training semantic segmentation for grayscale

Hi, I'm trying to modify the semantic segmetation example by Matlab to make it work with grayscale image

https://uk.mathworks.com/help/vision/examples/semantic-segmentation-using-deep-learning.html

I think it might involve modifying the nutral network to make it work with the grayscale image for training the model.

please suggest how :)

Answer 1

[Srivardhan Gadila]

From the layerGraph, replace the imageInputLayer with new imageInputLayer having value 1 in channel dimension. Make use of the function analyzeNetwork to check for any errors and replace the layers accordingly. Then convert all images of the dataset from RGB to Grayscale.

For the network used in the Example: Semantic Segmentation Using Deep Learning, make the following changes:

imageSize = [720 960 3];
lgraph = deeplabv3plusLayers(imageSize, numClasses, "resnet18");
layers = lgraph.Layers
analyzeNetwork(lgraph)
%% Replace layers and create new layerGraph
newlgraph = replaceLayer(lgraph,'data',imageInputLayer([720 960 1],'Name','input'));
newlgraph = replaceLayer(newlgraph,'conv1',convolution2dLayer(7,64,'stride',[2 2],'padding',[3 3 3 3],'Name','conv1'))
analyzeNetwork(newlgraph)

Use the newlgraph for training on grayscale images.

Answer 2

[KH TOHIDUL ISLAM]

You can use the following function!

pximdsResz = pixelLabelImageDatastore(imds,pxds,'OutputSize',imageSize,...

'ColorPreprocessing','gray2rgb');