ResNet50 on multi-spectral image segmentation

7 views (last 30 days)
Is there a way to use any pretrain network (not necessarily Resnet) to segment multispectral images in MATLAB?
deeplabv3plusLayers
only allows [height width 3] or [height width] input images. While I tried bypassing the error deeplabv3plusLayers returns, when I used trainNetwork I get an error referring to the wrong input data 224x224xN.
Can the first convolutional layer of the pretrained network be replaced to process more than 3 channels? An example done in python can be found here.

Accepted Answer

Srivardhan Gadila
Srivardhan Gadila on 13 Jul 2020
You can copy the layerGraph of the pretrained network and change the imageInputLayer, the first convolutionLayer to match the input image channel dimension & convolution filter dimensions. Then you can freeze/unfreeze the existing pretrained weights during training the new network accordingly.
You can do something like below:(N=50)
imageSize = [224 224 3];
% Specify the number of classes.
numClasses = 10;
N = 50;
% Create DeepLab v3+.
lgraph = deeplabv3plusLayers(imageSize, numClasses, "resnet50");
analyzeNetwork(lgraph)
layers = lgraph.Layers
%%
newlgraph = replaceLayer(lgraph,'input_1',imageInputLayer([224 224 N],'Name','input'));
newlgraph = replaceLayer(newlgraph,'conv1',convolution2dLayer(7,64,'stride',[2 2],'padding',[3 3 3 3],'Name','conv1'))
analyzeNetwork(newlgraph)

More Answers (0)

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!