Noise immunization of a neural fuzzy intelligent recognition system by the use of feature and rule extraction technique

The performance of a neural fuzzy intelligent recognition system (NFIRS) which recognizes varied levels of noise corrupted characters was investigated. The number of regions in the universe of discourse of the input space was first arbitrarily selected. Then, the centers of these regions were self organized by feeding the system with a 256-pixel alphabet and algebraic training samples to the Kohonen competitive learning network. Based on the reallocated centers, we tried several combinations of varied rule region product in order to generate a smaller set of fuzzy rules. We fixed the number of features for simulation, and to simplify and isolate the effect of rule extraction. Simulation results showed a NFIRS that uses a set of thirty six sampling data set as the training input will generate a set of thirty six if-then fuzzy rules which can be used to recognize a corrupted testing data set without sacrificing the rate of recognition under varied conditions.