Scikit-Image histogram matching doesn't work on numpy masked arrays. This gist is a workaround for this issue. It shows how to do histogram matching when you have an image and need to exclude certain values from being considered as part of the image histogram.

skimage-exposure-match-histograms-with-masked-array.py

from skimage import io, exposure
import numpy as np

# Histogram matching with masked image
def match_histograms(image, reference, image_mask, fill_value=0):
    masked_image = np.ma.array(image, mask=image_mask)
    matched = np.ma.array(np.empty(image.shape, dtype=image.dtype),
                          mask=image_mask, fill_value=fill_value)
    
    for channel in range(masked_image.shape[-1]):
        matched_channel = exposure.match_histograms(masked_image[...,channel].compressed(), 
                                                    reference[...,channel].ravel())
        
        # Re-insert masked background
        mask_ch = image_mask[...,channel]
        matched[..., channel][~mask_ch] = matched_channel.ravel() 
    
    return matched.filled()


# Load images
reference_img = io.imread("/path/to/reference_image.jpg") / 255
image = io.imread("/path/to/image.jpg") / 255

# Get a mask that matches image.shape, with mask being where pixel val across channels is 0
mask = np.repeat(np.expand_dims(np.all(img == 0, axis=2), axis=2), repeats=3, axis=2)

# Do the masked histogram matching
output = match_histograms(image, reference_img, mask, fill_value=0)

Tayden this is awesome! I actually need to go from a masked reference image to a masked target image. I'm going to use your gist here as a starting point to getting something working.

This is very cool and just what I was looking for! Just one thing I noticed. The fill value passed to the function is not relayed to the masked array, but set to 0.

Ah good catch! I'll update it

Thanks so much for sharing this!

I slightly modified the code, in case both images contain black background elements / masks:

# Histogram matching with masked image
def match_histograms(image, reference, image_mask, reference_mask, fill_value=0):
    # adapted from https://gist.github.com/tayden/dcc83424ce55bfb970f60db3d4ddad18
    # to include masks for the input and reference image
    masked_source_image = np.ma.array(image, mask=image_mask)
    masked_reference_image = np.ma.array(reference, mask=reference_mask)
    
    matched = np.ma.array(np.empty(image.shape, dtype=image.dtype),
                          mask=image_mask, fill_value=fill_value)
    
    for channel in range(masked_source_image.shape[-1]):
        matched_channel = exposure.match_histograms(masked_source_image[...,channel].compressed(), 
                                                    masked_reference_image[...,channel].compressed())
        
        # Re-insert masked background
        mask_ch = image_mask[...,channel]
        matched[..., channel][~mask_ch] = matched_channel.ravel() 
    
    return matched.filled()

Cool, why this is not implemented in the main exposure.match_histograms?