Announcement: scikit-image 0.19.0

We're happy to announce scikit-image v0.19.0!

scikit-image is an image processing toolbox for SciPy that includes algorithms for segmentation, geometric transformations, color space manipulation, analysis, filtering, morphology, feature detection, and more.

For more information, examples, and documentation, please visit our website: https://scikit-image.org

A highlight of this release is the addition of the popular scale-invariant feature transform (SIFT) feature detector and descriptor. This release also
introduces a perceptual blur metric, new pixel graph algorithms, and most functions now operate in single-precision when single-precision inputs are provided. Many other bug fixes, enhancements and performance improvements are detailed below.

A significant change in this release is in the treatment of multichannel images. The existing multichannel argument to functions has been deprecated in favor of a new channel_axis argument. channel_axis can be used to specify which axis of an array contains channel information (with channel_axis=None indicating a grayscale image).

scikit-image now uses "lazy loading", which enables users to access the functions from all skimage submodules without the overhead of eagerly importing all submodules. As a concrete example, after calling "import skimage" a user can directly call a function such as skimage.transform.warp whereas previously it would have been required to first "import skimage.transform".

An exciting change on the development side is the introduction of support for Pythran as an alternative to Cython for generating compiled code. We plan to keep Cython support as well going forward, so developers are free to use either one as appropriate. For those curious about Pythran, a good overview was given in the SciPy 2021 presentation, "Building SciPy Kernels with Pythran" (https://www.youtube.com/watch?v=6a9D9WL6ZjQ).

This release now supports 3.7-3.10. Apple M1 architecture (arm64) support is new to this release. MacOS 12 wheels are provided for Python 3.8-3.10.

New Features

• Added support for processing images with channels located along any array axis. This is in contrast to previous releases where channels were required to be the last axis of an image. See more info on the new channel_axis argument under the API section of the release notes.
• A no-reference measure of perceptual blur was added (skimage.measure.blur_effect).
• Non-local means (skimage.restoration.denoise_nl_means) now supports 3D multichannel, 4D and 4D multichannel data when fast_mode=True.
• An n-dimensional Fourier-domain Butterworth filter (skimage.filters.butterworth) was added.
• Color conversion functions now have a new channel_axis keyword argument that allows specification of which axis of an array corresponds to channels. For backwards compatibility, this parameter defaults to channel_axis=-1, indicating that channels are along the last axis.
• Added a new keyword only parameter random_state to morphology.medial_axis and restoration.unsupervised_wiener.
• Seeding random number generators will not give the same results as the underlying generator was updated to use numpy.random.Generator.
• Added saturation parameter to skimage.color.label2rgb
• Added normalized mutual information metric skimage.metrics.normalized_mutual_information
• threshold_local now supports n-dimensional inputs and anisotropic block_size
• New skimage.util.label_points function for assigning labels to points.
• Added nD support to several geometric transform classes
• Added skimage.metrics.hausdorff_pair to find points separated by the Hausdorff distance.
• Additional colorspace illuminants and observers parameter options were added to skimage.color.lab2rgb, skimage.color.rgb2lab,
  skimage.color.xyz2lab, skimage.color.lab2xyz, skimage.color.xyz2luv and skimage.color.luv2xyz.
• skimage.filters.threshold_multiotsu has a new hist keyword argument to allow use with a user-supplied histogram. (gh-5543)
• skimage.restoration.denoise_bilateral added support for images containing negative values. (gh-5527)
• The skimage.feature functions blob_dog, blob_doh and blob_log now support a threshold_rel keyword argument that can be used to specify a relative threshold (in range [0, 1]) rather than an absolute one. (gh-5517)
• Implement lazy submodule importing (gh-5101)
• Implement weighted estimation of geometric transform matrices (gh-5601)
• Added new pixel graph algorithms in skimage.graph: pixel_graph generates a graph (network) of pixels according to their adjacency, and central_pixel finds the geodesic center of the pixels. (gh-5602)
• scikit-image now supports use of Pythran in contributed code. (gh-3226)

Improvements

• Many more functions throughout the library now have single precision (float32) support.
• Biharmonic inpainting (skimage.restoration.inpaint_biharmonic) was refactored and is orders of magnitude faster than before.
• Salt-and-pepper noise generation with skimage.util.random_noise is now faster.
• The performance of the SLIC superpixels algorithm (skimage.segmentation.slice) was improved for the case where a mask is supplied by the user (gh-4903). The specific superpixels produced by masked SLIC will not be identical to those produced by prior releases.
• exposure.adjust_gamma has been accelerated for uint8 images thanks to a LUT (gh-4966).
• measure.label has been accelerated for boolean input images, by using scipy.ndimage's implementation for this case (gh-4945).
• util.apply_parallel now works with multichannel data (gh-4927).
• skimage.feature.peak_local_max supports now any Minkowski distance.
• Fast, non-Cython implementation for skimage.filters.correlate_sparse.
• For efficiency, the histogram is now precomputed within skimage.filters.try_all_threshold.
• Faster skimage.filters.find_local_max when given a finite num_peaks.
• All filters in the skimage.filters.rank module now release the GIL, enabling multithreaded use.
• skimage.restoration.denoise_tv_bregman and
  skimage.restoration.denoise_bilateral now release the GIL, enabling multithreaded use.
• A skimage.color.label2rgb performance regression was addressed.
• Improve numerical precision in CircleModel.estimate. (gh-5190)
• Add default keyword argument values to skimage.restoration.denoise_tv_bregman, skimage.measure.block_reduce, and skimage.filters.threshold_local. (gh-5454)
• Make matplotlib an optional dependency (gh-5990)
• single precision support in skimage.filters (gh-5354)
• Support nD images and labels in label2rgb (gh-5550)
• Regionprops table performance refactor (gh-5576)
• add regionprops benchmark script (gh-5579)
• remove use of apply_along_axes from greycomatrix & greycoprops (gh-5580)
• refactor gabor_kernel for efficiency (gh-5582)
• remove need for channel_as_last_axis decorator in skimage.filters (gh-5584)
• replace use of scipy.ndimage.gaussian_filter with skimage.filters.gaussian (gh-5872)
• add channel_axis argument to quickshift (gh-5987)
• add MacOS arm64 wheels (gh-6068)

API Changes

• The multichannel boolean argument has been deprecated. All functions with multichannel support now use an integer channel_axis to specify which axis corresponds to channels. Setting channel_axis to None is used to indicate that the image is grayscale. Specifically, existing code with multichannel=True should be updated to use channel_axis=-1 and code with multichannel=False should now specify channel_axis=None.
• Most functions now return float32 images when the input has float32 dtype.
• A default value has been added to measure.find_contours, corresponding to the half distance between the min and max values of the image (gh-4862).
• data.cat has been introduced as an alias of data.chelsea for a more descriptive name.
• The level parameter of measure.find_contours is now a keyword argument, with a default value set to (max(image) - min(image)) / 2.
• p_norm argument was added to skimage.feature.peak_local_max to add support for Minkowski distances.
• skimage.transforms.integral_image now promotes floating point inputs to double precision by default (for accuracy). A new dtype keyword argument can be used to override this behavior when desired.
• Color conversion functions now have a new channel_axis keyword argument (see New Features section).
• SLIC superpixel segmentation outputs may differ from previous versions for data that was not already scaled to [0, 1] range. There is now an automatic internal rescaling of the input to [0, 1] so that the compactness parameter has an effect that is independent of the input image's scaling.
• A bug fix to the phase normalization applied within skimage.register.phase_cross_correlation may result in a different result as compared to prior releases. The prior behavior of "unnormalized" cross correlation is still available by explicitly setting normalization=None. There is no change to the masked cross-correlation case, which uses a different algorithm.

Bugfixes

• Input labels argument renumbering in skimage.feature.peak_local_max is avoided (gh-5047).
• fix clip bug in resize when anti_aliasing is applied (gh-5202)
• Nonzero values at the image edge are no longer incorrectly marked as a boundary when using find_bounaries with mode='subpixel' (gh-5447).
• Fix return dtype of _label2rgb_avg function.
• Ensure skimage.color.separate_stains does not return negative values.
• Prevent integer overflow in EllipseModel.
• Fixed off-by one error in pixel bins in Hough line transform, skimage.transform.hough_line.
• Handle 1D arrays properly in skimage.filters.gaussian.
• Fix Laplacian matrix size bug in skimage.segmentation.random_walker.
• Regionprops table (skimage.measure.regionprops_table) dtype bugfix.
• Fix skimage.transform.rescale when using a small scale factor.
• Fix skimage.measure.label segfault.
• Watershed (skimage.segmentation.watershed): consider connectivity when calculating markers.
• Fix skimage.transform.warp output dtype when order=0.
• Fix multichannel intensity_image extra_properties in regionprops.
• Fix error message for skimage.metric.structural_similarity when image is too small.
• Do not mark image edges in 'subpixel' mode of skimage.segmentation.find_boundaries.
• Fix behavior of skimage.exposure.is_low_contrast for boolean inputs.
• Fix wrong syntax for the string argument of ValueError in skimage.metric.structural_similarity .
• Fixed NaN issue in skimage.filters.threshold_otsu.
• Fix skimage.feature.blob_dog docstring example and normalization.
• Fix uint8 overflow in skimage.exposure.adjust_gamma.
• Work with pooch 1.5.0 for fetching data (gh-5529).
• The offsets attribute of skimage.graph.MCP is now public. (gh-5547)
• Fix io.imread behavior with pathlib.Path inputs (gh-5543)
• Make scikit-image imports from Pooch, compatible with pooch >= 1.5.0 (gh-5529)
• Fix several broken doctests and restore doctesting on GitHub Actions (gh-5505)
• Fix broken doctests in skimage.exposure.histogram and skimage.measure.regionprops_table. (gh-5522)
• Rescale image consistently during SLIC superpixel segmentation. (gh-5518)
• Correct phase correlation in skimage.register.phase_cross_correlation. (gh-5461)
• Fix hidden attribute 'offsets' in skimage.graph.MCP (gh-5551)
• fix phase_cross_correlation for 3D with reference masks (gh-5559)
• fix return shape of blob_log and blob_dog when no peaks are found (gh-5567)
• Fix find contours key error (gh-5577)
• Refactor measure.ransac and add warning when the estimated model is not valid (gh-5583)
• Restore integer image rescaling for edge filters (gh-5589)
• trainable_segmentation: re-raise in error case (gh-5600)
• allow regionprops_table to be called with deprecated property names (gh-5908)
• Fix weight calculation in fast mode of non-local means (gh-5923)
• fix for #5948: lower boundary 1 for kernel_size in equalize_adapthist (gh-5949)
• convert pathlib.Path to str in imsave (gh-5971)
• Fix slic spacing (gh-5974)
• Add small regularization to avoid zero-division in richardson_lucy (gh-5976)
• Fix benchmark suite (watershed function was moved) (gh-5982)
• catch QhullError and return empty array (convex_hull) (gh-6008)
• add property getters for all newly deprecated regionprops names (gh-6000)
• Fix the estimation of ellipsoid axis lengths in the 3D case (gh-6013)
• Fix peak local max segfault (gh-6035)
• Avoid circular import errors when EAGER_IMPORT=1 (gh-6042)
• remove all use of the deprecated distutils package (gh-6044)
• Backport PR #6061 on branch v0.19.x (remove use of deprecated np.int in SIFT) (gh-6062)
• Backport PR #6060 on branch v0.19.x (Fix test failures observed with numpy 1.22rc0) (gh-6063)

Completed Deprecations from Prior Releases

• In measure.label, the deprecated neighbors parameter has been removed (use connectivity instead).
• The deprecated skimage.color.rgb2grey and skimage.color.grey2rgb functions have been removed (use skimage.color.rgb2gray and
  skimage.color.gray2rgb instead).
• skimage.color.rgb2gray no longer allows grayscale or RGBA inputs.
• The deprecated alpha parameter of skimage.color.gray2rgb has now been removed. Use skimage.color.gray2rgba for conversion to RGBA.
• Attempting to warp a boolean image with order > 0 now raises a ValueError.
• When warping or rescaling boolean images, setting anti-aliasing=True will raise a ValueError.
• The bg_label parameter of skimage.color.label2rgb is now 0.
• The deprecated filter parameter of skimage.transform.iradon has now been removed (use filter_name instead).
• The deprecated skimage.draw.circle function has been removed (use skimage.draw.disk instead).
• The deprecated skimage.feature.register_translation function has been removed (use skimage.registration.phase_cross_correlation instead).
• The deprecated skimage.feature.masked_register_translation function has been removed (use skimage.registration.phase_cross_correlation instead).
• The deprecated skimage.measure.marching_cubes_classic function has been removed (use skimage.measure.marching_cubes instead).
• The deprecated skimage.measure.marching_cubes_lewiner function has been removed (use skimage.measure.marching_cubes instead).
• The deprecated skimage.segmentation.circle_level_set function has been removed (use skimage.segmentation.disk_level_set instead).
• The deprecated inplace parameter of skimage.morphology.flood_fill has been removed (use in_place instead).
• The deprecated skimage.util.pad function has been removed (use numpy.pad instead).
• The default mode in skimage.filters.hessian is now 'reflect'.
• The default boundary mode in skimage.filters.sato is now 'reflect'.
• The default boundary mode in skimage.measure.profile_line is now 'reflect'.
• The default value of preserve_range in skimage.restoration.denoise_nl_means is now False.
• The default value of start_label in skimage.segmentation.slic is now 1.

Newly introduced deprecations:

• The multichannel argument is now deprecated throughout the library and will be removed in 1.0. The new channel_axis argument should be used instead. Existing code with multichannel=True should be updated to use channel_axis=-1 and code with multichannel=False should now specify channel_axis=None.

• skimage.feature.greycomatrix and skimage.feature.greycoprops are deprecated in favor of skimage.feature.graycomatrix and
  skimage.feature.graycoprops.

• The skimage.morphology.grey module has been renamed skimage.morphology.gray. The old name is deprecated.

• The skimage.morphology.greyreconstruct module has been renamed skimage.morphology.grayreconstruct. The old name is deprecated.

• see API Changes section regarding functions with deprecated argument names related to the number of iterations. num_iterations and
  max_num_iter are now used throughout the library.

• see API Changes section on deprecation of the selem argument in favor of footprint throughout the library

• Deprecate in_place in favor of the use of an explicit out argument in skimage.morphology.remove_small_objects,
  skimage.morphology.remove_small_holes and skimage.segmentation.clear_border

• The input argument of skimage.measure.label has been renamed label_image. The old name is deprecated.

• standardize on num_iter for paramters describing the number of iterations and max_num_iter for parameters specifying an iteration limit. Functions where the old argument names have now been deprecated are::

  Function Name
  skimage.filters.threshold_minimum
  skimage.morphology.thin
  skimage.restoration.denoise_tv_bregman
  skimage.restoration.richardson_lucy
  skimage.segmentation.active_contour
  skimage.segmentation.chan_vese
  skimage.segmentation.morphological_chan_vese
  skimage.segmentation.morphological_geodesic_active_contour
  skimage.segmentation.slic

• The names of several parameters in skimage.measure.regionprops
  have been updated so that properties are better grouped by the first
  word(s) of the name. The old names will continue to work for
  backwards compatibility. The specific names that were updated are:

  Old Name	New Name
  max_intensity	intensity_max
  mean_intensity	intensity_mean
  min_intensity	intensity_min
  bbox_area	area_bbox
  convex_area	area_convex
  filled_area	area_filled
  convex_image	image_convex
  filled_image	image_filled
  intensity_image	image_intensity
  local_centroid	centroid_local
  weighted_centroid	centroid_weighted
  weighted_local_centroid	centroid_weighted_local
  major_axis_length	axis_major_length
  minor_axis_length	axis_minor_length
  weighted_moments	moments_weighted
  weighted_moments_central	moments_weighted_central
  weighted_moments_hu	moments_weighted_hu
  weighted_moments_normalized	moments_weighted_normalized
  equivalent_diameter	equivalent_diameter_area

• The selem argument has been renamed to footprint throughout the library. The selem argument is now deprecated.

Additional details at: https://raw.githubusercontent.com/scikit-image/scikit-image/v0.19.0/doc/release/release_0.19.rst