What is the approach to select your deep neural network architecture? Number of layers, neurons, functions, etc?

Also, what architecture is suggested for a supervised learning application which has input and output images with the size bigger than 1000x1000 pixels?

The NN task is to convert an unfocused image to a focused one according to the samples it has been trained by. We have 2000 unfocused images and the corresponding focused images for the training. The purpose is to train the network with the input-output pair and ask NN to perform focusing on an unfocused image.

Similar questions and discussions