Calculating the co-occurrence matrix of part of an image from a mask

Question

I have the following code in which I am trying to get a portion of a image corresponding to a mask I am given. Then I would like to apply the skimage.feature.glcm on that portion. But I get the error:

glcm = greycomatrix(mancha, [2], [0], levels=None, symmetric = True, normed = True)
File "D:\WinPython-64bit-2.7.13.1ZeroNew\python-2.7.13.amd64\lib\site-packages\skimage\feature\texture.py", line 101, in greycomatrix
assert_nD(image, 2)
File "D:\WinPython-64bit-2.7.13.1ZeroNew\python-2.7.13.amd64\lib\site-packages\skimage\_shared\utils.py", line 178, in assert_nD
raise ValueError(msg_incorrect_dim % (arg_name, '-or-'.join([str(n) for n in ndim])))
ValueError: The parameter `image` must be a 2-dimensional array

The code is:

mask = cv2.imread(pathMask, 0)
cruda = cv2.imread(pathCruda, 0)
imaskc = mask > 0
mancha = cruda[imaskc]

glcm = greycomatrix(mancha, [2], [0], levels=None, symmetric = True, normed = True)
energy = greycoprops(glcm, 'energy')
homogeneity = greycoprops(glcm, 'homogeneity')

I have also tried unsuccesfully with:

labeled_image, nb_labels = ndimage.label(mascara)
blobs = ndimage.find_objects(labeled_image)

glcm = greycomatrix(cruda[blobs[0]]

Any ideas how to get this done?

Thanks!

Answer 1

You can't directly pass a masked image to greycomatrix. The good news is that you can compute the approximate value of Haralick features extracted from a region of interest of an image by introducing small changes in your code.

The basic idea consists in saving a grey level, say 0, to flag those image pixels that fall outside the region of interest (ROI). For this approach to work properly, you need to change the intensity of those pixels inside the ROI whose original intensity was 0 to a different (but similar) value, for example 1. Notice that modifying the image in this way inevitably introduces inaccuracies in the co-occurrence matrix, but as long as your image is sufficiently large and has a smooth histogram, you can safely assume that the obtained features are a pretty good approximation to the exact values. It is also important to note that you have to get rid of the 0^th row and the 0^th column of the co-occurrence matrix in order not to take into account the grey level used to flag non-ROI pixels.

To implement the workaround described above you just need to change the following two lines:

mancha = cruda[imaskc]
glcm = greycomatrix(mancha, [2], [0], levels=None, symmetric=True, normed=True)

to:

mancha = cruda.copy()
mancha[mancha == 0] = 1
mancha[~imaskc] = 0
glcm = greycomatrix(mancha, [2], [0], levels=None, symmetric=True, normed=True)[1:, 1:, :, :]