I have a pretty unique requirement for a ML model.
Essentially, I have a rules-based image transform that takes image X and creates Y.
I would like to fit a neural network that maps this image transform hence Y = H(X) where H is the neural network.
Now, obviously a CNN or a GAN or Unet would be useful here, but heres my primary requirement: The jacobian needs to be fast to compute. By Jacobian I mean the output pixel change with respect to the input pixel change. I need this value because the NN will then be used in an optimization task for the image. (The output image will be used to calculate an FOM, then everything will be backpropagated to the input image, which will then be adjusted.)
Im thinking of just using a shallow CNN, but if anyone has any better solutions that would be great.
