Repository files navigation

Mahotas is a library of fast computer vision algorithms (all implementedin C++ for speed) operating over numpy arrays.

Python versions 3.7+, are supported.

Notable algorithms:

• watershed
• convex points calculations.
• hit & miss, thinning.
• Zernike & Haralick, LBP, and TAS features.
• Speeded-Up Robust Features(SURF), a form of localfeatures.
• thresholding.
• convolution.
• Sobel edge detection.
• spline interpolation
• SLIC super pixels.

Mahotas currently has over 100 functions for image processing andcomputer vision and it keeps growing.

The release schedule is roughly one release a month and each releasebrings new functionality and improved performance. The interface is verystable, though, and code written using a version of mahotas from yearsback will work just fine in the current version, except it will befaster (some interfaces are deprecated and will be removed after a fewyears, but in the meanwhile, you only get a warning). In a fewunfortunate cases, there was a bug in the old code and your results willchange for the better.

Please citethe mahotas paper (seedetails below underCitation) if you use it in a publication.

This is a simple example (using an example file that is shipped withmahotas) of calling watershed using above threshold regions as a seed(we use Otsu to define threshold).

# import using ``mh`` abbreviation which is common:importmahotasasmh# Load one of the demo imagesim=mh.demos.load('nuclear')# Automatically compute a thresholdT_otsu=mh.thresholding.otsu(im)# Label the thresholded image (thresholding is done with numpy operationsseeds,nr_regions=mh.label(im>T_otsu)# Call seeded watershed to expand the thresholdlabeled=mh.cwatershed(im.max()-im,seeds)

Here is a very simple example of usingmahotas.distance (whichcomputes a distance map):

importpylabaspimportnumpyasnpimportmahotasasmhf=np.ones((256,256),bool)f[200:,240:]=Falsef[128:144,32:48]=False# f is basically True with the exception of two islands: one in the lower-right# corner, another, middle-leftdmap=mh.distance(f)p.imshow(dmap)p.show()

(This is undermahotas/demos/distance.py.)

How to invoke thresholding functions:

importmahotasasmhimportnumpyasnpfrompylabimportimshow,gray,show,subplotfromosimportpath# Load photo of mahotas' author in greyscalephoto=mh.demos.load('luispedro',as_grey=True)# Convert to integer values (using numpy operations)photo=photo.astype(np.uint8)# Compute Otsu thresholdT_otsu=mh.otsu(photo)thresholded_otsu= (photo>T_otsu)# Compute Riddler-Calvard thresholdT_rc=mh.rc(photo)thresholded_rc= (photo>T_rc)# Now call pylab functions to display the imagegray()subplot(2,1,1)imshow(thresholded_otsu)subplot(2,1,2)imshow(thresholded_rc)show()

As you can see, we rely on numpy/matplotlib for many operations.

If you are usingconda, you can install mahotas fromconda-forge using the following commands:

conda config --add channels conda-forgeconda install mahotas

You will need python (naturally), numpy, and a C++ compiler. Then youshould be able to use:

pip install mahotas

You can test your installation by running:

python -c"import mahotas as mh; mh.test()"

If you run into issues, the manual has moreextensive documentation onmahotasinstallation,including how to find pre-built for several platforms.

If you use mahotas on a published publication, please cite:

Luis Pedro Coelho Mahotas: Open source software for scriptablecomputer vision in Journal of Open Research Software, vol 1, 2013.[DOI]

In Bibtex format:

@article{mahotas,author = {Luis Pedro Coelho},title = {Mahotas: Open source software for scriptable computer vision},journal = {Journal of Open Research Software},year = {2013},doi = {https://dx.doi.org/10.5334/jors.ac},month = {July},volume = {1}}

You can access this information using themahotas.citation() function.

Development happens on github(https://github.com/luispedro/mahotas).

You can set theDEBUG environment variable before compilation to get adebug version:

export DEBUG=1python setup.pytest

You can set it to the value2 to get extra checks:

export DEBUG=2python setup.pytest

Be careful not to use this in production unless you are chasing a bug.Debug level 2 is very slow as it adds many runtime checks.

TheMakefile that is shipped with the source of mahotas can be usefultoo.make debug will create a debug build.make fast will create anon-debug build (you need tomake clean in between).make test willrun the test suite.

Documentation:https://mahotas.readthedocs.io/

Issue Tracker:github mahotasissues

Mailing List: Use thepythonvision mailinglist for questions,bug submissions, etc. Or ask onstackoverflow (tagmahotas)

Main Author & Maintainer:Luis Pedro Coelho(follow ontwitter orgithub).

Mahotas also includes code by Zachary Pincus [from scikits.image], PeterJ. Verveer [from scipy.ndimage], and Davis King [from dlib], ChristophGohlke, as well asothers.

Presentation about mahotas for bioimageinformatics

For more general discussion of computer vision in Python, thepythonvision mailinglist is a muchbetter venue and generates a public discussion log for others in thefuture. You can use it for mahotas or general computer vision in Pythonquestions.

• Fix bug in Haralick features and NumPy 2 (thanks to @Czaki, see#150)

• Fix bug that stopped mahotas from working on Windows

• update for NumPy 2
• Add deprecated warning for freeimage

• Update build system (thanks to @Czaki, see #147)

• Fix code for C++17 (issue #146)

• Fix freeimage testing (and make freeimage loading more robust, see #129)
• Add GIL fixed (which triggered crashes in newer NumPy versions)

• Update to newer NumPy
• Build wheels for Python 3.9 & 3.10

• Convert tests to pytest
• Fix testing for PyPy

• Build wheels automatically (PR #114 bynathanhillyer)

• Fix FreeImage detection (issue #108)

• Fix co-occurrence matrix computation (patch by @databaaz)

• Fix compilation on Windows

• Make watershed work for >2³¹ voxels (issue #102)
• Remove milk from demos
• Improve performance by avoid unnecessary array copies incwatershed(),majority_filter(), and color conversions
• Fix bug in interpolation

• Upgrade code to newer NumPy API (issue #95)

• Fix bug in Bernsen thresholding (issue #84)

• Fix distribution (add missingREADME.md file)

• Fixresize\_to return exactly the requested size
• Fix hard crash when computing texture on arrays with negative values (issue #72)
• Addeddistance argument to haralick features (pull request #76, byGuillaume Lemaitre)

• Addfilter\_labeled function
• Fix tests on 32 bit platforms and older versions of numpy

• Addedmahotas-features.py script
• Add short argument to citation() function
• Add max_iter argument to thin() function
• Fixed labeled.bbox when there is no background (issue #61, reportedby Daniel Haehn)
• bbox now allows dimensions greater than 2 (including when using theas_slice andborder arguments)
• Extended croptobbox for dimensions greater than 2
• Added use_x_minus_y_variance option to haralick features
• Add functionlbp_names

• Improve memory handling in freeimage.write_multipage
• Fix moments parameter swap
• Add labeled.bbox function
• Add return_mean and return_mean_ptp arguments to haralickfunction
• Add difference of Gaussians filter (by Jianyu Wang)
• Add Laplacian filter (by Jianyu Wang)
• Fix crash in median_filter when mismatched arguments are passed
• Fix gaussian_filter1d for ndim > 2

• Add PIL based IO

• Export mean_filter at top level
• Fix to Zernike moments computation (reported by Sergey Demurin)
• Fix compilation in platforms without npy_float128 (patch by GabiDavar)

• Add minlength argument to labeled_sum
• Generalize regmax/regmin to work with floating point images
• Allow floating point inputs tocwatershed()
• Correctly check for float16 & float128 inputs
• Make sobel into a pure function (i.e., do not normalize its input)
• Fix sobel filtering

• Explicitly set numpy.include_dirs() in setup.py [patch by AndrewStromnov]

• Export locmax|locmin at the mahotas namespace level
• Break away ellipse_axes from eccentricity code as it can be usefulon its own
• Addfind() function
• Addmean_filter() function
• Fixcwatershed() overflow possibility
• Make labeled functions more flexible in accepting more types
• Fix crash inclose_holes() with nD images (for n > 2)
• Remove matplotlibwrap
• Use standard setuptools for building (instead of numpy.distutils)
• Addoverlay() function

• Fix crash in close_holes() with nD images (for n > 2)

• Better error checking
• Fix interpolation of integer images using order 1
• Add resize_to & resize_rgb_to
• Add coveralls coverage
• Fix SLIC superpixels connectivity
• Add remove_regions_where function
• Fix hard crash in convolution
• Fix axis handling in convolve1d
• Add normalization to moments calculation

See theChangeLogfor older version.