Is there any "regression" output layer equivalent to pixelClass​ificationL​ayer?

Hi,

I am working on a denoising Neural Network that has an image as input and outputs a denoised version of the image as output. I have created the network in Matlab, using the Deep Learning Toolbox, and now I need to train it. However, trainNetwork does not accept a Ground Truth that is not a categorical. I've doubled check my network and the issue might be that my last layer is a Classification Layer (pixelClassificationLayer). However, in my case, the problem is a regression one, where the value of each output pixel can correspond to any level of gray. I wondered, is there any regression equivalent to pixelClassificationLayer, or plausible alternative for Regression issues in Deep Learning Imaging?

Thank you very much.

Virginia

[Binu]

There is a regression layer where you basically swap the last layer from classification to regression. Also there are transfer learning methods to convert from a classification to regression . However I have not tried this on a pixelLevel problems.

https://www.mathworks.com/help/deeplearning/examples/convert-classification-network-into-regression-network.html

Have you considered using your gray "levels" as a classes or bins and treat them as a classification problem?

Answer 1

[Sai Bhargav Avula]

Hi, I would recommend you trying the MATLAB way to network. Where you can convert to your layers to a dlnetwork

You can the following link for a detailed understanding of how to define it for your case

<https://www.mathworks.com/help/deeplearning/examples/train-generative-adversarial-network.html>

This way gives you more flexibility for defining your network

[supriya Naik]

Is this dlnetwork possible in matlab2019a ?? Because when I used this technique..it was showing an error that undefined dlnetwork

[Sai Bhargav Avula]

dlnetwork function was introduced in 2019b. Hence you are getting this error. Upgrade the MATLAB to atleast 2019b to use dlnetwork