optional
Standard deviation used for the Gaussian kernel, which is used as weighting function for the autocorrelation matrix.
 Returns:
 wndarray
Error ellipse sizes.
 qndarray
Roundness of error ellipse.
References
[1]Förstner, W., & Gülch, E. (1987, June). A fast operator for detection and precise location of distinct points, corners and centres of circular features. In Proc. ISPRS intercommission conference on fast processing of photogrammetric data (pp. 281305).
Examples
>>> from cucim.skimage.feature import corner_foerstner, corner_peaks >>> square = cp.zeros([10, 10]) >>> square[2:8, 2:8] = 1 >>> square.astype(int) array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]) >>> w, q = corner_foerstner(square) >>> accuracy_thresh = 0.5 >>> roundness_thresh = 0.3 >>> foerstner = (q > roundness_thresh) * (w > accuracy_thresh) * w >>> corner_peaks(foerstner, min_distance=1) array([[2, 2], [2, 7], [7, 2], [7, 7]])
 cucim.skimage.feature.corner_harris(image, method='k', k=0.05, eps=1e06, sigma=1)#
Compute Harris corner measure response image.
This corner detector uses information from the autocorrelation matrix A:
A = [(imx**2) (imx*imy)] = [Axx Axy] [(imx*imy) (imy**2)] [Axy Ayy]
Where imx and imy are first derivatives, averaged with a gaussian filter. The corner measure is then defined as:
det(A)  k * trace(A)**2
or:
2 * det(A) / (trace(A) + eps)
 Parameters:
 image(M, N) ndarray
Input image.
 method{‘k’, ‘eps’}, optional
Method to compute the response image from the autocorrelation matrix.
 kfloat, optional
Sensitivity factor to separate corners from edges, typically in range [0, 0.2]. Small values of k result in detection of sharp corners.
 epsfloat, optional
Normalisation factor (Noble’s corner measure).
 sigmafloat, optional
Standard deviation used for the Gaussian kernel, which is used as weighting function for the autocorrelation matrix.
 Returns:
 responsendarray
Harris response image.
References
Examples
>>> from cucim.skimage.feature import corner_harris, corner_peaks >>> square = cp.zeros([10, 10]) >>> square[2:8, 2:8] = 1 >>> square.astype(int) array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]) >>> corner_peaks(corner_harris(square), min_distance=1) array([[2, 2], [2, 7], [7, 2], [7, 7]])
 cucim.skimage.feature.corner_kitchen_rosenfeld(image, mode='constant', cval=0)#
Compute Kitchen and Rosenfeld corner measure response image.
The corner measure is calculated as follows:
(imxx * imy**2 + imyy * imx**2  2 * imxy * imx * imy) / (imx**2 + imy**2)
Where imx and imy are the first and imxx, imxy, imyy the second derivatives.
 Parameters:
 image(M, N) ndarray
Input image.
 mode{‘constant’, ‘reflect’, ‘wrap’, ‘nearest’, ‘mirror’}, optional
How to handle values outside the image borders.
 cvalfloat, optional
Used in conjunction with mode ‘constant’, the value outside the image boundaries.
 Returns:
 responsendarray
Kitchen and Rosenfeld response image.
References
[1]Kitchen, L., & Rosenfeld, A. (1982). Graylevel corner detection. Pattern recognition letters, 1(2), 95102. DOI:10.1016/01678655(82)900204
 cucim.skimage.feature.corner_peaks(image, min_distance=1, threshold_abs=None, threshold_rel=None, exclude_border=True, indices=True, num_peaks=inf, footprint=None, labels=None, *, num_peaks_per_label=inf, p_norm=inf)#
Find peaks in corner measure response image.
This differs from skimage.feature.peak_local_max in that it suppresses multiple connected peaks with the same accumulator value.
 Parameters:
 image(M, N) ndarray
Input image.
 min_distanceint, optional
The minimal allowed distance separating peaks.
 **
See
skimage.feature.peak_local_max()
. p_normfloat
Which Minkowski pnorm to use. Should be in the range [1, inf]. A finite large p may cause a ValueError if overflow can occur.
inf
corresponds to the Chebyshev distance and 2 to the Euclidean distance.
 Returns:
 outputndarray or ndarray of bools
If indices = True : (row, column, …) coordinates of peaks.
If indices = False : Boolean array shaped like image, with peaks represented by True values.
See also
Notes
Changed in version 0.18: The default value of threshold_rel has changed to None, which corresponds to letting skimage.feature.peak_local_max decide on the default. This is equivalent to threshold_rel=0.
The num_peaks limit is applied before suppression of connected peaks. To limit the number of peaks after suppression, set num_peaks=np.inf and postprocess the output of this function.
Examples
>>> from cucim.skimage.feature import peak_local_max >>> response = cp.zeros((5, 5)) >>> response[2:4, 2:4] = 1 >>> response array([[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.], [0., 0., 1., 1., 0.], [0., 0., 1., 1., 0.], [0., 0., 0., 0., 0.]]) >>> peak_local_max(response) array([[2, 2], [2, 3], [3, 2], [3, 3]]) >>> corner_peaks(response) array([[2, 2]])
 cucim.skimage.feature.corner_shi_tomasi(image, sigma=1)#
Compute ShiTomasi (KanadeTomasi) corner measure response image.
This corner detector uses information from the autocorrelation matrix A:
A = [(imx**2) (imx*imy)] = [Axx Axy] [(imx*imy) (imy**2)] [Axy Ayy]
Where imx and imy are first derivatives, averaged with a gaussian filter. The corner measure is then defined as the smaller eigenvalue of A:
((Axx + Ayy)  sqrt((Axx  Ayy)**2 + 4 * Axy**2)) / 2
 Parameters:
 image(M, N) ndarray
Input image.
 sigmafloat, optional
Standard deviation used for the Gaussian kernel, which is used as weighting function for the autocorrelation matrix.
 Returns:
 responsendarray
ShiTomasi response image.
References
Examples
>>> from cucim.skimage.feature import corner_shi_tomasi, corner_peaks >>> square = cp.zeros([10, 10]) >>> square[2:8, 2:8] = 1 >>> square.astype(int) array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]) >>> corner_peaks(corner_shi_tomasi(square), ... min_distance=1) array([[2, 2], [2, 7], [7, 2], [7, 7]])
 cucim.skimage.feature.daisy(image, step=4, radius=15, rings=3, histograms=8, orientations=8, normalization='l1', sigmas=None, ring_radii=None, visualize=False)#
Extract DAISY feature descriptors densely for the given image.
DAISY is a feature descriptor similar to SIFT formulated in a way that allows for fast dense extraction. Typically, this is practical for bagoffeatures image representations.
The implementation follows Tola et al. [1] but deviate on the following points:
Histogram bin contribution are smoothed with a circular Gaussian window over the tonal range (the angular range).
The sigma values of the spatial Gaussian smoothing in this code do not match the sigma values in the original code by Tola et al. [2]. In their code, spatial smoothing is applied to both the input image and the center histogram. However, this smoothing is not documented in [1] and, therefore, it is omitted.
 Parameters:
 image(M, N) array
Input image (grayscale).
 stepint, optional
Distance between descriptor sampling points.
 radiusint, optional
Radius (in pixels) of the outermost ring.
 ringsint, optional
Number of rings.
 histogramsint, optional
Number of histograms sampled per ring.
 orientationsint, optional
Number of orientations (bins) per histogram.
 normalization[ ‘l1’  ‘l2’  ‘daisy’  ‘off’ ], optional
How to normalize the descriptors
‘l1’: L1normalization of each descriptor.
‘l2’: L2normalization of each descriptor.
‘daisy’: L2normalization of individual histograms.
‘off’: Disable normalization.
 sigmas1D array of float, optional
Standard deviation of spatial Gaussian smoothing for the center histogram and for each ring of histograms. The array of sigmas should be sorted from the center and out. I.e. the first sigma value defines the spatial smoothing of the center histogram and the last sigma value defines the spatial smoothing of the outermost ring. Specifying sigmas overrides the following parameter.
rings = len(sigmas)  1
 ring_radii1D array of int, optional
Radius (in pixels) for each ring. Specifying ring_radii overrides the following two parameters.
rings = len(ring_radii)
radius = ring_radii[1]
If both sigmas and ring_radii are given, they must satisfy the following predicate since no radius is needed for the center histogram.
len(ring_radii) == len(sigmas) + 1
 visualizebool, optional
Generate a visualization of the DAISY descriptors
 Returns:
 descsarray
Grid of DAISY descriptors for the given image as an array dimensionality (P, Q, R) where
P = ceil((M  radius*2) / step)
Q = ceil((N  radius*2) / step)
R = (rings * histograms + 1) * orientations
 descs_img(M, N, 3) array (only if visualize==True)
Visualization of the DAISY descriptors.
References
 cucim.skimage.feature.hessian_matrix(image, sigma=1, mode='constant', cval=0, order='rc', use_gaussian_derivatives=None)#
Compute the Hessian matrix.
In 2D, the Hessian matrix is defined as:
H = [Hrr Hrc] [Hrc Hcc]
which is computed by convolving the image with the second derivatives of the Gaussian kernel in the respective r and cdirections.
The implementation here also supports ndimensional data.
 Parameters:
 imagendarray
Input image.
 sigmafloat
Standard deviation used for the Gaussian kernel, which is used as weighting function for the autocorrelation matrix.
 mode{‘constant’, ‘reflect’, ‘wrap’, ‘nearest’, ‘mirror’}, optional
How to handle values outside the image borders.
 cvalfloat, optional
Used in conjunction with mode ‘constant’, the value outside the image boundaries.
 order{‘rc’, ‘xy’}, optional
NOTE: ‘xy’ is only an option for 2D images, higher dimensions must always use ‘rc’ order. This parameter allows for the use of reverse or forward order of the image axes in gradient computation. ‘rc’ indicates the use of the first axis initially (Hrr, Hrc, Hcc), whilst ‘xy’ indicates the usage of the last axis initially (Hxx, Hxy, Hyy).
 use_gaussian_derivativesboolean, optional
Indicates whether the Hessian is computed by convolving with Gaussian derivatives, or by a simple finitedifference operation.
 Returns:
 H_elemslist of ndarray
Upperdiagonal elements of the hessian matrix for each pixel in the input image. In 2D, this will be a three element list containing [Hrr, Hrc, Hcc]. In nD, the list will contain
(n**2 + n) / 2
arrays.
Notes
The distributive property of derivatives and convolutions allows us to restate the derivative of an image, I, smoothed with a Gaussian kernel, G, as the convolution of the image with the derivative of G.
\[\frac{\partial }{\partial x_i}(I * G) = I * \left( \frac{\partial }{\partial x_i} G \right)\]When
use_gaussian_derivatives
isTrue
, this property is used to compute the second order derivatives that make up the Hessian matrix.When
use_gaussian_derivatives
isFalse
, simple finite differences on a Gaussiansmoothed image are used instead.Examples
>>> import cupy as cp >>> from cucim.skimage.feature import hessian_matrix >>> square = cp.zeros((5, 5)) >>> square[2, 2] = 4 >>> Hrr, Hrc, Hcc = hessian_matrix(square, sigma=0.1, order='rc', ... use_gaussian_derivatives=False) >>> Hrc array([[ 0., 0., 0., 0., 0.], [ 0., 1., 0., 1., 0.], [ 0., 0., 0., 0., 0.], [ 0., 1., 0., 1., 0.], [ 0., 0., 0., 0., 0.]])
 cucim.skimage.feature.hessian_matrix_det(image, sigma=1, approximate=True)#
Compute the approximate Hessian Determinant over an image.
The 2D approximate method uses box filters over integral images to compute the approximate Hessian Determinant.
 Parameters:
 imagendarray
The image over which to compute the Hessian Determinant.
 sigmafloat, optional
Standard deviation of the Gaussian kernel used for the Hessian matrix.
 approximatebool, optional
If
True
and the image is 2D, use a much faster approximate computation. This argument has no effect on 3D and higher images.
 Returns:
 outarray
The array of the Determinant of Hessians.
Notes
For 2D images when
approximate=True
, the running time of this method only depends on size of the image. It is independent of sigma as one would expect. The downside is that the result for sigma less than 3 is not accurate, i.e., not similar to the result obtained if someone computed the Hessian and took its determinant.References
[1]Herbert Bay, Andreas Ess, Tinne Tuytelaars, Luc Van Gool, “SURF: Speeded Up Robust Features” ftp://ftp.vision.ee.ethz.ch/publications/articles/eth_biwi_00517.pdf
 cucim.skimage.feature.hessian_matrix_eigvals(H_elems)#
Compute eigenvalues of Hessian matrix.
 Parameters:
 H_elemslist of ndarray
The upperdiagonal elements of the Hessian matrix, as returned by hessian_matrix.
 Returns:
 eigsndarray
The eigenvalues of the Hessian matrix, in decreasing order. The eigenvalues are the leading dimension. That is,
eigs[i, j, k]
contains the ithlargest eigenvalue at position (j, k).
Examples
>>> import cupy as cp >>> from cucim.skimage.feature import (hessian_matrix, ... hessian_matrix_eigvals) >>> square = cp.zeros((5, 5)) >>> square[2, 2] = 4 >>> H_elems = hessian_matrix(square, sigma=0.1, order='rc', ... use_gaussian_derivatives=False) >>> hessian_matrix_eigvals(H_elems)[0] array([[ 0., 0., 2., 0., 0.], [ 0., 1., 0., 1., 0.], [ 2., 0., 2., 0., 2.], [ 0., 1., 0., 1., 0.], [ 0., 0., 2., 0., 0.]])
 cucim.skimage.feature.match_descriptors(descriptors1, descriptors2, metric=None, p=2, max_distance=inf, cross_check=True, max_ratio=1.0)#
Bruteforce matching of descriptors.
For each descriptor in the first set this matcher finds the closest descriptor in the second set (and viceversa in the case of enabled crosschecking).
 Parameters:
 descriptors1(M, P) array
Descriptors of size P about M keypoints in the first image.
 descriptors2(N, P) array
Descriptors of size P about N keypoints in the second image.
 metric{‘euclidean’, ‘cityblock’, ‘minkowski’, ‘hamming’, …} , optional
The metric to compute the distance between two descriptors. See scipy.spatial.distance.cdist for all possible types. The hamming distance should be used for binary descriptors. By default the L2norm is used for all descriptors of dtype float or double and the Hamming distance is used for binary descriptors automatically.
 pint, optional
The pnorm to apply for
metric='minkowski'
. max_distancefloat, optional
Maximum allowed distance between descriptors of two keypoints in separate images to be regarded as a match.
 cross_checkbool, optional
If True, the matched keypoints are returned after cross checking i.e. a matched pair (keypoint1, keypoint2) is returned if keypoint2 is the best match for keypoint1 in second image and keypoint1 is the best match for keypoint2 in first image.
 max_ratiofloat, optional
Maximum ratio of distances between first and second closest descriptor in the second set of descriptors. This threshold is useful to filter ambiguous matches between the two descriptor sets. The choice of this value depends on the statistics of the chosen descriptor, e.g., for SIFT descriptors a value of 0.8 is usually chosen, see D.G. Lowe, “Distinctive Image Features from ScaleInvariant Keypoints”, International Journal of Computer Vision, 2004.
 Returns:
 matches(Q, 2) array
Indices of corresponding matches in first and second set of descriptors, where
matches[:, 0]
denote the indices in the first andmatches[:, 1]
the indices in the second set of descriptors.
 cucim.skimage.feature.match_template(image, template, pad_input=False, mode='constant', constant_values=0)#
Match a template to a 2D or 3D image using normalized correlation.
The output is an array with values between 1.0 and 1.0. The value at a given position corresponds to the correlation coefficient between the image and the template.
For pad_input=True matches correspond to the center and otherwise to the topleft corner of the template. To find the best match you must search for peaks in the response (output) image.
 Parameters:
 image(M, N[, D]) array
2D or 3D input image.
 template(m, n[, d]) array
Template to locate. It must be (m <= M, n <= N[, d <= D]).
 pad_inputbool
If True, pad image so that output is the same size as the image, and output values correspond to the template center. Otherwise, the output is an array with shape (M  m + 1, N  n + 1) for an (M, N) image and an (m, n) template, and matches correspond to origin (topleft corner) of the template.
 modesee numpy.pad, optional
Padding mode.
 constant_valuessee numpy.pad, optional
Constant values used in conjunction with
mode='constant'
.
 Returns:
 outputarray
Response image with correlation coefficients.
Notes
Details on the crosscorrelation are presented in [1]. This implementation uses FFT convolutions of the image and the template. Reference [2] presents similar derivations but the approximation presented in this reference is not used in our implementation.
This CuPy implementation does not force the image to float64 internally, but will use float32 for singleprecision inputs.
References
[1]J. P. Lewis, “Fast Normalized CrossCorrelation”, Industrial Light and Magic.
[2]Briechle and Hanebeck, “Template Matching using Fast Normalized Cross Correlation”, Proceedings of the SPIE (2001). DOI:10.1117/12.421129
Examples
>>> import cupy as cp >>> template = cp.zeros((3, 3)) >>> template[1, 1] = 1 >>> template array([[0., 0., 0.], [0., 1., 0.], [0., 0., 0.]]) >>> image = cp.zeros((6, 6)) >>> image[1, 1] = 1 >>> image[4, 4] = 1 >>> image array([[ 0., 0., 0., 0., 0., 0.], [ 0., 1., 0., 0., 0., 0.], [ 0., 0., 0., 0., 0., 0.], [ 0., 0., 0., 0., 0., 0.], [ 0., 0., 0., 0., 1., 0.], [ 0., 0., 0., 0., 0., 0.]]) >>> result = match_template(image, template) >>> cp.around(result, 3) array([[ 1. , 0.125, 0. , 0. ], [0.125, 0.125, 0. , 0. ], [ 0. , 0. , 0.125, 0.125], [ 0. , 0. , 0.125, 1. ]]) >>> result = match_template(image, template, pad_input=True) >>> cp.around(result, 3) array([[0.125, 0.125, 0.125, 0. , 0. , 0. ], [0.125, 1. , 0.125, 0. , 0. , 0. ], [0.125, 0.125, 0.125, 0. , 0. , 0. ], [ 0. , 0. , 0. , 0.125, 0.125, 0.125], [ 0. , 0. , 0. , 0.125, 1. , 0.125], [ 0. , 0. , 0. , 0.125, 0.125, 0.125]])
 cucim.skimage.feature.multiscale_basic_features(image, intensity=True, edges=True, texture=True, sigma_min=0.5, sigma_max=16, num_sigma=None, num_workers=None, *, channel_axis=None)#
Local features for a single or multichannel nd image.
Intensity, gradient intensity and local structure are computed at different scales thanks to Gaussian blurring.
 Parameters:
 imagendarray
Input image, which can be grayscale or multichannel.
 intensitybool, default True
If True, pixel intensities averaged over the different scales are added to the feature set.
 edgesbool, default True
If True, intensities of local gradients averaged over the different scales are added to the feature set.
 texturebool, default True
If True, eigenvalues of the Hessian matrix after Gaussian blurring at different scales are added to the feature set.
 sigma_minfloat, optional
Smallest value of the Gaussian kernel used to average local neighborhoods before extracting features.
 sigma_maxfloat, optional
Largest value of the Gaussian kernel used to average local neighborhoods before extracting features.
 num_sigmaint, optional
Number of values of the Gaussian kernel between sigma_min and sigma_max. If None, sigma_min multiplied by powers of 2 are used.
 num_workersint or None, optional
The number of parallel threads to use. If set to
None
, the full set of available cores are used. channel_axisint or None, optional
If None, the image is assumed to be a grayscale (single channel) image. Otherwise, this parameter indicates which axis of the array corresponds to channels.
 Returns:
 featuresnp.ndarray
Array of shape
image.shape + (n_features,)
. When channel_axis is not None, all channels are concatenated along the features dimension. (i.e.n_features == n_features_singlechannel * n_channels
)
 cucim.skimage.feature.peak_local_max(image, min_distance=1, threshold_abs=None, threshold_rel=None, exclude_border=True, num_peaks=inf, footprint=None, labels=None, num_peaks_per_label=inf, p_norm=inf)#
Find peaks in an image as coordinate list.
Peaks are the local maxima in a region of 2 * min_distance + 1 (i.e. peaks are separated by at least min_distance).
If both threshold_abs and threshold_rel are provided, the maximum of the two is chosen as the minimum intensity threshold of peaks.
Changed in version 0.18: Prior to version 0.18, peaks of the same height within a radius of min_distance were all returned, but this could cause unexpected behaviour. From 0.18 onwards, an arbitrary peak within the region is returned. See issue gh2592.
 Parameters:
 imagendarray
Input image.
 min_distanceint, optional
The minimal allowed distance separating peaks. To find the maximum number of peaks, use min_distance=1.
 threshold_absfloat or None, optional
Minimum intensity of peaks. By default, the absolute threshold is the minimum intensity of the image.
 threshold_relfloat or None, optional
Minimum intensity of peaks, calculated as
max(image) * threshold_rel
. exclude_borderint, tuple of ints, or bool, optional
If positive integer, exclude_border excludes peaks from within exclude_borderpixels of the border of the image. If tuple of nonnegative ints, the length of the tuple must match the input array’s dimensionality. Each element of the tuple will exclude peaks from within exclude_borderpixels of the border of the image along that dimension. If True, takes the min_distance parameter as value. If zero or False, peaks are identified regardless of their distance from the border.
 num_peaksint, optional
Maximum number of peaks. When the number of peaks exceeds num_peaks, return num_peaks peaks based on highest peak intensity.
 footprintndarray of bools, optional
If provided, footprint == 1 represents the local region within which to search for peaks at every point in image.
 labelsndarray of ints, optional
If provided, each unique region labels == value represents a unique region to search for peaks. Zero is reserved for background.
 num_peaks_per_labelint, optional
Maximum number of peaks for each label.
 p_normfloat
Which Minkowski pnorm to use. Should be in the range [1, inf]. A finite large p may cause a ValueError if overflow can occur.
inf
corresponds to the Chebyshev distance and 2 to the Euclidean distance.
 Returns:
 outputndarray
The coordinates of the peaks.
See also
Notes
The peak local maximum function returns the coordinates of local peaks (maxima) in an image. Internally, a maximum filter is used for finding local maxima. This operation dilates the original image. After comparison of the dilated and original images, this function returns the coordinates
Examples
>>> import cupy as cp >>> img1 = cp.zeros((7, 7)) >>> img1[3, 4] = 1 >>> img1[3, 2] = 1.5 >>> img1 array([[0. , 0. , 0. , 0. , 0. , 0. , 0. ], [0. , 0. , 0. , 0. , 0. , 0. , 0. ], [0. , 0. , 0. , 0. , 0. , 0. , 0. ], [0. , 0. , 1.5, 0. , 1. , 0. , 0. ], [0. , 0. , 0. , 0. , 0. , 0. , 0. ], [0. , 0. , 0. , 0. , 0. , 0. , 0. ], [0. , 0. , 0. , 0. , 0. , 0. , 0. ]])
>>> peak_local_max(img1, min_distance=1) array([[3, 2], [3, 4]])
>>> peak_local_max(img1, min_distance=2) array([[3, 2]])
>>> img2 = cp.zeros((20, 20, 20)) >>> img2[10, 10, 10] = 1 >>> img2[15, 15, 15] = 1 >>> peak_idx = peak_local_max(img2, exclude_border=0) >>> peak_idx array([[10, 10, 10], [15, 15, 15]])
>>> peak_mask = cp.zeros_like(img2, dtype=bool) >>> peak_mask[tuple(peak_idx.T)] = True >>> np.argwhere(peak_mask) array([[10, 10, 10], [15, 15, 15]])
 cucim.skimage.feature.shape_index(image, sigma=1, mode='constant', cval=0)#
Compute the shape index.
The shape index, as defined by Koenderink & van Doorn [1], is a single valued measure of local curvature, assuming the image as a 3D plane with intensities representing heights.
It is derived from the eigenvalues of the Hessian, and its value ranges from 1 to 1 (and is undefined (=NaN) in flat regions), with following ranges representing following shapes:
# Interval (s in …)
Shape
[ 1, 7/8)
Spherical cup
[7/8, 5/8)
Through
[5/8, 3/8)
Rut
[3/8, 1/8)
Saddle rut
[1/8, +1/8)
Saddle
[+1/8, +3/8)
Saddle ridge
[+3/8, +5/8)
Ridge
[+5/8, +7/8)
Dome
[+7/8, +1]
Spherical cap
 Parameters:
 image(M, N) ndarray
Input image.
 sigmafloat, optional
Standard deviation used for the Gaussian kernel, which is used for smoothing the input data before Hessian eigen value calculation.
 mode{‘constant’, ‘reflect’, ‘wrap’, ‘nearest’, ‘mirror’}, optional
How to handle values outside the image borders
 cvalfloat, optional
Used in conjunction with mode ‘constant’, the value outside the image boundaries.
 Returns:
 sndarray
Shape index
References
[1]Koenderink, J. J. & van Doorn, A. J., “Surface shape and curvature scales”, Image and Vision Computing, 1992, 10, 557564. DOI:10.1016/02628856(92)90076F
Examples
>>> from cucim.skimage.feature import shape_index >>> square = cp.zeros((5, 5)) >>> square[2, 2] = 4 >>> s = shape_index(square, sigma=0.1) >>> s array([[ nan, nan, 0.5, nan, nan], [ nan, 0. , nan, 0. , nan], [0.5, nan, 1. , nan, 0.5], [ nan, 0. , nan, 0. , nan], [ nan, nan, 0.5, nan, nan]])
 cucim.skimage.feature.structure_tensor(image, sigma=1, mode='constant', cval=0, order='rc')#
Compute structure tensor using sum of squared differences.
The (2dimensional) structure tensor A is defined as:
A = [Arr Arc] [Arc Acc]
which is approximated by the weighted sum of squared differences in a local window around each pixel in the image. This formula can be extended to a larger number of dimensions (see [1]).
 Parameters:
 imagendarray
Input image.
 sigmafloat or arraylike of float, optional
Standard deviation used for the Gaussian kernel, which is used as a weighting function for the local summation of squared differences. If sigma is an iterable, its length must be equal to image.ndim and each element is used for the Gaussian kernel applied along its respective axis.
 mode{‘constant’, ‘reflect’, ‘wrap’, ‘nearest’, ‘mirror’}, optional
How to handle values outside the image borders.
 cvalfloat, optional
Used in conjunction with mode ‘constant’, the value outside the image boundaries.
 order{‘rc’, ‘xy’}, optional
NOTE: ‘xy’ is only an option for 2D images, higher dimensions must always use ‘rc’ order. This parameter allows for the use of reverse or forward order of the image axes in gradient computation. ‘rc’ indicates the use of the first axis initially (Arr, Arc, Acc), whilst ‘xy’ indicates the usage of the last axis initially (Axx, Axy, Ayy).
 Returns:
 A_elemslist of ndarray
Upperdiagonal elements of the structure tensor for each pixel in the input image.
See also
References
Examples
>>> import cupy as cp >>> from cucim.skimage.feature import structure_tensor >>> square = cp.zeros((5, 5)) >>> square[2, 2] = 1 >>> Arr, Arc, Acc = structure_tensor(square, sigma=0.1, order="rc") >>> Acc array([[0., 0., 0., 0., 0.], [0., 1., 0., 1., 0.], [0., 4., 0., 4., 0.], [0., 1., 0., 1., 0.], [0., 0., 0., 0., 0.]])
 cucim.skimage.feature.structure_tensor_eigenvalues(A_elems)#
Compute eigenvalues of structure tensor.
 Parameters:
 A_elemslist of ndarray
The upperdiagonal elements of the structure tensor, as returned by structure_tensor.
 Returns:
 ndarray
The eigenvalues of the structure tensor, in decreasing order. The eigenvalues are the leading dimension. That is, the coordinate [i, j, k] corresponds to the ithlargest eigenvalue at position (j, k).
See also
Examples
>>> import cupy as cp >>> from cucim.skimage.feature import structure_tensor >>> from cucim.skimage.feature import structure_tensor_eigenvalues >>> square = cp.zeros((5, 5)) >>> square[2, 2] = 1 >>> A_elems = structure_tensor(square, sigma=0.1, order='rc') >>> structure_tensor_eigenvalues(A_elems)[0] array([[0., 0., 0., 0., 0.], [0., 2., 4., 2., 0.], [0., 4., 0., 4., 0.], [0., 2., 4., 2., 0.], [0., 0., 0., 0., 0.]])
filters#
 class cucim.skimage.filters.LPIFilter2D(impulse_response, **filter_params)#
Linear PositionInvariant Filter (2dimensional)
Methods
__call__
(data)Apply the filter to the given data.
 cucim.skimage.filters.apply_hysteresis_threshold(image, low, high)#
Apply hysteresis thresholding to
image
.This algorithm finds regions where
image
is greater thanhigh
ORimage
is greater thanlow
and that region is connected to a region greater thanhigh
. Parameters:
 imagearray, shape (M,[ N, …, P])
Grayscale input image.
 lowfloat, or array of same shape as
image
Lower threshold.
 highfloat, or array of same shape as
image
Higher threshold.
 Returns:
 thresholdedarray of bool, same shape as
image
Array in which
True
indicates the locations whereimage
was above the hysteresis threshold.
 thresholdedarray of bool, same shape as
References
[1]J. Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence. 1986; vol. 8, pp.679698. DOI:10.1109/TPAMI.1986.4767851
Examples
>>> import cupy as cp >>> from cucim.skimage.filters import apply_hysteresis_threshold >>> image = cp.asarray([1, 2, 3, 2, 1, 2, 1, 3, 2]) >>> apply_hysteresis_threshold(image, 1.5, 2.5).astype(int) array([0, 1, 1, 1, 0, 0, 0, 1, 1])
 cucim.skimage.filters.butterworth(image, cutoff_frequency_ratio=0.005, high_pass=True, order=2.0, channel_axis=None, *, squared_butterworth=True, npad=0)#
Apply a Butterworth filter to enhance high or low frequency features.
This filter is defined in the Fourier domain.
 Parameters:
 image(M[, N[, …, P]][, C]) ndarray
Input image.
 cutoff_frequency_ratiofloat, optional
Determines the position of the cutoff relative to the shape of the FFT. Receives a value between [0, 0.5].
 high_passbool, optional
Whether to perform a high pass filter. If False, a low pass filter is performed.
 orderfloat, optional
Order of the filter which affects the slope near the cutoff. Higher order means steeper slope in frequency space.
 channel_axisint, optional
If there is a channel dimension, provide the index here. If None (default) then all axes are assumed to be spatial dimensions.
 squared_butterworthbool, optional
When True, the square of a Butterworth filter is used. See notes below for more details.
 npadint, optional
Pad each edge of the image by npad pixels using numpy.pad’s
mode='edge'
extension.
 Returns:
 resultndarray
The Butterworthfiltered image.
Notes
A bandpass filter can be achieved by combining a highpass and lowpass filter. The user can increase npad if boundary artifacts are apparent.
The “Butterworth filter” used in image processing textbooks (e.g. [1], [2]) is often the square of the traditional Butterworth filters as described by [3], [4]. The squared version will be used here if squared_butterworth is set to
True
. The lowpass, squared Butterworth filter is given by the following expression for the lowpass case:\[H_{low}(f) = \frac{1}{1 + \left(\frac{f}{c f_s}\right)^{2n}}\]with the highpass case given by
\[H_{hi}(f) = 1  H_{low}(f)\]where \(f=\sqrt{\sum_{d=0}^{\mathrm{ndim}} f_{d}^{2}}\) is the absolute value of the spatial frequency, \(f_s\) is the sampling frequency, \(c\) the
cutoff_frequency_ratio
, and \(n\) is the filter order [1]. Whensquared_butterworth=False
, the square root of the above expressions are used instead.Note that
cutoff_frequency_ratio
is defined in terms of the sampling frequency, \(f_s\). The FFT spectrum covers the Nyquist range (\([f_s/2, f_s/2]\)) socutoff_frequency_ratio
should have a value between 0 and 0.5. The frequency response (gain) at the cutoff is 0.5 whensquared_butterworth
is true and \(1/\sqrt{2}\) when it is false.References
[2]Birchfield, Stan. Image Processing and Analysis. 2018. Cengage Learning.
[3]Butterworth, Stephen. “On the theory of filter amplifiers.” Wireless Engineer 7.6 (1930): 536541.
Examples
Apply a high pass and lowpass Butterworth filter to a grayscale and color image respectively:
>>> import cupy as cp >>> from skimage.data import camera, astronaut >>> from cucim.skimage.filters import butterworth >>> cam = cp.asarray(camera()) >>> astro = cp.asarray(astronaut()) >>> high_pass = butterworth(cam, 0.07, True, 8) >>> low_pass = butterworth(astro, 0.01, False, 4, channel_axis=1)
 cucim.skimage.filters.correlate_sparse(image, kernel, mode='reflect')#
Compute valid crosscorrelation of padded_array and kernel.
This function is fast when kernel is large with many zeros.
See
scipy.ndimage.correlate
for a description of crosscorrelation. Parameters:
 imagendarray, dtype float, shape (M, N,[ …,] P)
The input array. If mode is ‘valid’, this array should already be padded, as a margin of the same shape as kernel will be stripped off.
 kernelndarray, dtype float shape (Q, R,[ …,] S)
The kernel to be correlated. Must have the same number of dimensions as padded_array. For high performance, it should be sparse (few nonzero entries).
 modestring, optional
See scipy.ndimage.correlate for valid modes. Additionally, mode ‘valid’ is accepted, in which case no padding is applied and the result is the result for the smaller image for which the kernel is entirely inside the original data.
 Returns:
 resultarray of float, shape (M, N,[ …,] P)
The result of crosscorrelating image with kernel. If mode ‘valid’ is used, the resulting shape is (MQ+1, NR+1,[ …,] PS+1).
 cucim.skimage.filters.difference_of_gaussians(image, low_sigma, high_sigma=None, *, mode='nearest', cval=0, channel_axis=None, truncate=4.0)#
Find features between
low_sigma
andhigh_sigma
in size.This function uses the Difference of Gaussians method for applying bandpass filters to multidimensional arrays. The input array is blurred with two Gaussian kernels of differing sigmas to produce two intermediate, filtered images. The moreblurred image is then subtracted from the lessblurred image. The final output image will therefore have had highfrequency components attenuated by the smallersigma Gaussian, and low frequency components will have been removed due to their presence in the moreblurred intermediate.
 Parameters:
 imagendarray
Input array to filter.
 low_sigmascalar or sequence of scalars
Standard deviation(s) for the Gaussian kernel with the smaller sigmas across all axes. The standard deviations are given for each axis as a sequence, or as a single number, in which case the single number is used as the standard deviation value for all axes.
 high_sigmascalar or sequence of scalars, optional (default is None)
Standard deviation(s) for the Gaussian kernel with the larger sigmas across all axes. The standard deviations are given for each axis as a sequence, or as a single number, in which case the single number is used as the standard deviation value for all axes. If None is given (default), sigmas for all axes are calculated as 1.6 * low_sigma.
 mode{‘reflect’, ‘constant’, ‘nearest’, ‘mirror’, ‘wrap’}, optional
The
mode
parameter determines how the array borders are handled, wherecval
is the value when mode is equal to ‘constant’. Default is ‘nearest’. cvalscalar, optional
Value to fill past edges of input if
mode
is ‘constant’. Default is 0.0 channel_axisint or None, optional
If None, the image is assumed to be a grayscale (single channel) image. Otherwise, this parameter indicates which axis of the array corresponds to channels.
 truncatefloat, optional (default is 4.0)
Truncate the filter at this many standard deviations.
 Returns:
 filtered_imagendarray
the filtered array.
See also
skimage.feature.blog_dog
Notes
This function will subtract an array filtered with a Gaussian kernel with sigmas given by
high_sigma
from an array filtered with a Gaussian kernel with sigmas provided bylow_sigma
. The values forhigh_sigma
must always be greater than or equal to the corresponding values inlow_sigma
, or aValueError
will be raised.When
high_sigma
is none, the values forhigh_sigma
will be calculated as 1.6x the corresponding values inlow_sigma
. This ratio was originally proposed by Marr and Hildreth (1980) [1] and is commonly used when approximating the inverted Laplacian of Gaussian, which is used in edge and blob detection.Input image is converted according to the conventions of
img_as_float
.Except for sigma values, all parameters are used for both filters.
References
[1]Marr, D. and Hildreth, E. Theory of Edge Detection. Proc. R. Soc. Lond. Series B 207, 187217 (1980). https://doi.org/10.1098/rspb.1980.0020
Examples
Apply a simple Difference of Gaussians filter to a color image:
>>> from skimage.data import astronaut >>> from cucim.skimage.filters import difference_of_gaussians >>> astro = cp.asarray(astronaut()) >>> filtered_image = difference_of_gaussians(astro, 2, 10, ... channel_axis=1)
Apply a Laplacian of Gaussian filter as approximated by the Difference of Gaussians filter:
>>> filtered_image = difference_of_gaussians(astro, 2, ... channel_axis=1)
Apply a Difference of Gaussians filter to a grayscale image using different sigma values for each axis:
>>> from skimage.data import camera >>> cam = cp.array(camera()) >>> filtered_image = difference_of_gaussians(cam, (2,5), (3,20))
 cucim.skimage.filters.farid(image, mask=None, *, axis=None, mode='reflect', cval=0.0)#
Find the edge magnitude using the Farid transform.
 Parameters:
 imagecp.ndarray
The input image.
 maskcp.ndarray of bool, optional
Clip the output image to this mask. (Values where mask=0 will be set to 0.)
 axisint or sequence of int, optional
Compute the edge filter along this axis. If not provided, the edge magnitude is computed. This is defined as:
farid_mag = cp.sqrt(sum([farid(image, axis=i)**2 for i in range(image.ndim)]) / image.ndim)
The magnitude is also computed if axis is a sequence.
 modestr or sequence of str, optional
The boundary mode for the convolution. See scipy.ndimage.convolve for a description of the modes. This can be either a single boundary mode or one boundary mode per axis.
 cvalfloat, optional
When mode is
'constant'
, this is the constant used in values outside the boundary of the image data. Returns
 ——
 output2D array
The Farid edge map.
See also
Notes
Take the square root of the sum of the squares of the horizontal and vertical derivatives to get a magnitude that is somewhat insensitive to direction. Similar to the Scharr operator, this operator is designed with a rotation invariance constraint.
References
[1]Farid, H. and Simoncelli, E. P., “Differentiation of discrete multidimensional signals”, IEEE Transactions on Image Processing 13(4): 496508, 2004. DOI:10.1109/TIP.2004.823819
[2]Wikipedia, “Farid and Simoncelli Derivatives.” Available at: <https://en.wikipedia.org/wiki/Image_derivatives#Farid_and_Simoncelli_Derivatives>
Examples
>>> import cupy as cp >>> from skimage import data >>> camera = cp.array(data.camera()) >>> from cucim.skimage import filters >>> edges = filters.farid(camera)
 cucim.skimage.filters.farid_h(image, *, mask=None)#
Find the horizontal edges of an image using the Farid transform.
 Parameters:
 image2D array
Image to process.
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Farid edge map.
Notes
The kernel was constructed using the 5tap weights from [1].
References
[1]Farid, H. and Simoncelli, E. P., “Differentiation of discrete multidimensional signals”, IEEE Transactions on Image Processing 13(4): 496508, 2004. DOI:10.1109/TIP.2004.823819
[2]Farid, H. and Simoncelli, E. P. “Optimally rotationequivariant directional derivative kernels”, In: 7th International Conference on Computer Analysis of Images and Patterns, Kiel, Germany. Sep, 1997.
 cucim.skimage.filters.farid_v(image, *, mask=None)#
Find the vertical edges of an image using the Farid transform.
 Parameters:
 image2D array
Image to process.
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Farid edge map.
Notes
The kernel was constructed using the 5tap weights from [1].
References
[1]Farid, H. and Simoncelli, E. P., “Differentiation of discrete multidimensional signals”, IEEE Transactions on Image Processing 13(4): 496508, 2004. DOI:10.1109/TIP.2004.823819
 cucim.skimage.filters.filter_forward(data, impulse_response=None, filter_params=None, predefined_filter=None)#
Apply the given filter to data.
 Parameters:
 data(M, N) ndarray
Input data.
 impulse_responsecallable f(r, c, **filter_params)
Impulse response of the filter. See LPIFilter2D.__init__.
 filter_paramsdict, optional
Additional keyword parameters to the impulse_response function.
 Other Parameters:
 predefined_filterLPIFilter2D
If you need to apply the same filter multiple times over different images, construct the LPIFilter2D and specify it here.
Examples
Gaussian filter without normalization:
>>> def filt_func(r, c, sigma=1): ... return cp.exp((r**2 + c**2)/(2 * sigma**2)) >>> >>> from skimage import data >>> filtered = filter_forward(cp.array(data.coins()), filt_func)
 cucim.skimage.filters.filter_inverse(data, impulse_response=None, filter_params=None, max_gain=2, predefined_filter=None)#
Apply the filter in reverse to the given data.
 Parameters:
 data(M, N) ndarray
Input data.
 impulse_responsecallable f(r, c, **filter_params)
Impulse response of the filter. See
LPIFilter2D
. This is a required argument unless a predifined_filter is provided. filter_paramsdict, optional
Additional keyword parameters to the impulse_response function.
 max_gainfloat, optional
Limit the filter gain. Often, the filter contains zeros, which would cause the inverse filter to have infinite gain. High gain causes amplification of artefacts, so a conservative limit is recommended.
 Other Parameters:
 predefined_filterLPIFilter2D, optional
If you need to apply the same filter multiple times over different images, construct the LPIFilter2D and specify it here.
 cucim.skimage.filters.frangi(image, sigmas=range(1, 10, 2), scale_range=None, scale_step=None, alpha=0.5, beta=0.5, gamma=None, black_ridges=True, mode='reflect', cval=0)#
Filter an image with the Frangi vesselness filter.
This filter can be used to detect continuous ridges, e.g. vessels, wrinkles, rivers. It can be used to calculate the fraction of the whole image containing such objects.
Defined only for 2D and 3D images. Calculates the eigenvectors of the Hessian to compute the similarity of an image region to vessels, according to the method described in [1].
 Parameters:
 image(N, M[, P]) ndarray
Array with input image data.
 sigmasiterable of floats, optional
Sigmas used as scales of filter, i.e., np.arange(scale_range[0], scale_range[1], scale_step)
 scale_range2tuple of floats, optional
The range of sigmas used.
 scale_stepfloat, optional
Step size between sigmas.
 alphafloat, optional
Frangi correction constant that adjusts the filter’s sensitivity to deviation from a platelike structure.
 betafloat, optional
Frangi correction constant that adjusts the filter’s sensitivity to deviation from a bloblike structure.
 gammafloat, optional
Frangi correction constant that adjusts the filter’s sensitivity to areas of high variance/texture/structure. The default, None, uses half of the maximum Hessian norm.
 black_ridgesboolean, optional
When True (the default), the filter detects black ridges; when False, it detects white ridges.
 mode{‘constant’, ‘reflect’, ‘wrap’, ‘nearest’, ‘mirror’}, optional
How to handle values outside the image borders.
 cvalfloat, optional
Used in conjunction with mode ‘constant’, the value outside the image boundaries.
 Returns:
 out(N, M[, P]) ndarray
Filtered image (maximum of pixels across all scales).
Notes
Earlier versions of this filter were implemented by Marc Schrijver, (November 2001), D. J. Kroon, University of Twente (May 2009) [2], and D. G. Ellis (January 2017) [3].
References
[1]Frangi, A. F., Niessen, W. J., Vincken, K. L., & Viergever, M. A. (1998,). Multiscale vessel enhancement filtering. In International Conference on Medical Image Computing and ComputerAssisted Intervention (pp. 130137). Springer Berlin Heidelberg. DOI:10.1007/BFb0056195
[2]Kroon, D. J.: Hessian based Frangi vesselness filter.
[3]Ellis, D. G.: ellisdg/frangi3d
 cucim.skimage.filters.gabor(image, frequency, theta=0, bandwidth=1, sigma_x=None, sigma_y=None, n_stds=3, offset=0, mode='reflect', cval=0)#
Return real and imaginary responses to Gabor filter.
The real and imaginary parts of the Gabor filter kernel are applied to the image and the response is returned as a pair of arrays.
Gabor filter is a linear filter with a Gaussian kernel which is modulated by a sinusoidal plane wave. Frequency and orientation representations of the Gabor filter are similar to those of the human visual system. Gabor filter banks are commonly used in computer vision and image processing. They are especially suitable for edge detection and texture classification.
 Parameters:
 image2D array
Input image.
 frequencyfloat
Spatial frequency of the harmonic function. Specified in pixels.
 thetafloat, optional
Orientation in radians. If 0, the harmonic is in the xdirection.
 bandwidthfloat, optional
The bandwidth captured by the filter. For fixed bandwidth,
sigma_x
andsigma_y
will decrease with increasing frequency. This value is ignored ifsigma_x
andsigma_y
are set by the user. sigma_x, sigma_yfloat, optional
Standard deviation in x and ydirections. These directions apply to the kernel before rotation. If theta = pi/2, then the kernel is rotated 90 degrees so that
sigma_x
controls the vertical direction. n_stdsscalar, optional
The linear size of the kernel is n_stds (3 by default) standard deviations.
 offsetfloat, optional
Phase offset of harmonic function in radians.
 mode{‘constant’, ‘nearest’, ‘reflect’, ‘mirror’, ‘wrap’}, optional
Mode used to convolve image with a kernel, passed to ndi.convolve
 cvalscalar, optional
Value to fill past edges of input if
mode
of convolution is ‘constant’. The parameter is passed to ndi.convolve.
 Returns:
 real, imagarrays
Filtered images using the real and imaginary parts of the Gabor filter kernel. Images are of the same dimensions as the input one.
References
Examples
>>> import cupy as cp >>> from cucim.skimage.filters import gabor >>> from skimage import data, io >>> from matplotlib import pyplot as plt
>>> image = cp.array(data.coins()) >>> # detecting edges in a coin image >>> filt_real, filt_imag = gabor(image, frequency=0.6) >>> plt.figure() >>> io.imshow(cp.asnumpy(filt_real)) >>> io.show()
>>> # less sensitivity to finer details with the lower frequency kernel >>> filt_real, filt_imag = gabor(image, frequency=0.1) >>> plt.figure() >>> io.imshow(cp.asnumpy(filt_real) >>> io.show()
 cucim.skimage.filters.gabor_kernel(frequency, theta=0, bandwidth=1, sigma_x=None, sigma_y=None, n_stds=3, offset=0, dtype=None, *, float_dtype=None)#
Return complex 2D Gabor filter kernel.
Gabor kernel is a Gaussian kernel modulated by a complex harmonic function. Harmonic function consists of an imaginary sine function and a real cosine function. Spatial frequency is inversely proportional to the wavelength of the harmonic and to the standard deviation of a Gaussian kernel. The bandwidth is also inversely proportional to the standard deviation.
 Parameters:
 frequencyfloat
Spatial frequency of the harmonic function. Specified in pixels.
 thetafloat, optional
Orientation in radians. If 0, the harmonic is in the xdirection.
 bandwidthfloat, optional
The bandwidth captured by the filter. For fixed bandwidth,
sigma_x
andsigma_y
will decrease with increasing frequency. This value is ignored ifsigma_x
andsigma_y
are set by the user. sigma_x, sigma_yfloat, optional
Standard deviation in x and ydirections. These directions apply to the kernel before rotation. If theta = pi/2, then the kernel is rotated 90 degrees so that
sigma_x
controls the vertical direction. n_stdsscalar, optional
The linear size of the kernel is n_stds (3 by default) standard deviations
 offsetfloat, optional
Phase offset of harmonic function in radians.
 dtype{np.complex64, np.complex128}
Specifies if the filter is single or double precision complex.
 Returns:
 gcomplex array
Complex filter kernel.
References
Examples
>>> import cupy as cp >>> from cucim.skimage.filters import gabor_kernel >>> from skimage import io >>> from matplotlib import pyplot as plt
>>> gk = gabor_kernel(frequency=0.2) >>> plt.figure() >>> io.imshow(cp.asnumpy(gk.real)) >>> io.show()
>>> # more ripples (equivalent to increasing the size of the >>> # Gaussian spread) >>> gk = gabor_kernel(frequency=0.2, bandwidth=0.1) >>> plt.figure() >>> io.imshow(cp.asnumpy(gk.real)) >>> io.show()
 cucim.skimage.filters.gaussian(image, sigma=1, output=None, mode='nearest', cval=0, preserve_range=False, truncate=4.0, *, channel_axis=<ChannelAxisNotSet>)#
Multidimensional Gaussian filter.
 Parameters:
 imagearraylike
Input image (grayscale or color) to filter.
 sigmascalar or sequence of scalars, optional
Standard deviation for Gaussian kernel. The standard deviations of the Gaussian filter are given for each axis as a sequence, or as a single number, in which case it is equal for all axes.
 outputarray, optional
The
output
parameter passes an array in which to store the filter output. mode{‘reflect’, ‘constant’, ‘nearest’, ‘mirror’, ‘wrap’}, optional
The
mode
parameter determines how the array borders are handled, wherecval
is the value when mode is equal to ‘constant’. Default is ‘nearest’. cvalscalar, optional
Value to fill past edges of input if
mode
is ‘constant’. Default is 0.0 preserve_rangebool, optional
Whether to keep the original range of values. Otherwise, the input image is converted according to the conventions of
img_as_float
. Also see https://scikitimage.org/docs/dev/user_guide/data_types.html truncatefloat, optional
Truncate the filter at this many standard deviations.
 channel_axisint or None, optional
If None, the image is assumed to be a grayscale (single channel) image. Otherwise, this parameter indicates which axis of the array corresponds to channels.
Warning
Automatic detection of the color channel based on the old deprecated
multichannel=None
was broken in version 0.19. In 0.20 this behavior is recovered. The last axis of an image with dimensions (M, N, 3) is interpreted as a color channel if channel_axis is not set. Starting with 2023.03.06,channel_axis=None
will be used as the new default value.
 Returns:
 filtered_imagendarray
the filtered array
Notes
This function is a wrapper around
scipy.ndi.gaussian_filter()
.Integer arrays are converted to float.
The
output
should be floating point data type since gaussian converts to float providedimage
. Ifoutput
is not provided, another array will be allocated and returned as the result.The multidimensional filter is implemented as a sequence of onedimensional convolution filters. The intermediate arrays are stored in the same data type as the output. Therefore, for output types with a limited precision, the results may be imprecise because intermediate results may be stored with insufficient precision.
Examples
>>> import cupy as cp >>> a = cp.zeros((3, 3)) >>> a[1, 1] = 1 >>> a array([[0., 0., 0.], [0., 1., 0.], [0., 0., 0.]]) >>> gaussian(a, sigma=0.4) # mild smoothing array([[0.00163116, 0.03712502, 0.00163116], [0.03712502, 0.84496158, 0.03712502], [0.00163116, 0.03712502, 0.00163116]]) >>> gaussian(a, sigma=1) # more smoothing array([[0.05855018, 0.09653293, 0.05855018], [0.09653293, 0.15915589, 0.09653293], [0.05855018, 0.09653293, 0.05855018]]) >>> # Several modes are possible for handling boundaries >>> gaussian(a, sigma=1, mode='reflect') array([[0.08767308, 0.12075024, 0.08767308], [0.12075024, 0.16630671, 0.12075024], [0.08767308, 0.12075024, 0.08767308]]) >>> # For RGB images, each is filtered separately >>> from skimage.data import astronaut >>> image = cp.array(astronaut()) >>> filtered_img = gaussian(image, sigma=1, channel_axis=1)
 cucim.skimage.filters.hessian(image, sigmas=range(1, 10, 2), scale_range=None, scale_step=None, alpha=0.5, beta=0.5, gamma=15, black_ridges=True, mode='reflect', cval=0)#
Filter an image with the Hybrid Hessian filter.
This filter can be used to detect continuous edges, e.g. vessels, wrinkles, rivers. It can be used to calculate the fraction of the whole image containing such objects.
Defined only for 2D and 3D images. Almost equal to Frangi filter, but uses alternative method of smoothing. Refer to [1] to find the differences between Frangi and Hessian filters.
 Parameters:
 image(N, M[, P]) ndarray
Array with input image data.
 sigmasiterable of floats, optional
Sigmas used as scales of filter, i.e., np.arange(scale_range[0], scale_range[1], scale_step)
 scale_range2tuple of floats, optional
The range of sigmas used.
 scale_stepfloat, optional
Step size between sigmas.
 betafloat, optional
Frangi correction constant that adjusts the filter’s sensitivity to deviation from a bloblike structure.
 gammafloat, optional
Frangi correction constant that adjusts the filter’s sensitivity to areas of high variance/texture/structure.
 black_ridgesboolean, optional
When True (the default), the filter detects black ridges; when False, it detects white ridges.
 mode{‘constant’, ‘reflect’, ‘wrap’, ‘nearest’, ‘mirror’}, optional
How to handle values outside the image borders.
 cvalfloat, optional
Used in conjunction with mode ‘constant’, the value outside the image boundaries.
 Returns:
 out(N, M[, P]) ndarray
Filtered image (maximum of pixels across all scales).
Notes
Written by Marc Schrijver (November 2001) ReWritten by D. J. Kroon University of Twente (May 2009) [2]
References
[1]Ng, C. C., Yap, M. H., Costen, N., & Li, B. (2014,). Automatic wrinkle detection using hybrid Hessian filter. In Asian Conference on Computer Vision (pp. 609622). Springer International Publishing. DOI:10.1007/9783319168111_40
[2]Kroon, D. J.: Hessian based Frangi vesselness filter.
 cucim.skimage.filters.laplace(image, ksize=3, mask=None)#
Find the edges of an image using the Laplace operator.
 Parameters:
 imagendarray
Image to process.
 ksizeint, optional
Define the size of the discrete Laplacian operator such that it will have a size of (ksize,) * image.ndim.
 maskndarray, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 outputndarray
The Laplace edge map.
Notes
The Laplacian operator is generated using the function skimage.restoration.uft.laplacian().
 cucim.skimage.filters.median(image, footprint=None, out=None, mode='nearest', cval=0.0, behavior='ndimage', *, algorithm='auto', algorithm_kwargs={})#
Return local median of an image.
 Parameters:
 imagearraylike
Input image.
 footprintndarray, tuple of int, or None
If
None
,footprint
will be a ND array with 3 elements for each dimension (e.g., vector, square, cube, etc.). If footprint is a tuple of integers, it will be an array of ones with the given shape. Otherwise, ifbehavior=='rank'
,footprint
is a 2D array of 1’s and 0’s. Ifbehavior=='ndimage'
,footprint
is a ND array of 1’s and 0’s with the same number of dimension asimage
. Note that upstream scikitimage currently does not support supplying a tuple for footprint. It is added here to avoid overhead of generating a small weights array in cases where it is not needed. outndarray, (same dtype as image), optional
If None, a new array is allocated.
 mode{‘reflect’, ‘constant’, ‘nearest’, ‘mirror’,’‘wrap’}, optional
The mode parameter determines how the array borders are handled, where
cval
is the value when mode is equal to ‘constant’. Default is ‘nearest’.New in version 0.15:
mode
is used whenbehavior='ndimage'
. cvalscalar, optional
Value to fill past edges of input if mode is ‘constant’. Default is 0.0
New in version 0.15:
cval
was added in 0.15 is used whenbehavior='ndimage'
. behavior{‘ndimage’, ‘rank’}, optional
Either to use the old behavior (i.e., < 0.15) or the new behavior. The old behavior will call the
skimage.filters.rank.median()
. The new behavior will call thescipy.ndimage.median_filter()
. Default is ‘ndimage’.New in version 0.15:
behavior
is introduced in 0.15Changed in version 0.16: Default
behavior
has been changed from ‘rank’ to ‘ndimage’
 Returns:
 out2D array (same dtype as input image)
Output image.
 Other Parameters:
 algorithm{‘auto’, ‘histogram’, ‘sorting’}
Determines which algorithm is used to compute the median. The default of ‘auto’ will attempt to use a histogrambased algorithm for 2D images with 8 or 16bit integer data types. Otherwise a sortingbased algorithm will be used. Note: this paramter is cuCIMspecific and does not exist in upstream scikitimage.
 algorithm_kwargsdict
Any additional algorithmspecific keywords. Currently can only be used to set the number of parallel partitions for the ‘histogram’ algorithm. (e.g.
algorithm_kwargs={'partitions': 256}
). Note: this paramter is cuCIMspecific and does not exist in upstream scikitimage. selemDEPRECATED
Deprecated in favor of footprint.
Deprecated since version 22.02.00.
See also
skimage.filters.rank.median()
Rankbased implementation of the median filtering offering more flexibility with additional parameters but dedicated for unsigned integer images.
Notes
An efficient, histogrambased median filter as described in [1] is faster than the sorting based approach for larger kernel sizes (e.g. greater than 13x13 or so in 2D). It has nearconstant run time regardless of the kernel size. The algorithm presented in [1] has been adapted to additional bit depths here. When algorithm=’auto’, the histogrambased algorithm will be chosen for integervalued images with sufficiently large footprint size. Otherwise, the sortingbased approach is used.
References
[1] (1,2)O. Green, “Efficient Scalable Median Filtering Using HistogramBased Operations,” in IEEE Transactions on Image Processing, vol. 27, no. 5, pp. 22172228, May 2018, https://doi.org/10.1109/TIP.2017.2781375.
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage.morphology import disk >>> from cucim.skimage.filters import median >>> img = cp.array(data.camera()) >>> med = median(img, disk(5))
 cucim.skimage.filters.meijering(image, sigmas=range(1, 10, 2), alpha=None, black_ridges=True, mode='reflect', cval=0)#
Filter an image with the Meijering neuriteness filter.
This filter can be used to detect continuous ridges, e.g. neurites, wrinkles, rivers. It can be used to calculate the fraction of the whole image containing such objects.
Calculates the eigenvectors of the Hessian to compute the similarity of an image region to neurites, according to the method described in [1].
 Parameters:
 image(N, M[, …, P]) ndarray
Array with input image data.
 sigmasiterable of floats, optional
Sigmas used as scales of filter
 alphafloat, optional
Shaping filter constant, that selects maximally flat elongated features. The default, None, selects the optimal value 1/(ndim+1).
 black_ridgesboolean, optional
When True (the default), the filter detects black ridges; when False, it detects white ridges.
 mode{‘constant’, ‘reflect’, ‘wrap’, ‘nearest’, ‘mirror’}, optional
How to handle values outside the image borders.
 cvalfloat, optional
Used in conjunction with mode ‘constant’, the value outside the image boundaries.
 Returns:
 out(N, M[, …, P]) ndarray
Filtered image (maximum of pixels across all scales).
References
[1]Meijering, E., Jacob, M., Sarria, J. C., Steiner, P., Hirling, H., Unser, M. (2004). Design and validation of a tool for neurite tracing and analysis in fluorescence microscopy images. Cytometry Part A, 58(2), 167176. DOI:10.1002/cyto.a.20022
 cucim.skimage.filters.prewitt(image, mask=None, *, axis=None, mode='reflect', cval=0.0)#
Find the edge magnitude using the Prewitt transform.
 Parameters:
 imagearray
The input image.
 maskarray of bool, optional
Clip the output image to this mask. (Values where mask=0 will be set to 0.)
 axisint or sequence of int, optional
Compute the edge filter along this axis. If not provided, the edge magnitude is computed. This is defined as:
prw_mag = np.sqrt(sum([prewitt(image, axis=i)**2 for i in range(image.ndim)]) / image.ndim)
The magnitude is also computed if axis is a sequence.
 modestr or sequence of str, optional
The boundary mode for the convolution. See scipy.ndimage.convolve for a description of the modes. This can be either a single boundary mode or one boundary mode per axis.
 cvalfloat, optional
When mode is
'constant'
, this is the constant used in values outside the boundary of the image data.
 Returns:
 outputarray of float
The Prewitt edge map.
See also
Notes
The edge magnitude depends slightly on edge directions, since the approximation of the gradient operator by the Prewitt operator is not completely rotation invariant. For a better rotation invariance, the Scharr operator should be used. The Sobel operator has a better rotation invariance than the Prewitt operator, but a worse rotation invariance than the Scharr operator.
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage import filters >>> camera = cp.array(data.camera()) >>> edges = filters.prewitt(camera)
 cucim.skimage.filters.prewitt_h(image, mask=None)#
Find the horizontal edges of an image using the Prewitt transform.
 Parameters:
 image2D array
Image to process.
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Prewitt edge map.
Notes
We use the following kernel:
1/3 1/3 1/3 0 0 0 1/3 1/3 1/3
 cucim.skimage.filters.prewitt_v(image, mask=None)#
Find the vertical edges of an image using the Prewitt transform.
 Parameters:
 image2D array
Image to process.
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Prewitt edge map.
Notes
We use the following kernel:
1/3 0 1/3 1/3 0 1/3 1/3 0 1/3
 cucim.skimage.filters.rank_order(image)#
Return an image of the same shape where each pixel is the index of the pixel value in the ascending order of the unique values of
image
, aka the rankorder value. Parameters:
 imagecp.ndarray
 Returns:
 labelscp.ndarray of unsigned integers, of shape image.shape
New array where each pixel has the rankorder value of the corresponding pixel in
image
. Pixel values are between 0 and n  1, where n is the number of distinct unique values inimage
. The dtype of this array will be determined bycp.min_scalar_type(image.size)
. original_values1D cp.ndarray
Unique original values of
image
. This will have the same dtype asimage
.
Examples
>>> a = cp.asarray([[1, 4, 5], [4, 4, 1], [5, 1, 1]]) >>> a array([[1, 4, 5], [4, 4, 1], [5, 1, 1]]) >>> rank_order(a) (array([[0, 1, 2], [1, 1, 0], [2, 0, 0]], dtype=uint8), array([1, 4, 5])) >>> b = cp.asarray([1., 2.5, 3.1, 2.5]) >>> rank_order(b) (array([0, 1, 2, 1], dtype=uint8), array([1. , 2.5, 3.1]))
 cucim.skimage.filters.roberts(image, mask=None)#
Find the edge magnitude using Roberts’ cross operator.
 Parameters:
 image2D array
Image to process.
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Roberts’ Cross edge map.
See also
roberts_pos_diag
,roberts_neg_diag
diagonal edge detection.
sobel
,scharr
,prewitt
,cucim.skimage.feature.canny
Examples
>>> import cupy as cp >>> from skimage import data >>> camera = cp.array(data.camera()) >>> from cucim.skimage import filters >>> edges = filters.roberts(camera)
 cucim.skimage.filters.roberts_neg_diag(image, mask=None)#
Find the cross edges of an image using the Roberts’ Cross operator.
The kernel is applied to the input image to produce separate measurements of the gradient component one orientation.
 Parameters:
 image2D array
Image to process.
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Robert’s edge map.
Notes
We use the following kernel:
0 1 1 0
 cucim.skimage.filters.roberts_pos_diag(image, mask=None)#
Find the cross edges of an image using Roberts’ cross operator.
The kernel is applied to the input image to produce separate measurements of the gradient component one orientation.
 Parameters:
 image2D array
Image to process.
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Robert’s edge map.
Notes
We use the following kernel:
1 0 0 1
 cucim.skimage.filters.sato(image, sigmas=range(1, 10, 2), black_ridges=True, mode='reflect', cval=0)#
Filter an image with the Sato tubeness filter.
This filter can be used to detect continuous ridges, e.g. tubes, wrinkles, rivers. It can be used to calculate the fraction of the whole image containing such objects.
Defined only for 2D and 3D images. Calculates the eigenvectors of the Hessian to compute the similarity of an image region to tubes, according to the method described in [1].
 Parameters:
 image(N, M[, P]) ndarray
Array with input image data.
 sigmasiterable of floats, optional
Sigmas used as scales of filter.
 black_ridgesboolean, optional
When True (the default), the filter detects black ridges; when False, it detects white ridges.
 mode{‘constant’, ‘reflect’, ‘wrap’, ‘nearest’, ‘mirror’}, optional
How to handle values outside the image borders.
 cvalfloat, optional
Used in conjunction with mode ‘constant’, the value outside the image boundaries.
 Returns:
 out(N, M[, P]) ndarray
Filtered image (maximum of pixels across all scales).
References
[1]Sato, Y., Nakajima, S., Shiraga, N., Atsumi, H., Yoshida, S., Koller, T., …, Kikinis, R. (1998). Threedimensional multiscale line filter for segmentation and visualization of curvilinear structures in medical images. Medical image analysis, 2(2), 143168. DOI:10.1016/S13618415(98)800091
 cucim.skimage.filters.scharr(image, mask=None, *, axis=None, mode='reflect', cval=0.0)#
Find the edge magnitude using the Scharr transform.
 Parameters:
 imagearray
The input image.
 maskarray of bool, optional
Clip the output image to this mask. (Values where mask=0 will be set to 0.)
 axisint or sequence of int, optional
Compute the edge filter along this axis. If not provided, the edge magnitude is computed. This is defined as:
sch_mag = np.sqrt(sum([scharr(image, axis=i)**2 for i in range(image.ndim)]) / image.ndim)
The magnitude is also computed if axis is a sequence.
 modestr or sequence of str, optional
The boundary mode for the convolution. See scipy.ndimage.convolve for a description of the modes. This can be either a single boundary mode or one boundary mode per axis.
 cvalfloat, optional
When mode is
'constant'
, this is the constant used in values outside the boundary of the image data.
 Returns:
 outputarray of float
The Scharr edge map.
See also
Notes
The Scharr operator has a better rotation invariance than other edge filters such as the Sobel or the Prewitt operators.
References
[1]D. Kroon, 2009, Short Paper University Twente, Numerical Optimization of Kernel Based Image Derivatives.
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage import filters >>> camera = cp.array(data.camera()) >>> edges = filters.scharr(camera)
 cucim.skimage.filters.scharr_h(image, mask=None)#
Find the horizontal edges of an image using the Scharr transform.
 Parameters:
 image2D array
Image to process.
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Scharr edge map.
Notes
We use the following kernel:
3 10 3 0 0 0 3 10 3
References
[1]D. Kroon, 2009, Short Paper University Twente, Numerical Optimization of Kernel Based Image Derivatives.
 cucim.skimage.filters.scharr_v(image, mask=None)#
Find the vertical edges of an image using the Scharr transform.
 Parameters:
 image2D array
Image to process
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Scharr edge map.
Notes
We use the following kernel:
3 0 3 10 0 10 3 0 3
References
[1]D. Kroon, 2009, Short Paper University Twente, Numerical Optimization of Kernel Based Image Derivatives.
 cucim.skimage.filters.sobel(image, mask=None, *, axis=None, mode='reflect', cval=0.0)#
Find edges in an image using the Sobel filter.
 Parameters:
 imagearray
The input image.
 maskarray of bool, optional
Clip the output image to this mask. (Values where mask=0 will be set to 0.)
 axisint or sequence of int, optional
Compute the edge filter along this axis. If not provided, the edge magnitude is computed. This is defined as:
sobel_mag = np.sqrt(sum([sobel(image, axis=i)**2 for i in range(image.ndim)]) / image.ndim)
The magnitude is also computed if axis is a sequence.
 modestr or sequence of str, optional
The boundary mode for the convolution. See scipy.ndimage.convolve for a description of the modes. This can be either a single boundary mode or one boundary mode per axis.
 cvalfloat, optional
When mode is
'constant'
, this is the constant used in values outside the boundary of the image data.
 Returns:
 outputarray of float
The Sobel edge map.
See also
References
[1]D. Kroon, 2009, Short Paper University Twente, Numerical Optimization of Kernel Based Image Derivatives.
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage import filters >>> camera = cp.array(data.camera()) >>> edges = filters.sobel(camera)
 cucim.skimage.filters.sobel_h(image, mask=None)#
Find the horizontal edges of an image using the Sobel transform.
 Parameters:
 image2D array
Image to process.
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Sobel edge map.
Notes
We use the following kernel:
1 2 1 0 0 0 1 2 1
 cucim.skimage.filters.sobel_v(image, mask=None)#
Find the vertical edges of an image using the Sobel transform.
 Parameters:
 image2D array
Image to process.
 mask2D array, optional
An optional mask to limit the application to a certain area. Note that pixels surrounding masked regions are also masked to prevent masked regions from affecting the result.
 Returns:
 output2D array
The Sobel edge map.
Notes
We use the following kernel:
1 0 1 2 0 2 1 0 1
 cucim.skimage.filters.threshold_isodata(image=None, nbins=256, return_all=False, *, hist=None)#
Return threshold value(s) based on ISODATA method.
Histogrambased threshold, known as RidlerCalvard method or intermeans. Threshold values returned satisfy the following equality:
threshold = (image[image <= threshold].mean() + image[image > threshold].mean()) / 2.0
That is, returned thresholds are intensities that separate the image into two groups of pixels, where the threshold intensity is midway between the mean intensities of these groups.
For integer images, the above equality holds to within one; for floating point images, the equality holds to within the histogram binwidth.
Either image or hist must be provided. In case hist is given, the actual histogram of the image is ignored.
 Parameters:
 image(N, M[, …, P]) ndarray
Grayscale input image.
 nbinsint, optional
Number of bins used to calculate histogram. This value is ignored for integer arrays.
 return_allbool, optional
If False (default), return only the lowest threshold that satisfies the above equality. If True, return all valid thresholds.
 histarray, or 2tuple of arrays, optional
Histogram to determine the threshold from and a corresponding array of bin center intensities. Alternatively, only the histogram can be passed.
 Returns:
 thresholdfloat or int or array
Threshold value(s).
References
[1]Ridler, TW & Calvard, S (1978), “Picture thresholding using an iterative selection method” IEEE Transactions on Systems, Man and Cybernetics 8: 630632, DOI:10.1109/TSMC.1978.4310039
[2]Sezgin M. and Sankur B. (2004) “Survey over Image Thresholding Techniques and Quantitative Performance Evaluation” Journal of Electronic Imaging, 13(1): 146165, http://www.busim.ee.boun.edu.tr/~sankur/SankurFolder/Threshold_survey.pdf DOI:10.1117/1.1631315
[3]ImageJ AutoThresholder code, http://fiji.sc/wiki/index.php/Auto_Threshold
Examples
>>> from skimage.data import coins >>> image = cp.array(coins()) >>> thresh = threshold_isodata(image) >>> binary = image > thresh
 cucim.skimage.filters.threshold_li(image, *, tolerance=None, initial_guess=None, iter_callback=None)#
Compute threshold value by Li’s iterative Minimum Cross Entropy method.
 Parameters:
 image(N, M[, …, P]) ndarray
Grayscale input image.
 tolerancefloat, optional
Finish the computation when the change in the threshold in an iteration is less than this value. By default, this is half the smallest difference between intensity values in
image
. initial_guessfloat or Callable[[array[float]], float], optional
Li’s iterative method uses gradient descent to find the optimal threshold. If the image intensity histogram contains more than two modes (peaks), the gradient descent could get stuck in a local optimum. An initial guess for the iteration can help the algorithm find the globallyoptimal threshold. A float value defines a specific start point, while a callable should take in an array of image intensities and return a float value. Example valid callables include
numpy.mean
(default),lambda arr: numpy.quantile(arr, 0.95)
, or evenskimage.filters.threshold_otsu()
. iter_callbackCallable[[float], Any], optional
A function that will be called on the threshold at every iteration of the algorithm.
 Returns:
 thresholdfloat
Upper threshold value. All pixels with an intensity higher than this value are assumed to be foreground.
References
[1]Li C.H. and Lee C.K. (1993) “Minimum Cross Entropy Thresholding” Pattern Recognition, 26(4): 617625 DOI:10.1016/00313203(93)90115D
[2]Li C.H. and Tam P.K.S. (1998) “An Iterative Algorithm for Minimum Cross Entropy Thresholding” Pattern Recognition Letters, 18(8): 771776 DOI:10.1016/S01678655(98)000579
[3]Sezgin M. and Sankur B. (2004) “Survey over Image Thresholding Techniques and Quantitative Performance Evaluation” Journal of Electronic Imaging, 13(1): 146165 DOI:10.1117/1.1631315
[4]ImageJ AutoThresholder code, http://fiji.sc/wiki/index.php/Auto_Threshold
Examples
>>> from skimage.data import camera >>> image = cp.array(camera()) >>> thresh = threshold_li(image) >>> binary = image > thresh
 cucim.skimage.filters.threshold_local(image, block_size=3, method='gaussian', offset=0, mode='reflect', param=None, cval=0)#
Compute a threshold mask image based on local pixel neighborhood.
Also known as adaptive or dynamic thresholding. The threshold value is the weighted mean for the local neighborhood of a pixel subtracted by a constant. Alternatively the threshold can be determined dynamically by a given function, using the ‘generic’ method.
 Parameters:
 image(N, M[, …, P]) ndarray
Grayscale input image.
 block_sizeint or sequence of int
Odd size of pixel neighborhood which is used to calculate the threshold value (e.g. 3, 5, 7, …, 21, …).
 method{‘generic’, ‘gaussian’, ‘mean’, ‘median’}, optional
Method used to determine adaptive threshold for local neighborhood in weighted mean image.
‘generic’: use custom function (see
param
parameter)‘gaussian’: apply gaussian filter (see
param
parameter for custom sigma value)‘mean’: apply arithmetic mean filter
‘median’: apply median rank filter
By default the ‘gaussian’ method is used.
 offsetfloat, optional
Constant subtracted from weighted mean of neighborhood to calculate the local threshold value. Default offset is 0.
 mode{‘reflect’, ‘constant’, ‘nearest’, ‘mirror’, ‘wrap’}, optional
The mode parameter determines how the array borders are handled, where cval is the value when mode is equal to ‘constant’. Default is ‘reflect’.
 param{int, function}, optional
Either specify sigma for ‘gaussian’ method or function object for ‘generic’ method. This functions takes the flat array of local neighborhood as a single argument and returns the calculated threshold for the centre pixel.
 cvalfloat, optional
Value to fill past edges of input if mode is ‘constant’.
 Returns:
 threshold(N, M[, …, P]) ndarray
Threshold image. All pixels in the input image higher than the corresponding pixel in the threshold image are considered foreground.
References
[1]Gonzalez, R. C. and Wood, R. E. “Digital Image Processing (2nd Edition).” PrenticeHall Inc., 2002: 600–612. ISBN: 0201180758
Examples
>>> import cupy as cp >>> from skimage.data import camera >>> image = cp.array(camera()[:50, :50]) >>> binary_image1 = image > threshold_local(image, 15, 'mean')
 cucim.skimage.filters.threshold_mean(image)#
Return threshold value based on the mean of grayscale values.
 Parameters:
 image(N, M[, …, P]) ndarray
Grayscale input image.
 Returns:
 thresholdfloat
Upper threshold value. All pixels with an intensity higher than this value are assumed to be foreground.
References
[1]C. A. Glasbey, “An analysis of histogrambased thresholding algorithms,” CVGIP: Graphical Models and Image Processing, vol. 55, pp. 532537, 1993. DOI:10.1006/cgip.1993.1040
Examples
>>> from skimage.data import camera >>> image = cp.array(camera()) >>> thresh = threshold_mean(image) >>> binary = image > thresh
 cucim.skimage.filters.threshold_minimum(image=None, nbins=256, max_num_iter=10000, *, hist=None)#
Return threshold value based on minimum method.
The histogram of the input
image
is computed if not provided and smoothed until there are only two maxima. Then the minimum in between is the threshold value.Either image or hist must be provided. In case hist is given, the actual histogram of the image is ignored.
 Parameters:
 image(N, M[, …, P]) ndarray, optional
Grayscale input image.
 nbinsint, optional
Number of bins used to calculate histogram. This value is ignored for integer arrays.
 max_num_iterint, optional
Maximum number of iterations to smooth the histogram.
 histarray, or 2tuple of arrays, optional
Histogram to determine the threshold from and a corresponding array of bin center intensities. Alternatively, only the histogram can be passed.
 Returns:
 thresholdfloat
Upper threshold value. All pixels with an intensity higher than this value are assumed to be foreground.
 Other Parameters:
 max_iterDEPRECATED
Deprecated in favor of max_num_iter.
Deprecated since version 22.02.00.
 Raises:
 RuntimeError
If unable to find two local maxima in the histogram or if the smoothing takes more than 1e4 iterations.
References
[1]C. A. Glasbey, “An analysis of histogrambased thresholding algorithms,” CVGIP: Graphical Models and Image Processing, vol. 55, pp. 532537, 1993.
[2]Prewitt, JMS & Mendelsohn, ML (1966), “The analysis of cell images”, Annals of the New York Academy of Sciences 128: 10351053 DOI:10.1111/j.17496632.1965.tb11715.x
Examples
>>> from skimage.data import camera >>> image = cp.array(camera()) >>> thresh = threshold_minimum(image) >>> binary = image > thresh
 cucim.skimage.filters.threshold_multiotsu(image=None, classes=3, nbins=256, *, hist=None)#
Generate classes1 threshold values to divide gray levels in image, following Otsu’s method for multiple classes.
The threshold values are chosen to maximize the total sum of pairwise variances between the thresholded graylevel classes. See Notes and [1] for more details.
Either image or hist must be provided. If hist is provided, the actual histogram of the image is ignored.
 Parameters:
 image(N, M[, …, P]) ndarray, optional
Grayscale input image.
 classesint, optional
Number of classes to be thresholded, i.e. the number of resulting regions.
 nbinsint, optional
Number of bins used to calculate the histogram. This value is ignored for integer arrays.
 histarray, or 2tuple of arrays, optional
Histogram from which to determine the threshold, and optionally a corresponding array of bin center intensities. If no hist provided, this function will compute it from the image (see notes).
 Returns:
 thresharray
Array containing the threshold values for the desired classes.
 Raises:
 ValueError
If
image
contains less grayscale value then the desired number of classes.
Notes
This implementation relies on a Cython function whose complexity is \(O\left(\frac{Ch^{C1}}{(C1)!}\right)\), where \(h\) is the number of histogram bins and \(C\) is the number of classes desired.
If no hist is given, this function will make use of skimage.exposure.histogram, which behaves differently than np.histogram. While both allowed, use the former for consistent behaviour.
The input image must be grayscale.
References
[1]Liao, PS., Chen, TS. and Chung, PC., “A fast algorithm for multilevel thresholding”, Journal of Information Science and Engineering 17 (5): 713727, 2001. Available at: <https://ftp.iis.sinica.edu.tw/JISE/2001/200109_01.pdf> DOI:10.6688/JISE.2001.17.5.1
[2]Tosa, Y., “MultiOtsu Threshold”, a java plugin for ImageJ. Available at: <http://imagej.net/plugins/download/Multi_OtsuThreshold.java>
Examples
>>> import cupy as cp >>> from cucim.skimage.color import label2rgb >>> from skimage import data >>> image = cp.asarray(data.camera()) >>> thresholds = threshold_multiotsu(image) >>> regions = cp.digitize(image, bins=thresholds) >>> regions_colorized = label2rgb(regions)
 cucim.skimage.filters.threshold_niblack(image, window_size=15, k=0.2)#
Applies Niblack local threshold to an array.
A threshold T is calculated for every pixel in the image using the following formula:
T = m(x,y)  k * s(x,y)
where m(x,y) and s(x,y) are the mean and standard deviation of pixel (x,y) neighborhood defined by a rectangular window with size w times w centered around the pixel. k is a configurable parameter that weights the effect of standard deviation.
 Parameters:
 image(N, M[, …, P]) ndarray
Grayscale input image.
 window_sizeint, or iterable of int, optional
Window size specified as a single odd integer (3, 5, 7, …), or an iterable of length
image.ndim
containing only odd integers (e.g.(1, 5, 5)
). kfloat, optional
Value of parameter k in threshold formula.
 Returns:
 threshold(N, M) ndarray
Threshold mask. All pixels with an intensity higher than this value are assumed to be foreground.
Notes
This algorithm is originally designed for text recognition.
The Bradley threshold is a particular case of the Niblack one, being equivalent to
>>> from skimage import data >>> image = cp.array(data.page()) >>> q = 1 >>> threshold_image = threshold_niblack(image, k=0) * q
for some value
q
. By default, Bradley and Roth useq=1
.References
[1]W. Niblack, An introduction to Digital Image Processing, PrenticeHall, 1986.
[2]D. Bradley and G. Roth, “Adaptive thresholding using Integral Image”, Journal of Graphics Tools 12(2), pp. 1321, 2007. DOI:10.1080/2151237X.2007.10129236
Examples
>>> from skimage import data >>> image = cp.array(data.page()) >>> threshold_image = threshold_niblack(image, window_size=7, k=0.1)
 cucim.skimage.filters.threshold_otsu(image=None, nbins=256, *, hist=None)#
Return threshold value based on Otsu’s method.
Either image or hist must be provided. If hist is provided, the actual histogram of the image is ignored.
 Parameters:
 image(N, M[, …, P]) ndarray, optional
Grayscale input image.
 nbinsint, optional
Number of bins used to calculate histogram. This value is ignored for integer arrays.
 histarray, or 2tuple of arrays, optional
Histogram from which to determine the threshold, and optionally a corresponding array of bin center intensities. If no hist provided, this function will compute it from the image.
 Returns:
 thresholdfloat
Upper threshold value. All pixels with an intensity higher than this value are assumed to be foreground.
Notes
The input image must be grayscale.
References
[1]Wikipedia, https://en.wikipedia.org/wiki/Otsu’s_Method
Examples
>>> from skimage.data import camera >>> image = cp.array(camera()) >>> thresh = threshold_otsu(image) >>> binary = image <= thresh
 cucim.skimage.filters.threshold_sauvola(image, window_size=15, k=0.2, r=None)#
Applies Sauvola local threshold to an array. Sauvola is a modification of Niblack technique.
In the original method a threshold T is calculated for every pixel in the image using the following formula:
T = m(x,y) * (1 + k * ((s(x,y) / R)  1))
where m(x,y) and s(x,y) are the mean and standard deviation of pixel (x,y) neighborhood defined by a rectangular window with size w times w centered around the pixel. k is a configurable parameter that weights the effect of standard deviation. R is the maximum standard deviation of a grayscale image.
 Parameters:
 image(N, M[, …, P]) ndarray
Grayscale input image.
 window_sizeint, or iterable of int, optional
Window size specified as a single odd integer (3, 5, 7, …), or an iterable of length
image.ndim
containing only odd integers (e.g.(1, 5, 5)
). kfloat, optional
Value of the positive parameter k.
 rfloat, optional
Value of R, the dynamic range of standard deviation. If None, set to the half of the image dtype range.
 Returns:
 threshold(N, M) ndarray
Threshold mask. All pixels with an intensity higher than this value are assumed to be foreground.
Notes
This algorithm is originally designed for text recognition.
References
[1]J. Sauvola and M. Pietikainen, “Adaptive document image binarization,” Pattern Recognition 33(2), pp. 225236, 2000. DOI:10.1016/S00313203(99)000552
Examples
>>> from skimage import data >>> image = cp.array(data.page()) >>> t_sauvola = threshold_sauvola(image, window_size=15, k=0.2) >>> binary_image = image > t_sauvola
 cucim.skimage.filters.threshold_triangle(image, nbins=256)#
Return threshold value based on the triangle algorithm.
 Parameters:
 image(N, M[, …, P]) ndarray
Grayscale input image.
 nbinsint, optional
Number of bins used to calculate histogram. This value is ignored for integer arrays.
 Returns:
 thresholdfloat
Upper threshold value. All pixels with an intensity higher than this value are assumed to be foreground.
References
[1]Zack, G. W., Rogers, W. E. and Latt, S. A., 1977, Automatic Measurement of Sister Chromatid Exchange Frequency, Journal of Histochemistry and Cytochemistry 25 (7), pp. 741753 DOI:10.1177/25.7.70454
[2]ImageJ AutoThresholder code, http://fiji.sc/wiki/index.php/Auto_Threshold
Examples
>>> from skimage.data import camera >>> image = cp.array(camera()) >>> thresh = threshold_triangle(image) >>> binary = image > thresh
 cucim.skimage.filters.threshold_yen(image=None, nbins=256, *, hist=None)#
Return threshold value based on Yen’s method. Either image or hist must be provided. In case hist is given, the actual histogram of the image is ignored.
 Parameters:
 image(N, M[, …, P]) ndarray
Grayscale input image.
 nbinsint, optional
Number of bins used to calculate histogram. This value is ignored for integer arrays.
 histarray, or 2tuple of arrays, optional
Histogram from which to determine the threshold, and optionally a corresponding array of bin center intensities. An alternative use of this function is to pass it only hist.
 Returns:
 thresholdfloat
Upper threshold value. All pixels with an intensity higher than this value are assumed to be foreground.
References
[1]Yen J.C., Chang F.J., and Chang S. (1995) “A New Criterion for Automatic Multilevel Thresholding” IEEE Trans. on Image Processing, 4(3): 370378. DOI:10.1109/83.366472
[2]Sezgin M. and Sankur B. (2004) “Survey over Image Thresholding Techniques and Quantitative Performance Evaluation” Journal of Electronic Imaging, 13(1): 146165, DOI:10.1117/1.1631315 http://www.busim.ee.boun.edu.tr/~sankur/SankurFolder/Threshold_survey.pdf
[3]ImageJ AutoThresholder code, http://fiji.sc/wiki/index.php/Auto_Threshold
Examples
>>> from skimage.data import camera >>> image = cp.array(camera()) >>> thresh = threshold_yen(image) >>> binary = image <= thresh
 cucim.skimage.filters.try_all_threshold(image, figsize=(8, 5), verbose=True)#
Returns a figure comparing the outputs of different thresholding methods.
 Parameters:
 image(N, M) ndarray
Input image.
 figsizetuple, optional
Figure size (in inches).
 verbosebool, optional
Print function name for each method.
 Returns:
 fig, axtuple
Matplotlib figure and axes.
Notes
The following algorithms are used:
isodata
li
mean
minimum
otsu
triangle
yen
Examples
>>> from skimage.data import text >>> text_img = cp.array(text()) >>> fig, ax = try_all_threshold(text_img, figsize=(10, 6), verbose=False)
 cucim.skimage.filters.unsharp_mask(image, radius=1.0, amount=1.0, preserve_range=False, *, channel_axis=None)#
Unsharp masking filter.
The sharp details are identified as the difference between the original image and its blurred version. These details are then scaled, and added back to the original image.
 Parameters:
 image[P, …, ]M[, N][, C] ndarray
Input image.
 radiusscalar or sequence of scalars, optional
If a scalar is given, then its value is used for all dimensions. If sequence is given, then there must be exactly one radius for each dimension except the last dimension for multichannel images. Note that 0 radius means no blurring, and negative values are not allowed.
 amountscalar, optional
The details will be amplified with this factor. The factor could be 0 or negative. Typically, it is a small positive number, e.g. 1.0.
 preserve_rangebool, optional
Whether to keep the original range of values. Otherwise, the input image is converted according to the conventions of
img_as_float
. Also see https://scikitimage.org/docs/dev/user_guide/data_types.html channel_axisint or None, optional
If None, the image is assumed to be a grayscale (single channel) image. Otherwise, this parameter indicates which axis of the array corresponds to channels.
 Returns:
 output[P, …, ]M[, N][, C] ndarray of float
Image with unsharp mask applied.
Notes
Unsharp masking is an image sharpening technique. It is a linear image operation, and numerically stable, unlike deconvolution which is an illposed problem. Because of this stability, it is often preferred over deconvolution.
The main idea is as follows: sharp details are identified as the difference between the original image and its blurred version. These details are added back to the original image after a scaling step:
enhanced image = original + amount * (original  blurred)
When applying this filter to several color layers independently, color bleeding may occur. More visually pleasing result can be achieved by processing only the brightness/lightness/intensity channel in a suitable color space such as HSV, HSL, YUV, or YCbCr.
Unsharp masking is described in most introductory digital image processing books. This implementation is based on [1].
References
[1]Maria Petrou, Costas Petrou “Image Processing: The Fundamentals”, (2010), ed ii., page 357, ISBN 13: 9781119994398 DOI:10.1002/9781119994398
[2]Wikipedia. Unsharp masking https://en.wikipedia.org/wiki/Unsharp_masking
Examples
>>> import cupy as cp >>> array = cp.ones(shape=(5,5), dtype=cp.uint8)*100 >>> array[2,2] = 120 >>> array array([[100, 100, 100, 100, 100], [100, 100, 100, 100, 100], [100, 100, 120, 100, 100], [100, 100, 100, 100, 100], [100, 100, 100, 100, 100]], dtype=uint8) >>> cp.around(unsharp_mask(array, radius=0.5, amount=2),2) array([[0.39, 0.39, 0.39, 0.39, 0.39], [0.39, 0.39, 0.38, 0.39, 0.39], [0.39, 0.38, 0.53, 0.38, 0.39], [0.39, 0.39, 0.38, 0.39, 0.39], [0.39, 0.39, 0.39, 0.39, 0.39]])
>>> array = cp.ones(shape=(5,5), dtype=cp.int8)*100 >>> array[2,2] = 127 >>> cp.around(unsharp_mask(array, radius=0.5, amount=2),2) array([[0.79, 0.79, 0.79, 0.79, 0.79], [0.79, 0.78, 0.75, 0.78, 0.79], [0.79, 0.75, 1. , 0.75, 0.79], [0.79, 0.78, 0.75, 0.78, 0.79], [0.79, 0.79, 0.79, 0.79, 0.79]])
>>> cp.around(unsharp_mask(array, radius=0.5, amount=2, ... preserve_range=True), ... 2) array([[100. , 100. , 99.99, 100. , 100. ], [100. , 99.39, 95.48, 99.39, 100. ], [ 99.99, 95.48, 147.59, 95.48, 99.99], [100. , 99.39, 95.48, 99.39, 100. ], [100. , 100. , 99.99, 100. , 100. ]])
 cucim.skimage.filters.wiener(data, impulse_response=None, filter_params=None, K=0.25, predefined_filter=None)#
Minimum Mean Square Error (Wiener) inverse filter.
 Parameters:
 data(M,N) ndarray
Input data.
 Kfloat or (M,N) ndarray
Ratio between power spectrum of noise and undegraded image.
 impulse_responsecallable f(r, c, **filter_params)
Impulse response of the filter. See LPIFilter2D.__init__.
 filter_paramsdict, optional
Additional keyword parameters to the impulse_response function.
 Other Parameters:
 predefined_filterLPIFilter2D
If you need to apply the same filter multiple times over different images, construct the LPIFilter2D and specify it here.
 cucim.skimage.filters.window(window_type, shape, warp_kwargs=None)#
Return an ndimensional window of a given size and dimensionality.
 Parameters:
 window_typestring, float, or tuple
The type of window to be created. Any window type supported by
scipy.signal.get_window
is allowed here. See notes below for a current list, or the SciPy documentation for the version of SciPy on your machine. shapetuple of int or int
The shape of the window along each axis. If an integer is provided, a 1D window is generated.
 warp_kwargsdict
Keyword arguments passed to skimage.transform.warp (e.g.,
warp_kwargs={'order':3}
to change interpolation method).
 Returns:
 nd_windowndarray
A window of the specified
shape
.dtype
isnp.float64
.
Notes
This function is based on
scipy.signal.get_window
and thus can access all of the window types available to that function (e.g.,"hann"
,"boxcar"
). Note that certain window types require parameters that have to be supplied with the window name as a tuple (e.g.,("tukey", 0.8)
). If only a float is supplied, it is interpreted as the beta parameter of the Kaiser window.See https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.windows.get_window.html for more details.
Note that this function generates a double precision array of the specified
shape
and can thus generate very large arrays that consume a large amount of available memory.The approach taken here to create nD windows is to first calculate the Euclidean distance from the center of the intended nD window to each position in the array. That distance is used to sample, with interpolation, from a 1D window returned from
scipy.signal.get_window
. The method of interpolation can be changed with theorder
keyword argument passed to skimage.transform.warp.Some coordinates in the output window will be outside of the original signal; these will be filled in with zeros.
Window types:  boxcar  triang  blackman  hamming  hann  bartlett  flattop  parzen  bohman  blackmanharris  nuttall  barthann  kaiser (needs beta)  gaussian (needs standard deviation)  general_gaussian (needs power, width)  slepian (needs width)  dpss (needs normalized halfbandwidth)  chebwin (needs attenuation)  exponential (needs decay scale)  tukey (needs taper fraction)
References
[1]Twodimensional window design, Wikipedia, https://en.wikipedia.org/wiki/Two_dimensional_window_design
Examples
Return a Hann window with shape (512, 512):
>>> from cucim.skimage.filters import window >>> w = window('hann', (512, 512))
Return a Kaiser window with beta parameter of 16 and shape (256, 256, 35):
>>> w = window(16, (256, 256, 35))
Return a Tukey window with an alpha parameter of 0.8 and shape (100, 300):
>>> w = window(('tukey', 0.8), (100, 300))
measure#
 cucim.skimage.measure.approximate_polygon(coords, tolerance)#
Approximate a polygonal chain with the specified tolerance.
It is based on the DouglasPeucker algorithm.
Note that the approximated polygon is always within the convex hull of the original polygon.
 Parameters:
 coords(N, 2) array
Coordinate array.
 tolerancefloat
Maximum distance from original points of polygon to approximated polygonal chain. If tolerance is 0, the original coordinate array is returned.
 Returns:
 coords(M, 2) array
Approximated polygonal chain where M <= N.
References
 cucim.skimage.measure.block_reduce(image, block_size=2, func=<function sum>, cval=0, func_kwargs=None)#
Downsample image by applying function func to local blocks.
This function is useful for max and mean pooling, for example.
 Parameters:
 imagendarray
Ndimensional input image.
 block_sizearray_like or int
Array containing downsampling integer factor along each axis. Default block_size is 2.
 funccallable
Function object which is used to calculate the return value for each local block. This function must implement an
axis
parameter. Primary functions arenumpy.sum
,numpy.min
,numpy.max
,numpy.mean
andnumpy.median
. See also func_kwargs. cvalfloat
Constant padding value if image is not perfectly divisible by the block size.
 func_kwargsdict
Keyword arguments passed to func. Notably useful for passing dtype argument to
np.mean
. Takes dictionary of inputs, e.g.:func_kwargs={'dtype': np.float16})
.
 Returns:
 imagendarray
Downsampled image with same number of dimensions as input image.
Examples
>>> import cupy as cp >>> from cucim.skimage.measure import block_reduce >>> image = cp.arange(3*3*4).reshape(3, 3, 4) >>> image array([[[ 0, 1, 2, 3], [ 4, 5, 6, 7], [ 8, 9, 10, 11]], [[12, 13, 14, 15], [16, 17, 18, 19], [20, 21, 22, 23]], [[24, 25, 26, 27], [28, 29, 30, 31], [32, 33, 34, 35]]]) >>> block_reduce(image, block_size=(3, 3, 1), func=cp.mean) array([[[16., 17., 18., 19.]]]) >>> image_max1 = block_reduce(image, block_size=(1, 3, 4), func=cp.max) >>> image_max1 array([[[11]], [[23]], [[35]]]) >>> image_max2 = block_reduce(image, block_size=(3, 1, 4), func=cp.max) >>> image_max2 array([[[27], [31], [35]]])
 cucim.skimage.measure.blur_effect(image, h_size=11, channel_axis=None, reduce_func=<builtin function max>)#
Compute a metric that indicates the strength of blur in an image (0 for no blur, 1 for maximal blur).
 Parameters:
 imagendarray
RGB or grayscale nD image. The input image is converted to grayscale before computing the blur metric.
 h_sizeint, optional
Size of the reblurring filter.
 channel_axisint or None, optional
If None, the image is assumed to be grayscale (singlechannel). Otherwise, this parameter indicates which axis of the array corresponds to color channels.
 reduce_funccallable, optional
Function used to calculate the aggregation of blur metrics along all axes. If set to None, the entire list is returned, where the ith element is the blur metric along the ith axis. This function should be a host function that operates on standard python floats.
 Returns:
 blurfloat (0 to 1) or list of floats
Blur metric: by default, the maximum of blur metrics along all axes.
Notes
h_size must keep the same value in order to compare results between images. Most of the time, the default size (11) is enough. This means that the metric can clearly discriminate blur up to an average 11x11 filter; if blur is higher, the metric still gives good results but its values tend towards an asymptote.
References
[1]Frederique Crete, Thierry Dolmiere, Patricia Ladret, and Marina Nicolas “The blur effect: perception and estimation with a new noreference perceptual blur metric” Proc. SPIE 6492, Human Vision and Electronic Imaging XII, 64920I (2007) https://hal.archivesouvertes.fr/hal00232709 DOI:10.1117/12.702790
 cucim.skimage.measure.centroid(image, *, spacing=None)#
Return the (weighted) centroid of an image.
 Parameters:
 imagearray
The input image.
 spacing: tuple of float, shape (ndim, )
The pixel spacing along each axis of the image.
 Returns:
 centertuple of float, length
image.ndim
The centroid of the (nonzero) pixels in
image
.
 centertuple of float, length
Examples
>>> import cupy as cp >>> from cucim.skimage.measure import centroid >>> image = cp.zeros((20, 20), dtype=cp.float64) >>> image[13:17, 13:17] = 0.5 >>> image[10:12, 10:12] = 1 >>> centroid(image) array([13.16666667, 13.16666667])
 cucim.skimage.measure.inertia_tensor(image, mu=None, *, spacing=None)#
Compute the inertia tensor of the input image.
 Parameters:
 imagearray
The input image.
 muarray, optional
The precomputed central moments of
image
. The inertia tensor computation requires the central moments of the image. If an application requires both the central moments and the inertia tensor (for example, skimage.measure.regionprops), then it is more efficient to precompute them and pass them to the inertia tensor call. spacingtuple of float, optional
The pixel spacing along each axis of the image.
 Returns:
 Tarray, shape
(image.ndim, image.ndim)
The inertia tensor of the input image. \(T_{i, j}\) contains the covariance of image intensity along axes \(i\) and \(j\).
 Tarray, shape
References
[2]Bernd Jähne. SpatioTemporal Image Processing: Theory and Scientific Applications. (Chapter 8: Tensor Methods) Springer, 1993.
 cucim.skimage.measure.inertia_tensor_eigvals(image, mu=None, T=None, *, spacing=None)#
Compute the eigenvalues of the inertia tensor of the image.
The inertia tensor measures covariance of the image intensity along the image axes. (See inertia_tensor.) The relative magnitude of the eigenvalues of the tensor is thus a measure of the elongation of a (bright) object in the image.
 Parameters:
 imagearray
The input image.
 muarray, optional
The precomputed central moments of
image
. Tarray, shape
(image.ndim, image.ndim)
The precomputed inertia tensor. If
T
is given,mu
andimage
are ignored. spacingtuple of float, optional
The pixel spacing along each axis of the image.
 Returns:
 eigvalslist of float, length
image.ndim
The eigenvalues of the inertia tensor of
image
, in descending order.
 eigvalslist of float, length
Notes
Computing the eigenvalues requires the inertia tensor of the input image. This is much faster if the central moments (
mu
) are provided, or, alternatively, one can provide the inertia tensor (T
) directly.
 cucim.skimage.measure.intersection_coeff(image0_mask, image1_mask, mask=None)#
Fraction of a channel’s segmented binary mask that overlaps with a second channel’s segmented binary mask.
 Parameters:
 image0_mask(M, N) ndarray of dtype bool
Image mask of channel A.
 image1_mask(M, N) ndarray of dtype bool
Image mask of channel B. Must have same dimensions as image0_mask.
 mask(M, N) ndarray of dtype bool, optional
Only image0_mask and image1_mask pixels within this region of interest mask are included in the calculation. Must have same dimensions as image0_mask.
 Returns:
 Intersection coefficient, float
Fraction of image0_mask that overlaps with image1_mask.
 cucim.skimage.measure.label(label_image, background=None, return_num=False, connectivity=None)#
Label connected regions of an integer array.
Two pixels are connected when they are neighbors and have the same value. In 2D, they can be neighbors either in a 1 or 2connected sense. The value refers to the maximum number of orthogonal hops to consider a pixel/voxel a neighbor:
1connectivity 2connectivity diagonal connection closeup [ ] [ ] [ ] [ ] [ ]  \  /  < hop 2 [ ][x][ ] [ ][x][ ] [x][ ]  /  \ hop 1 [ ] [ ] [ ] [ ]
 Parameters:
 label_imagendarray of dtype int
Image to label.
 backgroundint, optional
Consider all pixels with this value as background pixels, and label them as 0. By default, 0valued pixels are considered as background pixels.
 return_numbool, optional
Whether to return the number of assigned labels.
 connectivityint, optional
Maximum number of orthogonal hops to consider a pixel/voxel as a neighbor. Accepted values are ranging from 1 to input.ndim. If
None
, a full connectivity ofinput.ndim
is used.
 Returns:
 labelsndarray of dtype int
Labeled array, where all connected regions are assigned the same integer value.
 numint, optional
Number of labels, which equals the maximum label index and is only returned if return_num is True.
See also
Notes
Currently the cucim implementation of this function always uses 32bit integers for the label array. This is done for performance. In the future 64bit integer support may also be added for better skimage compatibility.
References
[1]Christophe Fiorio and Jens Gustedt, “Two linear time UnionFind strategies for image processing”, Theoretical Computer Science 154 (1996), pp. 165181.
[2]Kensheng Wu, Ekow Otoo and Arie Shoshani, “Optimizing connected component labeling algorithms”, Paper LBNL56864, 2005, Lawrence Berkeley National Laboratory (University of California), http://repositories.cdlib.org/lbnl/LBNL56864
Examples
>>> import cupy as cp >>> x = cp.eye(3).astype(int) >>> print(x) [[1 0 0] [0 1 0] [0 0 1]] >>> print(label(x, connectivity=1)) [[1 0 0] [0 2 0] [0 0 3]] >>> print(label(x, connectivity=2)) [[1 0 0] [0 1 0] [0 0 1]] >>> print(label(x, background=1)) [[1 2 2] [2 1 2] [2 2 1]] >>> x = cp.asarray([[1, 0, 0], ... [1, 1, 5], ... [0, 0, 0]]) >>> print(label(x)) [[1 0 0] [1 1 2] [0 0 0]]
 cucim.skimage.measure.manders_coloc_coeff(image0, image1_mask, mask=None)#
Manders’ colocalization coefficient between two channels.
 Parameters:
 image0(M, N) ndarray
Image of channel A. All pixel values should be nonnegative.
 image1_mask(M, N) ndarray of dtype bool
Binary mask with segmented regions of interest in channel B. Must have same dimensions as image0.
 mask(M, N) ndarray of dtype bool, optional
Only image0 pixel values within this region of interest mask are included in the calculation. Must have same dimensions as image0.
 Returns:
 mccfloat
Manders’ colocalization coefficient.
Notes
Manders’ Colocalization Coefficient (MCC) is the fraction of total intensity of a certain channel (channel A) that is within the segmented region of a second channel (channel B) [1]. It ranges from 0 for no colocalisation to 1 for complete colocalization. It is also referred to as M1 and M2.
MCC is commonly used to measure the colocalization of a particular protein in a subceullar compartment. Typically a segmentation mask for channel B is generated by setting a threshold that the pixel values must be above to be included in the MCC calculation. In this implementation, the channel B mask is provided as the argument image1_mask, allowing the exact segmentation method to be decided by the user beforehand.
The implemented equation is:
\[r = \frac{\sum A_{i,coloc}}{\sum A_i}\] where
\(A_i\) is the value of the \(i^{th}\) pixel in image0 \(A_{i,coloc} = A_i\) if \(Bmask_i > 0\) \(Bmask_i\) is the value of the \(i^{th}\) pixel in mask
MCC is sensitive to noise, with diffuse signal in the first channel inflating its value. Images should be processed to remove out of focus and background light before the MCC is calculated [2].
References
[1]Manders, E.M.M., Verbeek, F.J. and Aten, J.A. (1993), Measurement of colocalization of objects in dualcolour confocal images. Journal of Microscopy, 169: 375382. https://doi.org/10.1111/j.13652818.1993.tb03313.x https://imagej.net/media/manders.pdf
[2]Dunn, K. W., Kamocka, M. M., & McDonald, J. H. (2011). A practical guide to evaluating colocalization in biological microscopy. American journal of physiology. Cell physiology, 300(4), C723–C742. https://doi.org/10.1152/ajpcell.00462.2010
 cucim.skimage.measure.manders_overlap_coeff(image0, image1, mask=None)#
Manders’ overlap coefficient
 Parameters:
 image0(M, N) ndarray
Image of channel A. All pixel values should be nonnegative.
 image1(M, N) ndarray
Image of channel B. All pixel values should be nonnegative. Must have same dimensions as image0
 mask(M, N) ndarray of dtype bool, optional
Only image0 and image1 pixel values within this region of interest mask are included in the calculation. Must have same dimensions as image0.
 Returns:
 moc: float
Manders’ Overlap Coefficient of pixel intensities between the two images.
Notes
Manders’ Overlap Coefficient (MOC) is given by the equation [1]:
\[r = \frac{\sum A_i B_i}{\sqrt{\sum A_i^2 \sum B_i^2}}\] where
\(A_i\) is the value of the \(i^{th}\) pixel in image0 \(B_i\) is the value of the \(i^{th}\) pixel in image1
It ranges between 0 for no colocalization and 1 for complete colocalization of all pixels.
MOC does not take into account pixel intensities, just the fraction of pixels that have positive values for both channels[Rb497c61262632]_ [3]. Its usefulness has been criticized as it changes in response to differences in both cooccurence and correlation and so a particular MOC value could indicate a wide range of colocalization patterns [4] [5].
References
[1]Manders, E.M.M., Verbeek, F.J. and Aten, J.A. (1993), Measurement of colocalization of objects in dualcolour confocal images. Journal of Microscopy, 169: 375382. https://doi.org/10.1111/j.13652818.1993.tb03313.x https://imagej.net/media/manders.pdf
[2]Dunn, K. W., Kamocka, M. M., & McDonald, J. H. (2011). A practical guide to evaluating colocalization in biological microscopy. American journal of physiology. Cell physiology, 300(4), C723–C742. https://doi.org/10.1152/ajpcell.00462.2010
[3]Bolte, S. and Cordelières, F.P. (2006), A guided tour into subcellular colocalization analysis in light microscopy. Journal of Microscopy, 224: 213232. https://doi.org/10.1111/j.13652818.2006.01
[4]Adler J, Parmryd I. (2010), Quantifying colocalization by correlation: the Pearson correlation coefficient is superior to the Mander’s overlap coefficient. Cytometry A. Aug;77(8):73342.https://doi.org/10.1002/cyto.a.20896
[5]Adler, J, Parmryd, I. Quantifying colocalization: The case for discarding the Manders overlap coefficient. Cytometry. 2021; 99: 910– 920. https://doi.org/10.1002/cyto.a.24336
 cucim.skimage.measure.moments(image, order=3, *, spacing=None)#
Calculate all raw image moments up to a certain order.
 The following properties can be calculated from raw image moments:
Area as:
M[0, 0]
.Centroid as: {
M[1, 0] / M[0, 0]
,M[0, 1] / M[0, 0]
}.
Note that raw moments are neither translation, scale nor rotation invariant.
 Parameters:
 imagenD double or uint8 array
Rasterized shape as image.
 orderint, optional
Maximum order of moments. Default is 3.
 spacing: tuple of float, shape (ndim, )
The pixel spacing along each axis of the image.
 Returns:
 m(
order + 1
,order + 1
) array Raw image moments.
 m(
References
[1]Wilhelm Burger, Mark Burge. Principles of Digital Image Processing: Core Algorithms. SpringerVerlag, London, 2009.
[2]B. Jähne. Digital Image Processing. SpringerVerlag, BerlinHeidelberg, 6. edition, 2005.
[3]T. H. Reiss. Recognizing Planar Objects Using Invariant Image Features, from Lecture notes in computer science, p. 676. Springer, Berlin, 1993.
Examples
>>> import cupy as cp >>> from cucim.skimage.measure import moments >>> image = cp.zeros((20, 20), dtype=cp.float64) >>> image[13:17, 13:17] = 1 >>> M = moments(image) >>> centroid = (M[1, 0] / M[0, 0], M[0, 1] / M[0, 0]) >>> centroid (array(14.5), array(14.5))
 cucim.skimage.measure.moments_central(image, center=None, order=3, *, spacing=None, **kwargs)#
Calculate all central image moments up to a certain order.
The center coordinates (cr, cc) can be calculated from the raw moments as: {
M[1, 0] / M[0, 0]
,M[0, 1] / M[0, 0]
}.Note that central moments are translation invariant but not scale and rotation invariant.
 Parameters:
 imagenD double or uint8 array
Rasterized shape as image.
 centertuple of float, optional
Coordinates of the image centroid. This will be computed if it is not provided.
 orderint, optional
The maximum order of moments computed.
 spacing: tuple of float, shape (ndim, )
The pixel spacing along each axis of the image.
 Returns:
 mu(
order + 1
,order + 1
) array Central image moments.
 mu(
References
[1]Wilhelm Burger, Mark Burge. Principles of Digital Image Processing: Core Algorithms. SpringerVerlag, London, 2009.
[2]B. Jähne. Digital Image Processing. SpringerVerlag, BerlinHeidelberg, 6. edition, 2005.
[3]T. H. Reiss. Recognizing Planar Objects Using Invariant Image Features, from Lecture notes in computer science, p. 676. Springer, Berlin, 1993.
Examples
>>> import cupy as cp >>> from cucim.skimage.measure import moments, moments_central >>> image = cp.zeros((20, 20), dtype=cp.float64) >>> image[13:17, 13:17] = 1 >>> M = moments(image) >>> centroid = (M[1, 0] / M[0, 0], M[0, 1] / M[0, 0]) >>> moments_central(image, centroid) array([[16., 0., 20., 0.], [ 0., 0., 0., 0.], [20., 0., 25., 0.], [ 0., 0., 0., 0.]])
 cucim.skimage.measure.moments_coords(coords, order=3)#
Calculate all raw image moments up to a certain order.
 The following properties can be calculated from raw image moments:
Area as:
M[0, 0]
.Centroid as: {
M[1, 0] / M[0, 0]
,M[0, 1] / M[0, 0]
}.
Note that raw moments are neither translation, scale nor rotation invariant.
 Parameters:
 coords(N, D) double or uint8 array
Array of N points that describe an image of D dimensionality in Cartesian space.
 orderint, optional
Maximum order of moments. Default is 3.
 Returns:
 M(
order + 1
,order + 1
, …) array Raw image moments. (D dimensions)
 M(
References
[1]Johannes Kilian. Simple Image Analysis By Moments. Durham University, version 0.2, Durham, 2001.
Examples
>>> import cupy as cp >>> from cucim.skimage.measure import moments_coords >>> coords = cp.array([[row, col] ... for row in range(13, 17) ... for col in range(14, 18)], dtype=cp.float64) >>> M = moments_coords(coords) >>> centroid = (M[1, 0] / M[0, 0], M[0, 1] / M[0, 0]) >>> centroid (array(14.5), array(15.5))
 cucim.skimage.measure.moments_coords_central(coords, center=None, order=3)#
Calculate all central image moments up to a certain order.
 The following properties can be calculated from raw image moments:
Area as:
M[0, 0]
.Centroid as: {
M[1, 0] / M[0, 0]
,M[0, 1] / M[0, 0]
}.
Note that raw moments are neither translation, scale nor rotation invariant.
 Parameters:
 coords(N, D) double or uint8 array
Array of N points that describe an image of D dimensionality in Cartesian space. A tuple of coordinates as returned by
cp.nonzero
is also accepted as input. centertuple of float, optional
Coordinates of the image centroid. This will be computed if it is not provided.
 orderint, optional
Maximum order of moments. Default is 3.
 Returns:
 Mc(
order + 1
,order + 1
, …) array Central image moments. (D dimensions)
 Mc(
References
[1]Johannes Kilian. Simple Image Analysis By Moments. Durham University, version 0.2, Durham, 2001.
Examples
>>> import cupy as cp >>> from cucim.skimage.measure import moments_coords_central >>> coords = cp.array([[row, col] ... for row in range(13, 17) ... for col in range(14, 18)]) >>> moments_coords_central(coords) array([[16., 0., 20., 0.], [ 0., 0., 0., 0.], [20., 0., 25., 0.], [ 0., 0., 0., 0.]])
As seen above, for symmetric objects, oddorder moments (columns 1 and 3, rows 1 and 3) are zero when centered on the centroid, or center of mass, of the object (the default). If we break the symmetry by adding a new point, this no longer holds:
>>> coords2 = cp.concatenate((coords, cp.array([[17, 17]])), axis=0) >>> cp.around(moments_coords_central(coords2), ... decimals=2) array([[17. , 0. , 22.12, 2.49], [ 0. , 3.53, 1.73, 7.4 ], [25.88, 6.02, 36.63, 8.83], [ 4.15, 19.17, 14.8 , 39.6 ]])
Image moments and central image moments are equivalent (by definition) when the center is (0, 0):
>>> cp.allclose(moments_coords(coords), ... moments_coords_central(coords, (0, 0))) array(True)
 cucim.skimage.measure.moments_hu(nu)#
Calculate Hu’s set of image moments (2Donly).
Note that this set of moments is proved to be translation, scale and rotation invariant.
 Parameters:
 nu(M, M) array
Normalized central image moments, where M must be >= 4.
 Returns:
 nu(7,) array
Hu’s set of image moments.
Notes
Due to the small array sizes, this function will be faster on the CPU. Consider transfering
nu
to the host and runningskimage.measure.moments_hu
if the moments are not needed on the device.References
[1]M. K. Hu, “Visual Pattern Recognition by Moment Invariants”, IRE Trans. Info. Theory, vol. IT8, pp. 179187, 1962
[2]Wilhelm Burger, Mark Burge. Principles of Digital Image Processing: Core Algorithms. SpringerVerlag, London, 2009.
[3]B. Jähne. Digital Image Processing. SpringerVerlag, BerlinHeidelberg, 6. edition, 2005.
[4]T. H. Reiss. Recognizing Planar Objects Using Invariant Image Features, from Lecture notes in computer science, p. 676. Springer, Berlin, 1993.
Examples
>>> import cupy as cp >>> from cucim.skimage.measure import (moments_central, moments_hu, ... moments_normalized) >>> image = cp.zeros((20, 20), dtype=np.float64) >>> image[13:17, 13:17] = 0.5 >>> image[10:12, 10:12] = 1 >>> mu = moments_central(image) >>> nu = moments_normalized(mu) >>> moments_hu(nu) array([7.45370370e01, 3.51165981e01, 1.04049179e01, 4.06442107e02, 2.64312299e03, 2.40854582e02, 6.50521303e19])
 cucim.skimage.measure.moments_normalized(mu, order=3, spacing=None)#
Calculate all normalized central image moments up to a certain order.
Note that normalized central moments are translation and scale invariant but not rotation invariant.
 Parameters:
 mu(M,[ …,] M) array
Central image moments, where M must be greater than or equal to
order
. orderint, optional
Maximum order of moments. Default is 3.
 Returns:
 nu(
order + 1
,[ …,]order + 1
) array Normalized central image moments.
 nu(
Notes
Differs from the scikitimage implementation in that any moments greater than the requested order will be set to
nan
.References
[1]Wilhelm Burger, Mark Burge. Principles of Digital Image Processing: Core Algorithms. SpringerVerlag, London, 2009.
[2]B. Jähne. Digital Image Processing. SpringerVerlag, BerlinHeidelberg, 6. edition, 2005.
[3]T. H. Reiss. Recognizing Planar Objects Using Invariant Image Features, from Lecture notes in computer science, p. 676. Springer, Berlin, 1993.
Examples
>>> import cupy as cp >>> from cucim.skimage.measure import (moments, moments_central, ... moments_normalized) >>> image = cp.zeros((20, 20), dtype=cp.float64) >>> image[13:17, 13:17] = 1 >>> m = moments(image) >>> centroid = (m[0, 1] / m[0, 0], m[1, 0] / m[0, 0]) >>> mu = moments_central(image, centroid) >>> moments_normalized(mu) array([[ nan, nan, 0.078125 , 0. ], [ nan, 0. , 0. , 0. ], [0.078125 , 0. , 0.00610352, 0. ], [0. , 0. , 0. , 0. ]])
 cucim.skimage.measure.pearson_corr_coeff(image0, image1, mask=None)#
Calculate Pearson’s Correlation Coefficient between pixel intensities in channels.
 Parameters:
 image0(M, N) ndarray
Image of channel A.
 image1(M, N) ndarray
Image of channel 2 to be correlated with channel B. Must have same dimensions as image0.
 mask(M, N) ndarray of dtype bool, optional
Only image0 and image1 pixels within this region of interest mask are included in the calculation. Must have same dimensions as image0.
 Returns:
 pccfloat
Pearson’s correlation coefficient of the pixel intensities between the two images, within the mask if provided.
 pvaluefloat
Twotailed pvalue.
Notes
Pearson’s Correlation Coefficient (PCC) measures the linear correlation between the pixel intensities of the two images. Its value ranges from 1 for perfect linear anticorrelation to +1 for perfect linear correlation. The calculation of the pvalue assumes that the intensities of pixels in each input image are normally distributed.
Scipy’s implementation of Pearson’s correlation coefficient is used. Please refer to it for further information and caveats [1].
\[r = \frac{\sum (A_i  m_A_i) (B_i  m_B_i)} {\sqrt{\sum (A_i  m_A_i)^2 \sum (B_i  m_B_i)^2}}\] where
\(A_i\) is the value of the \(i^{th}\) pixel in image0 \(B_i\) is the value of the \(i^{th}\) pixel in image1, \(m_A_i\) is the mean of the pixel values in image0 \(m_B_i\) is the mean of the pixel values in image1
A low PCC value does not necessarily mean that there is no correlation between the two channel intensities, just that there is no linear correlation. You may wish to plot the pixel intensities of each of the two channels in a 2D scatterplot and use Spearman’s rank correlation if a nonlinear correlation is visually identified [2]. Also consider if you are interested in correlation or cooccurence, in which case a method involving segmentation masks (e.g. MCC or intersection coefficient) may be more suitable [3] [4].
Providing the mask of only relevant sections of the image (e.g., cells, or particular cellular compartments) and removing noise is important as the PCC is sensitive to these measures [3] [4].
References
[3] (1,2)Dunn, K. W., Kamocka, M. M., & McDonald, J. H. (2011). A practical guide to evaluating colocalization in biological microscopy. American journal of physiology. Cell physiology, 300(4), C723–C742. https://doi.org/10.1152/ajpcell.00462.2010
[4] (1,2)Bolte, S. and Cordelières, F.P. (2006), A guided tour into subcellular colocalization analysis in light microscopy. Journal of Microscopy, 224: 213232. https://doi.org/10.1111/j.13652818.2006.01706.x
 cucim.skimage.measure.perimeter(image, neighborhood=4)#
Calculate total perimeter of all objects in binary image.
 Parameters:
 image(N, M) ndarray
2D binary image.
 neighborhood4 or 8, optional
Neighborhood connectivity for border pixel determination. It is used to compute the contour. A higher neighborhood widens the border on which the perimeter is computed.
 Returns:
 perimeterfloat
Total perimeter of all objects in binary image.
 Other Parameters:
 neighbourhoodDEPRECATED
Deprecated in favor of neighborhood.
Deprecated since version 2022.12.00.
References
[1]K. Benkrid, D. Crookes. Design and FPGA Implementation of a Perimeter Estimator. The Queen’s University of Belfast. http://www.cs.qub.ac.uk/~d.crookes/webpubs/papers/perimeter.doc
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage import util >>> from cucim.skimage.measure import label >>> # coins image (binary) >>> img_coins = cp.array(data.coins() > 110) >>> # total perimeter of all objects in the image >>> perimeter(img_coins, neighbourhood=4) array(7796.86799644) >>> perimeter(img_coins, neighbourhood=8) array(8806.26807333)
 cucim.skimage.measure.profile_line(image, src, dst, linewidth=1, order=None, mode='reflect', cval=0.0, *, reduce_func=<function mean>)#
Return the intensity profile of an image measured along a scan line.
 Parameters:
 imagendarray, shape (M, N[, C])
The image, either grayscale (2D array) or multichannel (3D array, where the final axis contains the channel information).
 srcarray_like, shape (2, )
The coordinates of the start point of the scan line.
 dstarray_like, shape (2, )
The coordinates of the end point of the scan line. The destination point is included in the profile, in contrast to standard numpy indexing.
 linewidthint, optional
Width of the scan, perpendicular to the line
 orderint in {0, 1, 2, 3, 4, 5}, optional
The order of the spline interpolation, default is 0 if image.dtype is bool and 1 otherwise. The order has to be in the range 05. See skimage.transform.warp for detail.
 mode{‘constant’, ‘nearest’, ‘reflect’, ‘mirror’, ‘wrap’}, optional
How to compute any values falling outside of the image.
 cvalfloat, optional
If mode is ‘constant’, what constant value to use outside the image.
 reduce_funccallable, optional
Function used to calculate the aggregation of pixel values perpendicular to the profile_line direction when linewidth > 1. If set to None the unreduced array will be returned.
 Returns:
 return_valuearray
The intensity profile along the scan line. The length of the profile is the ceil of the computed length of the scan line.
Examples
>>> import cupy as cp >>> x = cp.asarray([[1, 1, 1, 2, 2, 2]]) >>> img = cp.vstack([cp.zeros_like(x), x, x, x, cp.zeros_like(x)]) >>> img array([[0, 0, 0, 0, 0, 0], [1, 1, 1, 2, 2, 2], [1, 1, 1, 2, 2, 2], [1, 1, 1, 2, 2, 2], [0, 0, 0, 0, 0, 0]]) >>> profile_line(img, (2, 1), (2, 4)) array([1., 1., 2., 2.]) >>> profile_line(img, (1, 0), (1, 6), cval=4) array([1., 1., 1., 2., 2., 2., 2.])
The destination point is included in the profile, in contrast to standard numpy indexing. For example:
>>> profile_line(img, (1, 0), (1, 6)) # The final point is out of bounds array([1., 1., 1., 2., 2., 2., 2.]) >>> profile_line(img, (1, 0), (1, 5)) # This accesses the full first row array([1., 1., 1., 2., 2., 2.])
For different reduce_func inputs:
>>> profile_line(img, (1, 0), (1, 3), linewidth=3, reduce_func=cp.mean) array([0.66666667, 0.66666667, 0.66666667, 1.33333333]) >>> profile_line(img, (1, 0), (1, 3), linewidth=3, reduce_func=cp.max) array([1, 1, 1, 2]) >>> profile_line(img, (1, 0), (1, 3), linewidth=3, reduce_func=cp.sum) array([2, 2, 2, 4])
The unreduced array will be returned when reduce_func is None or when reduce_func acts on each pixel value individually.
>>> profile_line(img, (1, 2), (4, 2), linewidth=3, order=0, ... reduce_func=None) array([[1, 1, 2], [1, 1, 2], [1, 1, 2], [0, 0, 0]]) >>> profile_line(img, (1, 0), (1, 3), linewidth=3, reduce_func=cp.sqrt) array([[1. , 1. , 0. ], [1. , 1. , 0. ], [1. , 1. , 0. ], [1.41421356, 1.41421356, 0. ]])
 cucim.skimage.measure.regionprops(label_image, intensity_image=None, cache=True, *, extra_properties=None, spacing=None)#
Measure properties of labeled image regions.
 Parameters:
 label_image(M, N[, P]) ndarray
Labeled input image. Labels with value 0 are ignored.
Changed in version 0.14.1: Previously,
label_image
was processed bynumpy.squeeze
and so any number of singleton dimensions was allowed. This resulted in inconsistent handling of images with singleton dimensions. To recover the old behaviour, useregionprops(np.squeeze(label_image), ...)
. intensity_image(M, N[, P][, C]) ndarray, optional
Intensity (i.e., input) image with same size as labeled image, plus optionally an extra dimension for multichannel data. Currently, this extra channel dimension, if present, must be the last axis. Default is None.
Changed in version 0.18.0: The ability to provide an extra dimension for channels was added.
 cachebool, optional
Determine whether to cache calculated properties. The computation is much faster for cached properties, whereas the memory consumption increases.
 extra_propertiesIterable of callables
Add extra property computation functions that are not included with skimage. The name of the property is derived from the function name, the dtype is inferred by calling the function on a small sample. If the name of an extra property clashes with the name of an existing property the extra property will not be visible and a UserWarning is issued. A property computation function must take a region mask as its first argument. If the property requires an intensity image, it must accept the intensity image as the second argument.
 spacing: tuple of float, shape (ndim, )
The pixel spacing along each axis of the image.
 Returns:
 propertieslist of RegionProperties
Each item describes one labeled region, and can be accessed using the attributes listed below.
See also
Notes
The following properties can be accessed as attributes or keys:
 num_pixelsint
Number of foreground pixels.
 areafloat
Area of the region i.e. number of pixels of the region scaled by pixelarea.
 area_bboxfloat
Area of the bounding box i.e. number of pixels of bounding box scaled by pixelarea.
 area_convexfloat
Are of the convex hull image, which is the smallest convex polygon that encloses the region.
 area_filledfloat
Area of the region with all the holes filled in.
 axis_major_lengthfloat
The length of the major axis of the ellipse that has the same normalized second central moments as the region.
 axis_minor_lengthfloat
The length of the minor axis of the ellipse that has the same normalized second central moments as the region.
 bboxtuple
Bounding box
(min_row, min_col, max_row, max_col)
. Pixels belonging to the bounding box are in the halfopen interval[min_row; max_row)
and[min_col; max_col)
. centroidarray
Centroid coordinate tuple
(row, col)
. centroid_localarray
Centroid coordinate tuple
(row, col)
, relative to region bounding box. centroid_weightedarray
Centroid coordinate tuple
(row, col)
weighted with intensity image. centroid_weighted_localarray
Centroid coordinate tuple
(row, col)
, relative to region bounding box, weighted with intensity image. coords_scaled(N, 2) ndarray
Coordinate list
(row, col)``of the region scaled by ``spacing
. coords(N, 2) ndarray
Coordinate list
(row, col)
of the region. eccentricityfloat
Eccentricity of the ellipse that has the same secondmoments as the region. The eccentricity is the ratio of the focal distance (distance between focal points) over the major axis length. The value is in the interval [0, 1). When it is 0, the ellipse becomes a circle.
 equivalent_diameter_areafloat
The diameter of a circle with the same area as the region.
 euler_numberint
Euler characteristic of the set of nonzero pixels. Computed as number of connected components subtracted by number of holes (input.ndim connectivity). In 3D, number of connected components plus number of holes subtracted by number of tunnels.
 extentfloat
Ratio of pixels in the region to pixels in the total bounding box. Computed as
area / (rows * cols)
 feret_diameter_maxfloat
Maximum Feret’s diameter computed as the longest distance between points around a region’s convex hull contour as determined by
find_contours
. [5] image(H, J) ndarray
Sliced binary region image which has the same size as bounding box.
 image_convex(H, J) ndarray
Binary convex hull image which has the same size as bounding box.
 image_filled(H, J) ndarray
Binary region image with filled holes which has the same size as bounding box.
 image_intensityndarray
Image inside region bounding box.
 inertia_tensorndarray
Inertia tensor of the region for the rotation around its mass.
 inertia_tensor_eigvalstuple
The eigenvalues of the inertia tensor in decreasing order.
 intensity_maxfloat
Value with the greatest intensity in the region.
 intensity_meanfloat
Value with the mean intensity in the region.
 intensity_minfloat
Value with the least intensity in the region.
 labelint
The label in the labeled input image.
 moments(3, 3) ndarray
Spatial moments up to 3rd order:
m_ij = sum{ array(row, col) * row^i * col^j }
where the sum is over the row, col coordinates of the region.
 moments_central(3, 3) ndarray
Central moments (translation invariant) up to 3rd order:
mu_ij = sum{ array(row, col) * (row  row_c)^i * (col  col_c)^j }
where the sum is over the row, col coordinates of the region, and row_c and col_c are the coordinates of the region’s centroid.
 moments_hutuple
Hu moments (translation, scale and rotation invariant).
 moments_normalized(3, 3) ndarray
Normalized moments (translation and scale invariant) up to 3rd order:
nu_ij = mu_ij / m_00^[(i+j)/2 + 1]
where m_00 is the zeroth spatial moment.
 moments_weighted(3, 3) ndarray
Spatial moments of intensity image up to 3rd order:
wm_ij = sum{ array(row, col) * row^i * col^j }
where the sum is over the row, col coordinates of the region.
 moments_weighted_central(3, 3) ndarray
Central moments (translation invariant) of intensity image up to 3rd order:
wmu_ij = sum{ array(row, col) * (row  row_c)^i * (col  col_c)^j }
where the sum is over the row, col coordinates of the region, and row_c and col_c are the coordinates of the region’s weighted centroid.
 moments_weighted_hutuple
Hu moments (translation, scale and rotation invariant) of intensity image.
 moments_weighted_normalized(3, 3) ndarray
Normalized moments (translation and scale invariant) of intensity image up to 3rd order:
wnu_ij = wmu_ij / wm_00^[(i+j)/2 + 1]
where
wm_00
is the zeroth spatial moment (intensityweighted area). orientationfloat
Angle between the 0th axis (rows) and the major axis of the ellipse that has the same second moments as the region, ranging from pi/2 to pi/2 counterclockwise.
 perimeterfloat
Perimeter of object which approximates the contour as a line through the centers of border pixels using a 4connectivity.
 perimeter_croftonfloat
Perimeter of object approximated by the Crofton formula in 4 directions.
 slicetuple of slices
A slice to extract the object from the source image.
 solidityfloat
Ratio of pixels in the region to pixels of the convex hull image.
Each region also supports iteration, so that you can do:
for prop in region: print(prop, region[prop])
References
[1]Wilhelm Burger, Mark Burge. Principles of Digital Image Processing: Core Algorithms. SpringerVerlag, London, 2009.
[2]B. Jähne. Digital Image Processing. SpringerVerlag, BerlinHeidelberg, 6. edition, 2005.
[3]T. H. Reiss. Recognizing Planar Objects Using Invariant Image Features, from Lecture notes in computer science, p. 676. Springer, Berlin, 1993.
[5]W. Pabst, E. Gregorová. Characterization of particles and particle systems, pp. 2728. ICT Prague, 2007. https://old.vscht.cz/sil/keramika/Characterization_of_particles/CPPS%20_English%20version_.pdf
Examples
>>> from skimage import data, util >>> from cucim.skimage.measure import label, regionprops >>> img = cp.asarray(util.img_as_ubyte(data.coins()) > 110) >>> label_img = label(img, connectivity=img.ndim) >>> props = regionprops(label_img) >>> # centroid of first labeled object >>> props[0].centroid (22.72987986048314, 81.91228523446583) >>> # centroid of first labeled object >>> props[0]['centroid'] (22.72987986048314, 81.91228523446583)
Add custom measurements by passing functions as
extra_properties
>>> from skimage import data, util >>> from cucim.skimage.measure import label, regionprops >>> import numpy as np >>> img = cp.asarray(util.img_as_ubyte(data.coins()) > 110) >>> label_img = label(img, connectivity=img.ndim) >>> def pixelcount(regionmask): ... return np.sum(regionmask) >>> props = regionprops(label_img, extra_properties=(pixelcount,)) >>> props[0].pixelcount array(7741) >>> props[1]['pixelcount'] array(42)
 cucim.skimage.measure.regionprops_table(label_image, intensity_image=None, properties=('label', 'bbox'), *, cache=True, separator='', extra_properties=None, spacing=None)#
Compute image properties and return them as a pandascompatible table.
The table is a dictionary mapping column names to value arrays. See Notes section below for details.
New in version 0.16.
 Parameters:
 label_image(N, M[, P]) ndarray
Labeled input image. Labels with value 0 are ignored.
 intensity_image(M, N[, P][, C]) ndarray, optional
Intensity (i.e., input) image with same size as labeled image, plus optionally an extra dimension for multichannel data. Currently, this extra channel dimension, if present, must be the last axis. Default is None.
Changed in version 0.18.0: The ability to provide an extra dimension for channels was added.
 propertiestuple or list of str, optional
Properties that will be included in the resulting dictionary For a list of available properties, please see
regionprops()
. Users should remember to add “label” to keep track of region identities. cachebool, optional
Determine whether to cache calculated properties. The computation is much faster for cached properties, whereas the memory consumption increases.
 separatorstr, optional
For nonscalar properties not listed in OBJECT_COLUMNS, each element will appear in its own column, with the index of that element separated from the property name by this separator. For example, the inertia tensor of a 2D region will appear in four columns:
inertia_tensor00
,inertia_tensor01
,inertia_tensor10
, andinertia_tensor11
(where the separator is
).Object columns are those that cannot be split in this way because the number of columns would change depending on the object. For example,
image
andcoords
. extra_propertiesIterable of callables
Add extra property computation functions that are not included with skimage. The name of the property is derived from the function name, the dtype is inferred by calling the function on a small sample. If the name of an extra property clashes with the name of an existing property the extra property will not be visible and a UserWarning is issued. A property computation function must take a region mask as its first argument. If the property requires an intensity image, it must accept the intensity image as the second argument.
 spacing: tuple of float, shape (ndim, )
The pixel spacing along each axis of the image.
 Returns:
 out_dictdict
Dictionary mapping property names to an array of values of that property, one value per region. This dictionary can be used as input to pandas
DataFrame
to map property names to columns in the frame and regions to rows. If the image has no regions, the arrays will have length 0, but the correct type.
Notes
Each column contains either a scalar property, an object property, or an element in a multidimensional array.
Properties with scalar values for each region, such as “eccentricity”, will appear as a float or int array with that property name as key.
Multidimensional properties of fixed size for a given image dimension, such as “centroid” (every centroid will have three elements in a 3D image, no matter the region size), will be split into that many columns, with the name {property_name}{separator}{element_num} (for 1D properties), {property_name}{separator}{elem_num0}{separator}{elem_num1} (for 2D properties), and so on.
For multidimensional properties that don’t have a fixed size, such as “image” (the image of a region varies in size depending on the region size), an object array will be used, with the corresponding property name as the key.
Examples
>>> from skimage import data, util, measure >>> image = data.coins() >>> label_image = measure.label(image > 110, connectivity=image.ndim) >>> props = measure.regionprops_table(label_image, image, ... properties=['label', 'inertia_tensor', ... 'inertia_tensor_eigvals']) >>> props {'label': array([ 1, 2, ...]), ... 'inertia_tensor00': array([ 4.012...e+03, 8.51..., ...]), ... ..., 'inertia_tensor_eigvals1': array([ 2.67...e+02, 2.83..., ...])}
The resulting dictionary can be directly passed to pandas, if installed, to obtain a clean DataFrame:
>>> import pandas as pd >>> data = pd.DataFrame(props) >>> data.head() label inertia_tensor00 ... inertia_tensor_eigvals1 0 1 4012.909888 ... 267.065503 1 2 8.514739 ... 2.834806 2 3 0.666667 ... 0.000000 3 4 0.000000 ... 0.000000 4 5 0.222222 ... 0.111111
[5 rows x 7 columns]
If we want to measure a feature that does not come as a builtin property, we can define custom functions and pass them as
extra_properties
. For example, we can create a custom function that measures the intensity quartiles in a region:>>> from skimage import data, util, measure >>> import numpy as np >>> def quartiles(regionmask, intensity): ... return np.percentile(intensity[regionmask], q=(25, 50, 75)) >>> >>> image = data.coins() >>> label_image = measure.label(image > 110, connectivity=image.ndim) >>> props = measure.regionprops_table(label_image, intensity_image=image, ... properties=('label',), ... extra_properties=(quartiles,)) >>> import pandas as pd >>> pd.DataFrame(props).head() label quartiles0 quartiles1 quartiles2 0 1 117.00 123.0 130.0 1 2 111.25 112.0 114.0 2 3 111.00 111.0 111.0 3 4 111.00 111.5 112.5 4 5 112.50 113.0 114.0
 cucim.skimage.measure.shannon_entropy(image, base=2)#
Calculate the Shannon entropy of an image.
The Shannon entropy is defined as S = sum(pk * log(pk)), where pk are frequency/probability of pixels of value k.
 Parameters:
 image(N, M) ndarray
Grayscale input image.
 basefloat, optional
The logarithmic base to use.
 Returns:
 entropy0dimensional float cupy.ndarray
Notes
The returned value is measured in bits or shannon (Sh) for base=2, natural unit (nat) for base=np.e and hartley (Hart) for base=10.
References
Examples
>>> import cupy as cp >>> from skimage import data >>> from cucim.skimage.measure import shannon_entropy >>> shannon_entropy(cp.array(data.camera())) array(7.23169501)
 cucim.skimage.measure.subdivide_polygon(coords, degree=2, preserve_ends=False)#
Subdivision of polygonal curves using BSplines.
Note that the resulting curve is always within the convex hull of the original polygon. Circular polygons stay closed after subdivision.
 Parameters:
 coords(N, 2) array
Coordinate array.
 degree{1, 2, 3, 4, 5, 6, 7}, optional
Degree of BSpline. Default is 2.
 preserve_endsbool, optional
Preserve first and last coordinate of noncircular polygon. Default is False.
 Returns:
 coords(M, 2) array
Subdivided coordinate array.
References
metrics#
 cucim.skimage.metrics.adapted_rand_error(image_true=None, image_test=None, *, table=None, ignore_labels=(0,), alpha=0.5)#
Compute Adapted Rand error as defined by the SNEMI3D contest. [1]
 Parameters:
 image_truecp.ndarray of int
Groundtruth label image, same shape as im_test.
 image_testcp.ndarray of int
Test image.
 tablecupyx.scipy.sparse array in csr format, optional
A contingency table built with skimage.evaluate.contingency_table. If None, it will be computed on the fly.
 ignore_labelssequence of int, optional
Labels to ignore. Any part of the true image labeled with any of these values will not be counted in the score.
 alphafloat, optional
Relative weight given to precision and recall in the adapted Rand error calculation.
 Returns:
 arefloat
The adapted Rand error.
 precfloat
The adapted Rand precision: this is the number of pairs of pixels that have the same label in the test label image and in the true image, divided by the number in the test image.
 recfloat
The adapted Rand recall: this is the number of pairs of pixels that have the same label in the test label image and in the true image, divided by the number in the true image.
Notes
Pixels with label 0 in the true segmentation are ignored in the score.
The adapted Rand error is calculated as follows:
\(1  \frac{\sum_{ij} p_{ij}^{2}}{\alpha \sum_{k} s_{k}^{2} + (1\alpha)\sum_{k} t_{k}^{2}}\), where \(p_{ij}\) is the probability that a pixel has the same label in the test image and in the true image, \(t_{k}\) is the probability that a pixel has label \(k\) in the true image, and \(s_{k}\) is the probability that a pixel has label \(k\) in the test image.
Default behavior is to weight precision and recall equally in the adapted Rand error calculation. When alpha = 0, adapted Rand error = recall. When alpha = 1, adapted Rand error = precision.
References
[1]ArgandaCarreras I, Turaga SC, Berger DR, et al. (2015) Crowdsourcing the creation of image segmentation algorithms for connectomics. Front. Neuroanat. 9:142. DOI:10.3389/fnana.2015.00142
 cucim.skimage.metrics.contingency_table(im_true, im_test, *, ignore_labels=None, normalize=False)#
Return the contingency table for all regions in matched segmentations.
 Parameters:
 im_truendarray of int
Groundtruth label image, same shape as im_test.
 im_testndarray of int
Test image.
 ignore_labelssequence of int, optional
Labels to ignore. Any part of the true image labeled with any of these values will not be counted in the score.
 normalizebool
Determines if the contingency table is normalized by pixel count.
 Returns:
 contscipy.sparse.csr_matrix
A contingency table. cont[i, j] will equal the number of voxels labeled i in im_true and j in im_test.
 cucim.skimage.metrics.mean_squared_error(image0, image1)#
Compute the meansquared error between two images.
 Parameters:
 image0, image1ndarray
Images. Any dimensionality, must have same shape.
 Returns:
 msefloat
The meansquared error (MSE) metric.
Notes
Changed in version 0.16: This function was renamed from
skimage.measure.compare_mse
toskimage.metrics.mean_squared_error
.
 cucim.skimage.metrics.normalized_mutual_information(image0, image1, *, bins=100)#
Compute the normalized mutual information (NMI).
The normalized mutual information of \(A\) and \(B\) is given by:
.. math::
Y(A, B) = frac{H(A) + H(B)}{H(A, B)}
where \(H(X) :=  \sum_{x \in X}{x \log x}\) is the entropy.
It was proposed to be useful in registering images by Colin Studholme and colleagues [1]. It ranges from 1 (perfectly uncorrelated image values) to 2 (perfectly correlated image values, whether positively or negatively).
 Parameters:
 image0, image1ndarray
Images to be compared. The two input images must have the same number of dimensions.
 binsint or sequence of int, optional
The number of bins along each axis of the joint histogram.
 Returns:
 nmifloat
The normalized mutual information between the two arrays, computed at the granularity given by
bins
. Higher NMI implies more similar input images.
 Raises:
 ValueError
If the images don’t have the same number of dimensions.
Notes
If the two input images are not the same shape, the smaller image is padded with zeros.
References
[1]C. Studholme, D.L.G. Hill, & D.J. Hawkes (1999). An overlap invariant entropy measure of 3D medical image alignment. Pattern Recognition 32(1):7186 DOI:10.1016/S00313203(98)000910
 cucim.skimage.metrics.normalized_root_mse(image_true, image_test, *, normalization='euclidean')#
Compute the normalized root meansquared error (NRMSE) between two images.
 Parameters:
 image_truendarray
Groundtruth image, same shape as im_test.
 image_testndarray
Test image.
 normalization{‘euclidean’, ‘minmax’, ‘mean’}, optional
Controls the normalization method to use in the denominator of the NRMSE. There is no standard method of normalization across the literature [1]. The methods available here are as follows:
‘euclidean’ : normalize by the averaged Euclidean norm of
im_true
:NRMSE = RMSE * sqrt(N) /  im_true 
where  .  denotes the Frobenius norm and
N = im_true.size
. This result is equivalent to:NRMSE =  im_true  im_test  /  im_true .
‘minmax’ : normalize by the intensity range of
im_true
.‘mean’ : normalize by the mean of
im_true
 Returns:
 nrmsefloat
The NRMSE metric.
Notes
Changed in version 0.16: This function was renamed from
skimage.measure.compare_nrmse
toskimage.metrics.normalized_root_mse
.References
 cucim.skimage.metrics.peak_signal_noise_ratio(image_true, image_test, *, data_range=None)#
Compute the peak signal to noise ratio (PSNR) for an image.
 Parameters:
 image_truendarray
Groundtruth image, same shape as im_test.
 image_testndarray
Test image.
 data_rangeint, optional
The data range of the input image (distance between minimum and maximum possible values). By default, this is estimated from the image datatype.
 Returns:
 psnrfloat
The PSNR metric.
Notes
Changed in version 0.16: This function was renamed from
skimage.measure.compare_psnr
toskimage.metrics.peak_signal_noise_ratio
.References
 cucim.skimage.metrics.structural_similarity(im1, im2, *, win_size=None, gradient=False, data_range=None, channel_axis=None, gaussian_weights=False, full=False, **kwargs)#
Compute the mean structural similarity index between two images. Please pay attention to the data_range parameter with floatingpoint images.
 Parameters:
 im1, im2ndarray
Images. Any dimensionality with same shape.
 win_sizeint or None, optional
The sidelength of the sliding window used in comparison. Must be an odd value. If gaussian_weights is True, this is ignored and the window size will depend on sigma.
 gradientbool, optional
If True, also return the gradient with respect to im2.
 data_rangefloat, optional
The data range of the input image (distance between minimum and maximum possible values). By default, this is estimated from the image data type. This estimate may be wrong for floatingpoint image data. Therefore it is recommended to always pass this value explicitly (see note below).
 channel_axisint or None, optional
If None, the image is assumed to be a grayscale (single channel) image. Otherwise, this parameter indicates which axis of the array corresponds to channels.
 gaussian_weightsbool, optional
If True, each patch has its mean and variance spatially weighted by a normalized Gaussian kernel of width sigma=1.5.
 fullbool, optional
If True, also return the full structural similarity image.
 Returns:
 mssimfloat
The mean structural similarity index over the image.
 gradndarray
The gradient of the structural similarity between im1 and im2 [2]. This is only returned if gradient is set to True.
 Sndarray
The full SSIM image. This is only returned if full is set to True.
 Other Parameters:
 use_sample_covariancebool
If True, normalize covariances by N1 rather than, N where N is the number of pixels within the sliding window.
 K1float
Algorithm parameter, K1 (small constant, see [1]).
 K2float
Algorithm parameter, K2 (small constant, see [1]).
 sigmafloat
Standard deviation for the Gaussian when gaussian_weights is True.
Notes
If data_range is not specified, the range is automatically guessed based on the image data type. However for floatingpoint image data, this estimate yields a result double the value of the desired range, as the dtype_range in skimage.util.dtype.py has defined intervals from 1 to +1. This yields an estimate of 2, instead of 1, which is most often required when working with image data (as negative light intentsities are nonsensical). In case of working with YCbCrlike color data, note that these ranges are different per channel (Cb and Cr have double the range of Y), so one cannot calculate a channelaveraged SSIM with a single call to this function, as identical ranges are assumed for each channel.
To match the implementation of Wang et al. [1], set gaussian_weights to True, sigma to 1.5, use_sample_covariance to False, and specify the data_range argument.
Changed in version 0.16: This function was renamed from
skimage.measure.compare_ssim
toskimage.metrics.structural_similarity
.References
[1] (1,2,3)Wang, Z., Bovik, A. C., Sheikh, H. R., & Simoncelli, E. P. (2004). Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13, 600612. https://ece.uwaterloo.ca/~z70wang/publications/ssim.pdf, DOI:10.1109/TIP.2003.819861
[2]Avanaki, A. N. (2009). Exact global histogram specification optimized for structural similarity. Optical Review, 16, 613621. arXiv:0901.0065 DOI:10.1007/s100430090119z
 cucim.skimage.metrics.variation_of_information(image0=None, image1=None, *, table=None, ignore_labels=())#
Return symmetric conditional entropies associated with the VI. [1]
The variation of information is defined as VI(X,Y) = H(XY) + H(YX). If X is the groundtruth segmentation, then H(XY) can be interpreted as the amount of undersegmentation and H(XY) as the amount of oversegmentation. In other words, a perfect oversegmentation will have H(XY)=0 and a perfect undersegmentation will have H(YX)=0.
 Parameters:
 image0, image1cp.ndarray of int
Label images / segmentations, must have same shape.
 tablecupyx.scipy.sparse array in csr format, optional
A contingency table built with skimage.evaluate.contingency_table. If None, it will be computed with skimage.evaluate.contingency_table. If given, the entropies will be computed from this table and any images will be ignored.
 ignore_labelssequence of int, optional
Labels to ignore. Any part of the true image labeled with any of these values will not be counted in the score.
 Returns:
 vicp.ndarray of float, shape (2,)
The conditional entropies of image1image0 and image0image1.
References
[1]Marina Meilă (2007), Comparing clusterings—an information based distance, Journal of Multivariate Analysis, Volume 98, Issue 5, Pages 873895, ISSN 0047259X, DOI:10.1016/j.jmva.2006.11.013.
morphology#
 cucim.skimage.morphology.ball(radius, dtype=<class 'numpy.uint8'>, *, strict_radius=True, decomposition=None)#
Generates a ballshaped footprint.
This is the 3D equivalent of a disk. A pixel is within the neighborhood if the Euclidean distance between it and the origin is no greater than radius.
 Parameters:
 radiusint
The radius of the ballshaped footprint.
 Returns:
 footprintcupy.ndarray
The footprint where elements of the neighborhood are 1 and 0 otherwise.
 Other Parameters:
 dtypedatatype, optional
The data type of the footprint.
 strict_radiusbool, optional
If False, extend the radius by 0.5. This allows the circle to expand further within a cube that remains of size
2 * radius + 1
along each axis. This parameter is ignored if decomposition is not None. decomposition{None, ‘sequence’}, optional
If None, a single array is returned. For ‘sequence’, a tuple of smaller footprints is returned. Applying this series of smaller footprints will given a result equivalent to a single, larger footprint, but with better computational performance. For ball footprints, the sequence decomposition is not exactly equivalent to decomposition=None. See Notes for more details.
Notes
The disk produced by the decomposition=’sequence’ mode is not identical to that with decomposition=None. Here we extend the approach taken in [1] for disks to the 3D case, using 3dimensional extensions of the “square”, “diamond” and “tshaped” elements from that publication. All of these elementary elements have size
(3,) * ndim
. We numerically computed the number of repetitions of each element that gives the closest match to the ball computed with kwargsstrict_radius=False, decomposition=None
.Empirically, the equivalent composite footprint to the sequence decomposition approaches a rhombicuboctahedron (26faces [2]).
References
[1]Park, H and Chin R.T. Decomposition of structuring elements for optimal implementation of morphological operations. In Proceedings: 1997 IEEE Workshop on Nonlinear Signal and Image Processing, London, UK. https://www.iwaenc.org/proceedings/1997/nsip97/pdf/scan/ns970226.pdf
 cucim.skimage.morphology.binary_closing(image, footprint=None, out=None)#
Return fast binary morphological closing of an image.
This function returns the same result as grayscale closing but performs faster for binary images.
The morphological closing on an image is defined as a dilation followed by an erosion. Closing can remove small dark spots (i.e. “pepper”) and connect small bright cracks. This tends to “close” up (dark) gaps between (bright) features.
 Parameters:
 imagendarray
Binary input image.
 footprintndarray or tuple, optional
The neighborhood expressed as a 2D array of 1’s and 0’s. If None, use a crossshaped footprint (connectivity=1). The footprint can also be provided as a sequence of smaller footprints as described in the notes below.
 outndarray of bool, optional
The array to store the result of the morphology. If None, is passed, a new array will be allocated.
 Returns:
 closingndarray of bool
The result of the morphological closing.
 Other Parameters:
 selemDEPRECATED
Deprecated in favor of footprint.
Deprecated since version 22.02.00.
Notes
The footprint can also be a provided as a sequence of 2tuples where the first element of each 2tuple is a footprint ndarray and the second element is an integer describing the number of times it should be iterated. For example
footprint=[(cp.ones((9, 1)), 1), (cp.ones((1, 9)), 1)]
would apply a 9x1 footprint followed by a 1x9 footprint resulting in a net effect that is the same asfootprint=cp.ones((9, 9))
, but with lower computational cost. Most of the builtin footprints such asskimage.morphology.disk
provide an option to automically generate a footprint sequence of this type.
 cucim.skimage.morphology.binary_dilation(image, footprint=None, out=None)#
Return fast binary morphological dilation of an image.
This function returns the same result as grayscale dilation but performs faster for binary images.
Morphological dilation sets a pixel at
(i,j)
to the maximum over all pixels in the neighborhood centered at(i,j)
. Dilation enlarges bright regions and shrinks dark regions. Parameters:
 imagendarray
Binary input image.
 footprintndarray or tuple, optional
The neighborhood expressed as a 2D array of 1’s and 0’s. If None, use a crossshaped footprint (connectivity=1). The footprint can also be provided as a sequence of smaller footprints as described in the notes below.
 outndarray of bool, optional
The array to store the result of the morphology. If None is passed, a new array will be allocated.
 Returns:
 dilatedndarray of bool or uint
The result of the morphological dilation with values in
[False, True]
.
 Other Parameters:
 selemDEPRECATED
Deprecated in favor of footprint.
Deprecated since version 22.02.00.
Notes
The footprint can also be a provided as a sequence of 2tuples where the first element of each 2tuple is a footprint ndarray and the second element is an integer describing the number of times it should be iterated. For example
footprint=[(cp.ones((9, 1)), 1), (cp.ones((1, 9)), 1)]
would apply a 9x1 footprint followed by a 1x9 footprint resulting in a net effect that is the same asfootprint=cp.ones((9, 9))
, but with lower computational cost. Most of the builtin footprints such asskimage.morphology.disk
provide an option to automically generate a footprint sequence of this type.
 cucim.skimage.morphology.binary_erosion(image, footprint=None, out=None)#
Return fast binary morphological erosion of an image.
This function returns the same result as grayscale erosion but performs faster for binary images.
Morphological erosion sets a pixel at
(i,j)
to the minimum over all pixels in the neighborhood centered at(i,j)
. Erosion shrinks bright regions and enlarges dark regions. Parameters:
 imagendarray
Binary input image.
 footprintndarray or tuple, optional
The neighborhood expressed as a 2D array of 1’s and 0’s. If None, use a crossshaped footprint (connectivity=1). The footprint can also be provided as a sequence of smaller footprints as described in the notes below.
 outndarray of bool, optional
The array to store the result of the morphology. If None is passed, a new array will be allocated.
 Returns:
 erodedndarray of bool or uint
The result of the morphological erosion taking values in
[False, True]
.
 Other Parameters:
 selemDEPRECATED
Deprecated in favor of footprint.
Deprecated since version 22.02.00.
Notes
The footprint can also be a provided as a sequence of 2tuples where the first element of each 2tuple is a footprint ndarray and the second element is an integer describing the number of times it should be iterated. For example
footprint=[(cp.ones((9, 1)), 1), (cp.ones((1, 9)), 1)]
would apply a 9x1 footprint followed by a 1x9 footprint resulting in a net effect that is the same asfootprint=cp.ones((9, 9))
, but with lower computational cost. Most of the builtin footprints such asskimage.morphology.disk
provide an option to automically generate a footprint sequence of this type.
 cucim.skimage.morphology.binary_opening(image, footprint=None, out=None)#
Return fast binary morphological opening of an image.
This function returns the same result as grayscale opening but performs faster for binary images.
The morphological opening on an image is defined as an erosion followed by a dilation. Opening can remove small bright spots (i.e. “salt”) and connect small dark cracks. This tends to “open” up (dark) gaps between (bright) features.
 Parameters:
 imagendarray
Binary input image.
 footprintndarray or tuple, optional
The neighborhood expressed as a 2D array of 1’s and 0’s. If None, use a crossshaped footprint (connectivity=1). The footprint can also be provided as a sequence of smaller footprints as described in the notes below.
 outndarray of bool, optional
The array to store the result of the morphology. If None is passed, a new array will be allocated.
 Returns:
 openingndarray of bool
The result of the morphological opening.
 Other Parameters:
 selemDEPRECATED
Deprecated in favor of footprint.
Deprecated since version 22.02.00.
Notes
The footprint can also be a provided as a sequence of 2tuples where the first element of each 2tuple is a footprint ndarray and the second element is an integer describing the number of times it should be iterated. For example
footprint=[(cp.ones((9, 1)), 1), (cp.ones((1, 9)), 1)]
would apply a 9x1 footprint followed by a 1x9 footprint resulting in a net effect that is the same asfootprint=cp.ones((9, 9))
, but with lower computational cost. Most of the builtin footprints such asskimage.morphology.disk
provide an option to automically generate a footprint sequence of this type.
 cucim.skimage.morphology.black_tophat(image, footprint=None, out=None)#
Return black top hat of an image.
The black top hat of an image is defined as its morphological closing minus the original image. This operation returns the dark spots of the image that are smaller than the footprint. Note that dark spots in the original image are bright spots after the black top hat.
 Parameters:
 imagecupy.ndarray
Image array.
 footprintcupy.ndarray, optional
The neighborhood expressed as a 2D array of 1’s and 0’s. If None, use a crossshaped footprint (connectivity=1). The footprint can also be provided as a sequence of smaller footprints as described in the notes below.
 outcupy.ndarray, optional
The array to store the result of the morphology. If None is passed, a new array will be allocated.
 Returns:
 outcupy.ndarray, same shape and type as image
The result of the morphological black top hat.
 Other Parameters:
 selemDEPRECATED
Deprecated in favor of footprint.
Deprecated since version 22.02.00.
See also
Notes
The footprint can also be a provided as a sequence of 2tuples where the first element of each 2tuple is a footprint ndarray and the second element is an integer describing the number of times it should be iterated. For example
footprint=[(cp.ones((9, 1)), 1), (cp.ones((1, 9)), 1)]
would apply a 9x1 footprint followed by a 1x9 footprint resulting in a net effect that is the same asfootprint=cp.ones((9, 9))
, but with lower computational cost. Most of the builtin footprints such asskimage.morphology.disk
provide an option to automically generate a footprint sequence of this type.References
Examples
>>> # Change dark peak to bright peak and subtract background >>> import cupy as cp >>> from cucim.skimage.morphology import square >>> dark_on_grey = cp.asarray([[7, 6, 6, 6, 7], ... [6, 5, 4, 5, 6], ... [6, 4, 0, 4, 6], ... [6, 5, 4, 5, 6], ... [7, 6, 6, 6, 7]], dtype=cp.uint8) >>> black_tophat(dark_on_grey, square(3)) array([[0, 0, 0, 0, 0], [0, 0, 1, 0, 0], [0, 1, 5, 1, 0], [0, 0, 1, 0, 0], [0, 0, 0, 0, 0]], dtype=uint8)
 cucim.skimage.morphology.closing(image, footprint=None, out=None)#
Return grayscale morphological closing of an image.
The morphological closing of an image is defined as a dilation followed by an erosion. Closing can remove small dark spots (i.e. “pepper”) and connect small bright cracks. This tends to “close” up (dark) gaps between (bright) features.
 Parameters:
 imagecupy.ndarray
Image array.
 footprintcupy.ndarray, optional
The neighborhood expressed as a 2D array of 1’s and 0’s. If None, use a crossshaped footprint (connectivity=1). The footprint can also be provided as a sequence of smaller footprints as described in the notes below.
 outcupy.ndarray, optional
The array to store the result of the morphology. If None, a new array will be allocated.
 Returns:
 closingcupy.ndarray, same shape and type as image
The result of the morphological closing.
 Other Parameters:
 selemDEPRECATED
Deprecated in favor of footprint.
Deprecated since version 22.02.00.
Notes
The footprint can also be a provided as a sequence of 2tuples where the first element of each 2tuple is a footprint ndarray and the second element is an integer describing the number of times it should be iterated. For example
footprint=[(cp.ones((9, 1)), 1), (cp.ones((1, 9)), 1)]
would apply a 9x1 footprint followed by a 1x9 footprint resulting in a net effect that is the same asfootprint=cp.ones((9, 9))
, but with lower computational cost. Most of the builtin footprints such asskimage.morphology.disk
provide an option to automically generate a footprint sequence of this type.Examples
>>> # Close a gap between two bright lines >>> import cupy as cp >>> from cucim.skimage.morphology import square >>> broken_line = cp.asarray([[0, 0, 0, 0, 0], ... [0, 0, 0, 0, 0], ... [1, 1, 0, 1, 1], ... [0, 0, 0, 0, 0], ... [0, 0, 0, 0, 0]], dtype=cp.uint8) >>> closing(broken_line, square(3)) array([[0, 0, 0, 0, 0], [0, 0, 0, 0, 0], [1, 1, 1, 1, 1], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0]], dtype=uint8)
 cucim.skimage.morphology.cube(width, dtype=None, *, decomposition=None)#
Generates a cubeshaped footprint.
This is the 3D equivalent of a square. Every pixel along the perimeter has a chessboard distance no greater than radius (radius=floor(width/2)) pixels.
 Parameters:
 widthint
The width, height and depth of the cube.
 Returns:
 footprintcupy.ndarray
The footprint where elements of the neighborhood are 1 and 0 otherwise. When decomposition is None, this is just a numpy.ndarray. Otherwise, this will be a tuple whose length is equal to the number of unique structuring elements to apply (see Notes for more detail)
 Other Parameters:
 dtypedatatype or None, optional
The data type of the footprint. When None, a tuple will be returned in place of the actual footprint array. This can be be passed to grayscale and binary morphology functions in place of an explicit array to avoid array allocation overhead.
 decomposition{None, ‘separable’, ‘sequence’}, optional
If None, a single array is returned. For ‘sequence’, a tuple of smaller footprints is returned. Applying this series of smaller footprints will given an identical result to a single, larger footprint, but often with better computational performance. See Notes for more details.
Notes
When decomposition is not None, each element of the footprint tuple is a 2tuple of the form
(ndarray, num_iter)
that specifies a footprint array and the number of iterations it is to be applied.For binary morphology, using
decomposition='sequence'
was observed to give better performance, with the magnitude of the performance increase rapidly increasing with footprint size. For grayscale morphology with square footprints, it is recommended to usedecomposition=None
since the internal SciPy functions that are called already have a fast implementation based on separable 1D sliding windows.The ‘sequence’ decomposition mode only supports odd valued width. If width is even, the sequence used will be identical to the ‘separable’ mode.
 cucim.skimage.morphology.diamond(radius, dtype=<class 'numpy.uint8'>, *, decomposition=None)#
Generates a flat, diamondshaped footprint.
A pixel is part of the neighborhood (i.e. labeled 1) if the city block/Manhattan distance between it and the center of the neighborhood is no greater than radius.
 Parameters:
 radiusint
The radius of the diamondshaped footprint.
 Returns:
 footprintcupy.ndarray
The footprint where elements of the neighborhood are 1 and 0 otherwise. When decomposition is None, this is just a numpy.ndarray. Otherwise, this will be a tuple whose length is equal to the number of unique structuring elements to apply (see Notes for more detail)
 Other Parameters:
 dtypedatatype, optional
The data type of the footprint.
 decomposition{None, ‘sequence’}, optional
If None, a single array is returned. For ‘sequence’, a tuple of smaller footprints is returned. Applying this series of smaller footprints will given an identical result to a single, larger footprint, but with better computational performance. See Notes for more details.
Notes
When decomposition is not None, each element of the footprint tuple is a 2tuple of the form
(ndarray, num_iter)
that specifies a footprint array and the number of iterations it is to be applied.For either binary or grayscale morphology, using
decomposition='sequence'
was observed to have a performance benefit, with the magnitude of the benefit increasing with increasing footprint size.
 cucim.skimage.morphology.dilation(image, footprint=None, out=None, shift_x=False, shift_y=False)#
Return grayscale morphological dilation of an image.
Morphological dilation sets the value of a pixel to the maximum over all pixel values within a local neighborhood centered about it. The values where the footprint is 1 define this neighborhood. Dilation enlarges bright regions and shrinks dark regions.
 Parameters:
 imagecupy.ndarray
Image array.
 footprintcupy.ndarray, optional
The neighborhood expressed as a 2D array of 1’s and 0’s. If None, use a crossshaped footprint (connectivity=1). The footprint can also be provided as a sequence of smaller footprints as described in the notes below.
 outcupy.ndarray, optional
The array to store the result of the morphology. If None is passed, a new array will be allocated.
 shift_x, shift_ybool, optional
Shift footprint about center point. This only affects 2D eccentric footprints (i.e., footprints with evennumbered sides).
 Returns:
 dilatedcupy.ndarray, same shape and type as image
The result of the morphological dilation.
 Other Parameters:
 selemDEPRECATED
Deprecated in favor of footprint.
Deprecated since version 22.02.00.
Notes
For uint8 (and uint16 up to a certain bitdepth) data, the lower algorithm complexity makes the skimage.filters.rank.maximum function more efficient for larger images and footprints.
The footprint can also be a provided as a sequence of 2tuples where the first element of each 2tuple is a footprint ndarray and the second element is an integer describing the number of times it should be iterated. For example
footprint=[(cp.ones((9, 1)), 1), (cp.ones((1, 9)), 1)]
would apply a 9x1 footprint followed by a 1x9 footprint resulting in a net effect that is the same asfootprint=cp.ones((9, 9))
, but with lower computational cost. Most of the builtin footprints such asskimage.morphology.disk
provide an option to automically generate a footprint sequence of this type.Examples
>>> # Dilation enlarges bright regions >>> import cupy as cp >>> from cucim.skimage.morphology import square >>> bright_pixel = cp.asarray([[0, 0, 0, 0, 0], ... [0, 0, 0, 0, 0], ... [0, 0, 1, 0, 0], ... [0, 0, 0, 0, 0], ... [0, 0, 0, 0, 0]], dtype=cp.uint8) >>> dilation(bright_pixel, square(3)) array([[0, 0, 0, 0, 0], [0, 1, 1, 1, 0], [0, 1, 1, 1, 0], [0, 1, 1, 1, 0], [0, 0, 0, 0, 0]], dtype=uint8)
 cucim.skimage.morphology.disk(radius, dtype=<class 'numpy.uint8'>, *, strict_radius=True, decomposition=None)#
Generates a flat, diskshaped footprint.
A pixel is within the neighborhood if the Euclidean distance between it and the origin is no greater than radius (This is only approximately True, when decomposition == ‘sequence’).
 Parameters:
 radiusint
The radius of the diskshaped footprint.
 Returns:
 footprintcupy.ndarray
The footprint where elements of the neighborhood are 1 and 0 otherwise.
 Other Parameters:
 dtypedatatype, optional
The data type of the footprint.
 strict_radiusbool, optional
If False, extend the radius by 0.5. This allows the circle to expand further within a cube that remains of size
2 * radius + 1
along each axis. This parameter is ignored if decomposition is not None. decomposition{None, ‘sequence’, ‘crosses’}, optional
If None, a single array is returned. For ‘sequence’, a tuple of smaller footprints is returned. Applying this series of smaller footprints will given a result equivalent to a single, larger footprint, but with better computational performance. For disk footprints, the ‘sequence’ or ‘crosses’ decompositions are not always exactly equivalent to
decomposition=None
. See Notes for more details.
Notes
When decomposition is not None, each element of the footprint tuple is a 2tuple of the form
(ndarray, num_iter)
that specifies a footprint array and the number of iterations it is to be applied.The disk produced by the
decomposition='sequence'
mode may not be identical to that withdecomposition=None
. A disk footprint can be approximated by applying a series of smaller footprints of extent 3 along each axis. Specific solutions for this are given in [1] for the case of 2D disks with radius 2 through 10. Here, we numerically computed the number of repetitions of each element that gives the closest match to the disk computed with kwargsstrict_radius=False, decomposition=None
.Empirically, the series decomposition at large radius approaches a hexadecagon (a 16sided polygon [2]). In [3], the authors demonstrate that a hexadecagon is the closest approximation to a disk that can be achieved for decomposition with footprints of shape (3, 3).
The disk produced by the
decomposition='crosses'
is often but not always identical to that withdecomposition=None
. It tends to give a closer approximation thandecomposition='sequence'
, at a performance that is fairly comparable. The individual crossshaped elements are not limited to extent (3, 3) in size. Unlike the ‘seqeuence’ decomposition, the ‘crosses’ decomposition can also accurately approximate the shape of disks withstrict_radius=True
. The method is based on an adaption of algorithm 1 given in [4].References
[1]Park, H and Chin R.T. Decomposition of structuring elements for optimal implementation of morphological operations. In Proceedings: 1997 IEEE Workshop on Nonlinear Signal and Image Processing, London, UK. https://www.iwaenc.org/proceedings/1997/nsip97/pdf/scan/ns970226.pdf
[3]Vanrell, M and Vitrià, J. Optimal 3 × 3 decomposable disks for morphological transformations. Image and Vision Computing, Vol. 15, Issue 11, 1997. DOI:10.1016/S02628856(97)000267
[4]Li, D. and Ritter, G.X. Decomposition of Separable and Symmetric Convex Templates. Proc. SPIE 1350, Image Algebra and Morphological Image Processing, (1 November 1990). DOI:10.1117/12.23608
 cucim.skimage.morphology.erosion(image, footprint=None, out=None, shift_x=False, shift_y=False)#
Return grayscale morphological erosion of an image.
Morphological erosion sets a pixel at (i,j) to the minimum over all pixels in the neighborhood centered at (i,j). Erosion shrinks bright regions and enlarges dark regions.
 Parameters:
 imagecupy.ndarray
Image array.
 footprintcupy.ndarray, optional
The neighborhood expressed as a 2D array of 1’s and 0’s. If None, use a crossshaped footprint (connectivity=1). The footprint can also be provided as a sequence of smaller footprints as described in the notes below.
 outcupy.ndarray, optional
The array to store the result of the morphology. If None is passed, a new array will be allocated.
 shift_x, shift_ybool, optional
shift footprint about center point. This only affects eccentric footprints (i.e. footprint with even numbered sides).
 Returns:
 erodedcupy.ndarray, same shape as image
The result of the morphological erosion.
 Other Parameters:
 selemDEPRECATED
Deprecated in favor of footprint.
Deprecated since version 22.02.00.
Notes
For
uint8
(anduint16
up to a certain bitdepth) data, the lower algorithm complexity makes the skimage.filters.rank.minimum function more efficient for larger images and footprints.The footprint can also be a provided as a sequence of 2tuples where the first element of each 2tuple is a footprint ndarray and the second element is an integer describing the number of times it should be iterated. For example
footprint=[(cp.ones((9, 1)), 1), (cp.ones((1, 9)), 1)]
would apply a 9x1 footprint followed by a 1x9 footprint resulting in a net effect that is the same asfootprint=cp.ones((9, 9))
, but with lower computational cost. Most of the builtin footprints such asskimage.morphology.disk
provide an option to automically generate a footprint sequence of this type.Examples
>>> # Erosion shrinks bright regions >>> import cupy as cp >>> from cucim.skimage.morphology import square >>> bright_square = cp.asarray([[0, 0, 0, 0, 0], ... [0, 1, 1, 1, 0], ... [0, 1, 1, 1, 0], ... [0, 1, 1, 1, 0], ... [0, 0, 0, 0, 0]], dtype=cp.uint8) >>> erosion(bright_square, square(3)) array([[0, 0, 0, 0, 0], [0, 0, 0, 0, 0], [0, 0, 1, 0, 0], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0]], dtype=uint8)
 cucim.skimage.morphology.isotropic_closing(image, radius, out=None, spacing=None)#
Return binary morphological closing of an image.
This function returns the same result as binary
skimage.morphology.binary_closing()
but performs faster for large circular structuring elements. This works by thresholding the exact Euclidean distance map [1], [2]. The implementation is based on: func:cucim.core.operations.morphology.distance_transform_edt. Parameters:
 imagendarray
Binary input image.
 radiusfloat
The radius with which the regions should be closed.
 outndarray of bool, optional
The array to store the result of the morphology. If None, is passed, a new array will be allocated.
 spacingfloat, or sequence of float, optional
Spacing of elements along each dimension. If a sequence, must be of length equal to the input’s dimension (number of axes). If a single number, this value is used for all axes. If not specified, a grid spacing of unity is implied.
 Returns:
 closedndarray of bool
The result of the morphological closing.
Notes
Empirically, on an RTX A6000 GPU, it was observed that
isotropic_closing
is faster thanbinary_closing
withdecomposition=None
at radius 12 in 2D and radius 3 in 3D. It becomes faster thanbinary_erosion
withdecomposition="sequence"
at radius 14 in 2D and radius 5 in 3D. In practice, the exact point at which these isotropic functions become faster than their binary counterparts will also be dependent on image shape and content.References
[1]Cuisenaire, O. and Macq, B., “Fast Euclidean morphological operators using local distance transformation by propagation, and applications,” Image Processing And Its Applications, 1999. Seventh International Conference on (Conf. Publ. No. 465), 1999, pp. 856860 vol.2. DOI:10.1049/cp:19990446
[2]Ingemar Ragnemalm, Fast erosion and dilation by contour processing and thresholding of distance maps, Pattern Recognition Letters, Volume 13, Issue 3, 1992, Pages 161166. DOI:10.1016/01678655(92)900555
 cucim.skimage.morphology.isotropic_dilation(image, radius, out=None, spacing=None)#
Return binary morphological dilation of an image.
This function returns the same result as
skimage.morphology.binary_dilation()
but performs faster for large circular structuring elements. This works by applying a threshold to the exact Euclidean distance map of the inverted image [1], [2]. The implementation is based on: func:cucim.core.operations.morphology.distance_transform_edt. Parameters:
 imagendarray
Binary input image.
 radiusfloat
The radius by which regions should be dilated.
 outndarray of bool, optional
The array to store the result of the morphology. If None is passed, a new array will be allocated.
 spacingfloat, or sequence of float, optional
Spacing of elements along each dimension. If a sequence, must be of length equal to the input’s dimension (number of axes). If a single number, this value is used for all axes. If not specified, a grid spacing of unity is implied.
 Returns:
 dilatedndarray of bool
The result of the morphological dilation with values in
[False, True]
.
Notes
Empirically, on an RTX A6000 GPU, it was observed that
isotropic_dilation
is faster thanbinary_dilation
withdecomposition=None
at radius 12 in 2D and radius 3 in 3D. It becomes faster thanbinary_dilation
withdecomposition="sequence"
at radius 14 in 2D and radius 5 in 3D. In practice, the exact point at which these isotropic functions become faster than their binary counterparts will also be dependent on image shape and content.References
[1]Cuisenaire, O. and Macq, B., “Fast Euclidean morphological operators using local distance transformation by propagation, and applications,” Image Processing And Its Applications, 1999. Seventh International Conference on (Conf. Publ. No. 465), 1999, pp. 856860 vol.2. DOI:10.1049/cp:19990446
[2]Ingemar Ragnemalm, Fast erosion and dilation by contour processing and thresholding of distance maps, Pattern Recognition Letters, Volume 13, Issue 3, 1992, Pages 161166. DOI:10.1016/01678655(92)900555
 cucim.skimage.morphology.isotropic_erosion(image, radius, out=None, spacing=None)#
Return binary morphological erosion of an image.
This function returns the same result as
skimage.morphology.binary_erosion()
but performs faster for large circular structuring elements. This works by applying a threshold to the exact Euclidean distance map of the image [1], [2]. The implementation is based on: func:cucim.core.operations.morphology.distance_transform_edt. Parameters:
 imagendarray
Binary input image.
 radiusfloat
The radius by which regions should be eroded.
 outndarray of bool, optional
The array to store the result of the morphology. If None, a new array will be allocated.
 spacingfloat, or sequence of float, optional
Spacing of elements along each dimension. If a sequence, must be of length equal to the input’s dimension (number of axes). If a single number, this value is used for all axes. If not specified, a grid spacing of unity is implied.
 Returns:
 erodedndarray of bool
The result of the morphological erosion taking values in
[False, True]
.
Notes
Empirically, on an RTX A6000 GPU, it was observed that
isotropic_erosion
is faster thanbinary_erosion
withdecomposition=None
at radius 12 in 2D and radius 3 in 3D. It becomes faster thanbinary_erosion
withdecomposition="sequence"
at radius 14 in 2D and radius 5 in 3D. In practice, the exact point at which these isotropic functions become faster than their binary counterparts will also be dependent on image shape and content.References
[1]Cuisenaire, O. and Macq, B., “Fast Euclidean morphological operators using local distance transformation by propagation, and applications,” Image Processing And Its Applications, 1999. Seventh International Conference on (Conf. Publ. No. 465), 1999, pp. 856860 vol.2. DOI:10.1049/cp:19990446
[2]Ingemar Ragnemalm, Fast erosion and dilation by contour processing and thresholding of distance maps, Pattern Recognition Letters, Volume 13, Issue 3, 1992, Pages 161166. DOI:10.1016/01678655(92)900555
 cucim.skimage.morphology.isotropic_opening(image, radius, out=None, spacing=None)#
Return binary morphological opening of an image.
This function returns the same result as
skimage.morphology.binary_opening()
but performs faster for large circular structuring elements. This works by thresholding the exact Euclidean distance map [1], [2]. The implementation is based on: func:cucim.core.operations.morphology.distance_transform_edt. Parameters:
 imagendarray
Binary input image.
 radiusfloat
The radius with which the regions should be opened.
 outndarray of bool, optional
The array to store the result of the morphology. If None is passed, a new array will be allocated.
 spacingfloat, or sequence of float, optional
Spacing of elements along each dimension. If a sequence, must be of length equal to the input’s dimension (number of axes). If a single number, this value is used for all axes. If not specified, a grid spacing of unity is implied.
 Returns:
 openedndarray of bool
The result of the morphological opening.
Notes
Empirically, on an RTX A6000 GPU, it was observed that
isotropic_opening
is faster thanbinary_opening
withdecomposition=None
at radius 12 in 2D and radius 3 in 3D. It becomes faster thanbinary_erosion
withdecomposition="sequence"
at radius 14 in 2D and radius 5 in 3D. In practice, the exact point at which these isotropic functions become faster than their binary counterparts will also be dependent on image shape and content.References
[1]Cuisenaire, O. and Macq, B., “Fast Euclidean morphological operators using local distance transformation by propagation, and applications,” Image Processing And Its Applications, 1999. Seventh International Conference on (Conf. Publ. No. 465), 1999, pp. 856860 vol.2. DOI:10.1049/cp:19990446
[2]Ingemar Ragnemalm, Fast erosion and dilation by contour processing and thresholding of distance maps, Pattern Recognition Letters, Volume 13, Issue 3, 1992, Pages 161166. DOI:10.1016/01678655(92)900555
 cucim.skimage.morphology.medial_axis(image, mask=None, return_distance=False, *, seed=None)#
Compute the medial axis transform of a binary image.
 Parameters:
 imagebinary ndarray, shape (M, N)
The image of the shape to be skeletonized.
 maskbinary ndarray, shape (M, N), optional
If a mask is given, only those elements in image with a true value in mask are used for computing the medial axis.
 return_distancebool, optional
If true, the distance transform is returned as well as the skeleton.
 seed{None, int, numpy.random.Generator}, optional
If seed is None, the numpy.random.Generator singleton is used. If seed is an int, a new
Generator
instance is used, seeded with seed. If seed is already aGenerator
instance, then that instance is used.New in version 0.19.
 Returns:
 outndarray of bools
Medial axis transform of the image
 distndarray of ints, optional
Distance transform of the image (only returned if return_distance is True)
 Other Parameters:
 random_stateDEPRECATED
Deprecated in favor of seed.
Deprecated since version 23.08.
See also
skeletonize()
Notes
This algorithm computes the medial axis transform of an image as the ridges of its distance transform.
 The different steps of the algorithm are as follows
A lookup table is used, that assigns 0 or 1 to each configuration of the 3x3 binary square, whether the central pixel should be removed or kept. We want a point to be removed if it has more than one neighbor and if removing it does not change the number of connected components.
The distance transform to the background is computed, as well as the cornerness of the pixel.
The foreground (value of 1) points are ordered by the distance transform, then the cornerness.
A cython function is called to reduce the image to its skeleton. It processes pixels in the order determined at the previous step, and removes or maintains a pixel according to the lookup table. Because of the ordering, it is possible to process all pixels in only one pass.
Examples
>>> square = np.zeros((7, 7), dtype=np.uint8) >>> square[1:1, 2:2] = 1 >>> square array([[0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0]], dtype=uint8) >>> medial_axis(square).astype(np.uint8) array([[0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 1, 0, 0], [0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0], [0, 0, 1, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0]], dtype=uint8)
 cucim.skimage.morphology.octagon(m, n, dtype=<clas