I was wondering if anyone knows of a library that can do the following: I need to be able to recognize an arbitrary symbol. It does not need to translate to any known characters, just recognize the symbol as unique, and recall that unique symbol when it encounters it again. Which is basically just OCR trained differently. The other thing is it does not need to scan pages of symbols and try to discern distinct symbols. It will be fed them one at a time. For example, someone will write a symbol on a tablet. It will be then told, what they just wrote is a symbol, and then start over. If someone writes it again it should ideally recognize it. Is this beyond our current OCR abilities?

Try looking into gesture recognition; it shares many of the underlying concepts of OCR.

You can look into Principle Component Analysis as well. It provides a way to classify a vector in terms similarity to a training set. I've heard it's being used to scan handwriting off letters by the US post office.

Neural networks might work too, but it may be a little harder since the symbols are arbitrary and you can't train them yourself.