kilmer@hq.af.mil (09/19/90)
I am looking for information on which learning architecture is best suited for feature extraction. I know it's hard to judge one from another except in individual cases...so I'll briefly give you a summary of what I'm trying to do. I've been delving into the various aspects of OCR. I first started by creating a backprop net and training it on a font consisting of 94 characters. The network converged on the solution, and was able to identify the 8x14 character matrix. It even worked well at identifying those characters with 5 to 10 percent noise. Unfortunately, it works on that size character only. I was wondering how one would go about training a net to recognize features. I think (although I'm very new to NN) that the hidden layers store features in their particular weights and thresholds, or at least generalizations of the input vectors...??? If this were the case, you could train the net with characters that have similar features (such as a 'p' and a 'q') and have those inputs match with a feature identifying output: e.g: p-> 0010100 q-> 0001100 / \ / \ Desc bar circle(arc) Desc bar circle(arc) on left on right of circle of circle Would this work?? If so, what net would be best for this type of association?? Also, what about feature deformation, translation, and size independence? Any help would be appreciated. Richard. -- .-------------------------------------------------------------------------. | Richard Kilmer Kilmer@Opsnet-Pentagon.af.mil | | VAX Systems Analyst (AKA Kilmer@26.24.0.26) | `-------------------------------------------------------------------------' -- -------------------------------------------------------------------------. | Richard Kilmer Kilmer@Opsnet-Pentagon.af.mil | | VAX Systems Analyst (AKA Kilmer@26.24.0.26) | `-------------------------------------------------------------------------'