scikit-image 0.15.0
We're happy to announce the release of scikit-image v0.15.0!
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.
0.15 is the first scikit-image release that is only compatible with Python 3.5 and above. Python 2.7 users should strongly consider upgrading to Python 3.5+, or use the 0.14 long term support releases.
New Features
N-dimensional flood fill, with tolerance
Attribute operators
Extension of register_translation to enable subpixel precision in 3D and optionally disable error calculation
unsharp mask filtering
New options connectivity, indices and allow_borders for skimage.morphology.local_maxima and local_minima.
Image translation registration for masked data (skimage.feature.masked_register_translation)
Frangi (vesselness), Meijering (neuriteness), and Sato (tubeness) filters
Allow float->float conversion of any range
Let lower precision float arrays pass through img_as_float
Lazy apply_parallel (allows optimization of dask array operations)
Add range option for histogram.
Add histogram matching
Improvements
Replace morphology.local_maxima with faster flood-fill based Cython version
skivi is now using qtpy for Qt4/Qt5/PySide/PySide2 compatibility (a new optional dependency).
Performance is now monitored by Airspeed Velocity. Benchmark results will appear at https://pandas.pydata.org/speed/
Speed up inner loop of GLCM
Allow tuple to define kernel in threshold_niblack and threshold_sauvola
Add support for anisotropic blob detection in blob_log and blob_dog
API Changes
skimage.transform.seam_carve has been removed because the algorithm is patented.
Parameter dynamic_range in skimage.measure.compare_psnr has been removed. Use parameter data_range instead.
imageio is now the preferred plugin for reading and writing images.
imageio is now a dependency of scikit-image.
regular_grid now returns a tuple instead of a list for compatibility with numpy 1.15
colorconv.separate_stains and colorconv.combine_stains now uses base10 instead of the natural logarithm as discussed in issue 2995.
Default value of clip_negative parameter in skimage.util.dtype_limits has been set to False.
Default value of circle parameter in skimage.transform.radon has been set to True.
Default value of circle parameter in skimage.transform.iradon has been set to True.
Default value of mode parameter in skimage.transform.swirl has been set to reflect.
Deprecated skimage.filters.threshold_adaptive has been removed. Use skimage.filters.threshold_local instead.
Default value of multichannel parameter in skimage.restoration.denoise_bilateral has been set to False.
Default value of multichannel parameter in skimage.restoration.denoise_nl_means has been set to False.
Default value of mode parameter in skimage.transform.resize and skimage.transform.rescale has been set to reflect.
Default value of anti_aliasing parameter in skimage.transform.resize and skimage.transform.rescale has been set to True.
Removed the skimage.test function. This functionality can be achieved by calling pytest directly.
morphology.local_maxima now returns a boolean array
Bugfixes
Correct bright ridge detection for Frangi filter
skimage.morphology.local_maxima and skimage.morphology.local_minima no longer raise an error if any dimension of the image is smaller 3 and the keyword allow_borders was false.
skimage.morphology.local_maxima and skimage.morphology.local_minima will return a boolean array instead of an array of 0s and 1s if the parameter indices was false.
When compare_ssim is used with gaussian_weights set to True, the boundary crop used when computing the mean structural similarity will now exactly match the width of the Gaussian used. The Gaussian filter window is also now truncated at 3.5 rather than 4.0 standard deviations to exactly match the original publication on the SSIM. These changes should produce only a very small change in the computed SSIM value. There is no change to the existing behavior when gaussian_weights is False.
erroneous use of cython wrap around
Speed up block reduce by providing the appropriate parameters to numpy
Add urllib.request again
Repeat pixels in reflect mode when image has dimension 1
Improve Li thresholding
38ad795f14