This is indeed neural networks area of work. The biggest problem is finding a (hopefully readily available) training set of symbols to recognize. Coding the neural network itself should be fairly easy and simple as it's just floating point math in the end. I just wrote an k-ary classifier single layer perceptron (not the bleeding edge, I know, but still sufficient for the task) from scratch in just 120 lines of Pascal, minus the training and testing data, and it still includes a bunch of Write(Ln) statements for progress visualization.