How can i do multi-output for u-net model

Hi, i just want to know did images regression that create with u-net model can train/predict multi-output? If it's can how can i do.

For my model, i created and trained the network by using Train Convolutional Neural Network for Regression as a guideline which is it have only one output but I need to create the model with the same artchitecture but more output. The multi-ouput that i need to train create from the original output which split into groups based on the values in the array as shown in the picture below.

Note: My old model predict the original image and my new model that i need to try, i want it predict the output 1 to output 4.

Answer 1

[Aishwarya]

Hello,

As per my understanding, you have created a U-Net model using the below mentioned document as reference. Now you wish to create a multi-class U-Net model and split the output mask for each class.

https://www.mathworks.com/help/deeplearning/ug/train-a-convolutional-neural-network-for-regression.html

As the mentioned documentation provides a simple convolution neural network for regression which outputs a single continuous value, I assume you have added up-sampling layers and skip-connections to make the network into U-Net architecture. To modify the network into multi-class U-Net model, consider changing the the last few layers after the last “relu” layer as show in example code below:

conv = convolution2dLayer(1, numClasses); % numClasses is the number of output classes 
softmax = softmaxLayer(); 
outputLayer = pixelClassificationLayer(); 

After getting the output mask (“labelled_mask”) from the U-Net model, each class output can be separated using the example code below.

% Extract each class into a separate image 
num_classes = 4; 
for i = 1:num_classes    
    % Create binary mask for current class 
    class_mask = (labeled_mask == i); 
    
    % Apply binary mask to input image 
    class_img = img .* uint8(class_mask); 
    
    % Show output image 
    figure, 
    imshow(class_img); 
end 

Please refer to below MathWorks documentation for more details about the functions used:

• “softmaxLayer” function: https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.softmaxlayer.html
• “pixelClassificationLayer” function: https://www.mathworks.com/help/vision/ref/nnet.cnn.layer.pixelclassificationlayer.html

I hope this helps!