Is it possible to use deep network designer App to design CNN for image enhancement?

I have matlab R2020a and I want to use deep network designer to design my own network for image enhancement. However, I found that the available layers and network examples in deep network designer App are for the task of image classification or regression. ( i.e. the network input is an image and the network output is a label not an image). and I want to design and train a network for image enhancement in which that the network recieves image as an input and outputs an image also not a label. Does anyone know is it possible to do that and how?

[Image Analyst]

You can build your own network from scratch if you want. I believe you just drag and drop components onto the design layout.

[ghada sandoub]

I already tried to build the network from the scratch. My network needs only some convolutional layers, and ReLUs. however, I found some limitations in this App like :

(1) I should start the network with imageinput layer with specified input size, However, I want the network to process input image of any size without resizing the input size.

(2) I need the output to be an image not a label. I found all output layers in the deep network designer related to classification or regression in which the output is a label.

for more details: I want to use CNN for image enhancement, there is some stages to do that like: feature extractin, non-linear mapping and reconstraction. the input for the CNN is an image and the output should be the enhanced image. I want to know how to do this in matlab using deep network designer. Any help will be appreciated.

Answer 1

[Aniket]

Hi @ghada sandoub,

To address your first question, the size of input data needs to be provided to “ImageInputlayer”. If you want the network to process input images of any size, you can resize the image before calling the network. Refer to the documentation on "imresize" function in MATLAB:

https://www.mathworks.com/help/releases/R2020a/matlab/ref/imresize.html

For the second question regarding image as an output, you can use “RegressionOutputLayer” in Deep Network Designer:

https://www.mathworks.com/help/releases/R2020a/deeplearning/ref/nnet.cnn.layer.regressionoutputlayer.html

The following example shows an image-to-image regression in Deep Network Designer to generate a super resolution image as output: https://www.mathworks.com/help/releases/R2020b/deeplearning/ug/image-to-image-regression-in-deep-network-designer.html

I hope this helps you!