mcgrant@walt.cc.utexas.edu (Michael Charles Grant) (12/20/89)
Hello there. I am an undergraduate student at the University of Texas, and for my Senior Design Laboratory I was considering trying to construct a fingerprint recognition system. Unfortunately, I am relatively inexperienced with neural nets...Okay, the fact of the matter is that I have never worked with them, only read about them. Assuming I can construct the hardware for this project, I was considering feeding the neural net a digitized image of the fingerprint, with a bit of contrast enhancement thrown in to improve the distinction between the peaks and valleys. The net would then vote "yes" or "no", telling me that the fingerprint was recognized, and, if possible, specify which trained fingerprint it matched. In any case, there are several questions to ask: 1) In what form would the input be best input (i.e., how many bits of grey, and how many pixels)? I was considering supplying just two bits of grey, centered around the average darkness of the original 6- or 8- bit image. (to account for different skin colors). 2) What should be the structure of the net? I thought that, in order to be flexible, the net should be at least three layers, but the more layers the slower the computation. 3) What level of computing power will I need? The less power I need, the better, because if possible I am going to attempt to implement this as a stand-alone unit which perhaps would be connected to another computer only for training purposes. My final objective is to be able to recognize the fingerprint in five seconds or less. 4) How should it be trained? Will it be flexible enough to recognize the inevitable slight distortions of the fingerprints? Should I train the net with several different images of the same fingerprint? Of course, I will be doing some research for this on my own, and I will be consulting the resident net expert (Dr. Joydeep Ghosh, if you recognize the name), but I thought that if some of you had time to lead me in the right direction, it would be quite helpful. Even if you just refer me to a few books on the subject, I would greatly appreciate your help. (If you mail me a message and I don't reply for awhile, however, it is probably because I will be away on Christmas vacation). Thanks in advance for your help. Michael C. Grant, undergraduate electrical engineer The University of Texas at Austin mcgrant@snowwhite.utexas.edu (Internet) /512/926-7964, /512/471-3238