Interesting work, and here are a few thoughts (in no particular order). 10 seconds seems very high for what's being done here. We can get realtime facial recog

Interesting work, and here are a few thoughts (in no particular order). 10 seconds seems very high for what's being done here. We can get realtime facial recognition (at like 30-60fps). Recognizing chess pieces ought to be trivial by comparison. If you're taking the rules seriously, there should be a lot more rules: there are many (many) invalid board states. It seems strange focusing on just a couple. In fact, it seems strange looking at rules altogether. Why not simply have your algorithm figure out what piece is at what position? SURF or SIFT might be useful here; where you're looking for features that identify specific pieces. In other words, I'm not totally sure if ML is necessary. I would consider processing the image with something like a canny filter and then running a shape detection on the contours. Either way, really cool project, and great write-up!