Can I apply a graylevel mask on a color image?

Hi everyone, I have an RGB image(n by m by 3) that i segmented it using watershed algorithm. The output of the segmentation is a graylevel image (n by m) which contains the labels assigned to each region of the image. I want to calculate the average of each region's color of the original image but i don't find how to define each region of the original color image. Is it possible to consider the output of the watershed (the graylevel image) as a mask of my original image? If not, can anyone please suggest another solution? Thank you for your answer.

[Adam]

Sounds like it should just be usable as a mask if it is segmented, but you'd need to apply the mask to each layer independently or work out the indices for the full 3d mask, which also isn't too difficult.

[As Bdg]

Thank you for your reply Adam ; but how can i apply the mask to each layer when it is not binary , it contains the labels of the regions, and i want to conserve these labels.

[Adam]

You would have to take 1 label at a time to create a binary mask for that particular label. Or just follow Guillaume's answer. I almost never use regionprops so I don't know much about it, but Guillaume's answers generally work fine.

Answer 1

[Guillaume]

It can be done any numbers of ways. Probably the simplest is to use regionprops on your label image and ask for the 'PixelIdxList' of each blob. This is gives you the linear pixel index of each blob in each channel. You could then just loop over the blobs and channels to compute the mean:

blobprops = regionprops(yourlabelimage, 'PixelIdxList');
channelmeans = zeros(numel(blobprops), size(yourimage, 3));  %using size(img, 3) so that it works with rgb and greyscale images
for channel = 1:size(yourimage, 3)
    imagechannel = yourimage(:, :, channel);
    for blob = 1:numel(blobprops)
        channelmeans(blob, channel) = mean(imagechannel(blobprops(blob).PixelIdxList));
    end
end

Or you can use the 'PixelList' instead, repmat that for each colour channel, convert to linear index, extract the pixels and reshape into 3D to avoid the channel loop:

blobprops = regionprops(yourlabelimage, 'PixelList');
channelmeans = zeros(numel(blobprops), size(yourimage, 3));  %using size(img, 3) so that it 
for blob = 1:numel(blobprops)
   allpixelsidx = reshape(sub2ind(size(yourlabelimage), [repmat(blobprops(blob).PixelList, [3 1]), repelem([1;2;3], size(blobprops(blob).PixelList, 1)]), [], 3);
   channelmeans(blob, :) = mean(yourimage(allpixelsidx), 2);
end

[As Bdg]

Thank you very much Guillaume. In fact , i want to save the mean values in a matrix of the same size as that of my original image and that of the label one. It's like i'm going to modify the values of the label image with the mean of each blob's color.

[Guillaume]

Once you've got the mean it's trivial to replace the blob pixels by the mean value. You've already got the list of blob pixels in PixelIdxList or PixelList. e.g. with the first method:

blobprops = regionprops(yourlabelimage, 'PixelIdxList');
for channel = 1:size(yourimage, 3)
    imagechannel = yourimage(:, :, channel);
    for blob = 1:numel(blobprops)
       imagechannel(blobprops(blob).PixelIdxList) = mean(imagechannel(blobprops(blob).PixelIdxList));  %replace blob pixels by mean for channel          
    end
    yourimage(:, :, channel) = imagechannel;  %and put channel back into image
end

I've just though of a method which is even simpler, using accumarray:

numblobs = max(yourlabelimage(:)) + 1;  %also include backgrounds
subs = [yourlabelimage(:) + 1; yourlabelimage(:) + numblobs + 1; yourlabelimage(:) + 2*numblobs + 1];  %add numblobs to accumulate green, 2*numblobs to accumulate blue
channelmeans = reshape(accumarray(subs, yourimage(:), [], @mean), [], 3);

Note that row 1 of channelmeans is the background mean. row 2 is label 1, etc.

With the above, it is also trivial to assign the mean back to each blob (and background mean):

meanimage = channelmeans(reshape(subs, size(yourimage)));  %That's all!

[As Bdg]

Sorry for the disruption, but it seems that i don't well your answer. I need to have a (m by n) matrix that will be obtained by replacing each label of the initial watershed segmentation by the mean colorvalue of the corresponding region in my original image (m by n by 3). and the watershed lines are removed by setting the new matrix(n by m) = the original one (n by m by 3) at all regions borders (i.e. the watershed lines are replaced by the corresponding colorvalues of the original color image. Can you help me more please? I really don't know how to proceed. Thanks again