I'm training a custom CNN (built for academic purpose) to perform Super-Resolution. I based my work on this review.
The input of the network is a RGB color image, so 3 channels of size image_width x image_height.
It is processed through 3 successive convolution layers, as described in the article above :
input > conv9x9 > ReLu > conv1x1 > ReLu > conv5x5 > output
My layers have less filters than described in the article, because of memory limitations.
The last layers has a 3 filters, so it produces a 3-channel output. Then, I convert this 3-channel output back as an image.
During training, no matter what, the colors slowly disappear from the output image. After a few cycles, the output always ends up in black and white. So, I guess, the 3 output channels end up identical.
I don't understand why.