since the dice are always in a fixed position and they are always oriented the same, it should be fairly easy to just compare the background image (without dice) witht eh new image (with dice) and pick out the delta ( the part that has changed, i.e. the dice). Get shots of each of the combinations and then feed them randomly into a simple perceptron network over a weekend, just randomly shuffle the images and do a slow train (error *.01). Use each pixel as an input into the network, with 11 hidden layer neurons and 11 output neurons. You should end up with a fairly fast and accurate NN that recognizes the result.
