I'm trying to implement a neural network that is able to generate an image indicating territory occupation given a board state for GO (a strategy board game). Input images are 19x19x1 grayscale images, with white pixels indicating white pieces, black pixels indicating black pieces, and gray pixels indicating unoccupied areas. Output images are 19x19x1 grayscale images with white pixels indicating white territory, black pixels indicating black territory, and gray areas indicating unassigned territories. A sample input and desired output image is as follows:
The images are quite small, so just to give an overview of trends I noticed: - Pixels surrounded by pixels of opposite colors are 'captured' pieces and therefore part of opponent territory - Two 'eyes' or closed groups of pieces comprising at least two open intersections are invincible or confirmed territory
While I'm not looking for exact specifications of network layers etc., I was hoping I could be given some direction as to what type of network to use, and what it should comprise. Looking at MATLAB documentation, I've found info about semantic segmentation, and autoencoder networks but neither of these seem particularly helpful. I know the question is a little broad, but I just need some direction more than anything. This kind of image recognition problem is a first for me.