Trying to understand masking

Question

I've been trying to understand masking and how it works with image filters. I'm using the following code to try to develop my understanding.

import scipy.ndimage as ndi
import matplotlib.pyplot as plt
import numpy as np

#   Generate a random binary mask
np.random.seed(seed=182)
mask = np.random.randint(2, size=(901, 877))

img = np.random.rand(901, 877)

img_masked = np.ma.masked_array(img, mask = mask)
img_masked_filtered = ndi.median_filter(img_masked, size=10)
img_unmasked_filtered = ndi.median_filter(img, size=10)

median_masked = np.ma.median(img_masked)
median_unmasked = np.ma.median(img)

In the results, median_unmasked != median_masked as I expect, but img_masked_filtered == img_unmasked_filtered which I don't want. scipy.ndimage.median_filter does exactly the job I need, but it doesn't work with masked images. What can I use that will do the same thing as the median filter, but which will work on a masked image?

The weird size I'm using for the array is because that's the size of the image I eventually want to filter.

Answer 1

The ndimage filters do not respect masked arrays' masks. Instead, "mask" an ordinary NumPy array with nan values, and then use ndimage.generic_filter to call np.nanmedian:

import scipy.ndimage as ndi
import matplotlib.pyplot as plt
import numpy as np

np.random.seed(seed=182)
# h, w = 901, 877
h, w = 10, 10
mask = np.random.randint(2, size=(h, w))
img = np.random.rand(h, w)
img_masked = np.where(mask, img, np.nan)

size = 3
img_masked_median = ndi.generic_filter(img_masked, np.nanmedian, size=size)
img_unmasked_median = ndi.median_filter(img, size=size)

fig, ax = plt.subplots(nrows=2, ncols=2)
ax[0,0].imshow(img)
ax[0,0].set_title('img')
ax[0,1].imshow(img_masked)
ax[0,1].set_title('img_masked')
ax[1,0].imshow(img_unmasked_median)
ax[1,0].set_title('img_unmasked_median')
ax[1,1].imshow(img_masked_median)
ax[1,1].set_title('img_masked_median')
plt.show()