How can I use semantic segmentation for gray scale images?

Hi,

I labeled gray scale images using Pixel Label in Image Labeler app (MATLAB 2018a). I am using semantic segmentation (https://www.mathworks.com/help/vision/examples/semantic-segmentation-using-deep-learning.html) to classify different objects in the images. When I ran this part of the code (I changed the third element in the imageSize to 1 because I am using the gray scale images):

imageSize = [144 176 1];
numClasses = numel(classes);
lgraph = segnetLayers(imageSize,numClasses,'vgg16');

I got this error:

imageSize must have three elements and the third element must be 3 when creating SegNet based off of VGG-16 or VGG-19.

    Error in segnetLayers (line 170)
        iCheckImageSizeHasThreeElementsForVGG(args.imageSize);

    Error in semanticSegmentation (line 23)
    lgraph = segnetLayers(imageSize,numClasses,'vgg16');

Does SegNet support gray scale images? If yes, how can I solve the issue?

Thank you,

Abbas

[Aydin Sümer]

Did you solve this problem ? I'm getting same errror.

Answer 1

[Andrey Gizdov]

Hi,

I was having exactly the same issue as you. To solve it, I created a simple functions which checks for the number of color channels in each of image of the 'imds' ImageDatastore from your example. The function is as follows:

function [] = checkDataSize(imds)
%Check if all provided training images are in RGB format

imagesPath = [imds.Files];

for i=1:numel(imagesPath)
   currentImage = readimage(imds, i);
   [~, ~, colorChannels] = size(currentImage);
   if colorChannels == 1 || colorChannels == 2
       newImage = cat(3, currentImage, currentImage, currentImage);
       imwrite(newImage, imagesPath{i});
       fprintf("Found and overwrote image with a single color channel in path %s \n", imagesPath{i});
   end
end

end

Probably not the best solution, but it works around the problem and did the trick for me.

Hope that helps!

[soheib hadj moussa]

i have the same problem have u solution?