Currently I implemented only neural networks with back propagation. I only tried sigmoid as activation function by now.
Neural network has 2 inputs (x, y coordinates), some hidden layers and 3 outputs (color represented as RGB). It is possible to train network on an image so it remembers it. If you’re not dummy and trained x, y being normalized then it’s obviously possible to do upscaling.
NOTE: it’s possible to upscale bigger images but it requires bigger networks and thus more time to train. My OCaml is not so good so we’re training small models. For now.
