Subset Methods for Recognizing Distorted Patterns

At an application-independent level of generality, the problem of recognizing noisy distorted patterns is discussed. Practical techniques for recognizing, for instance, speech, characters, vector-cardiograms, and fingerprints, are designed to tolerate minor distortions of patterns, but not to tolerate any distortion that changes any pattern into a further pattern that belongs to a different recognition class. In a given practical application, only a particular class of distortions, which we call admissible distortions, should be tolerated, and this class must somehow be defined. Different methods, in which definitions of sets of admissible distortions of parts of patterns are used for deciding whether distortions of entire patterns are or are not admissible, are compared. These methods are of interest because they are more economical than other known general methods. Theory suggests that lower recognition error rates should be obtained with an iterative, rather than with a structurally comparable noniterative, method of discriminating between admissible and nonadmissible distortions. To test this experimentally, at least in character recognition, the work has been taken through a phase of practical development, and computer simulation results are reported.