That might be doable. One of my fellow students designed a jigsaw puzzle solver that worked by looking at the shape of the piece, rather than a picture. However, I have worked in this kind of area and if you can guarantee little rotation, working out the translational offset isn't too bad. I'll give this some thought.