Weighted-Area Correlation and Bidirectional Flow of Information in Pattern Recognition

摘要

This paper describes an investigation of two problems in pattern recognition. The first was to estimate empirically the error rate to be expected using weighted-area correlation on single font, uncontrolled, typed characters. A computer program used a learning procedure to create new templates based on those inputs it was unsure of. A total of 650 inputs were used, during which the error rate dropped by almost an order of magnitude to under 1%. The second problem was to test the power of simple bidirectional flow of information in pattern recognition. Even when weighted-area correlation could not positively identify a sample, it could reduce the ambiguity to a list of candidates. This list was never over six out of a possible 26 and averaged between one and three. An unsophisticated procedure involving no learning was used to compare all pairs of elements on the candidate list in an attempt to eliminate one of the pair. The results show the value of bidirectional flow of information. Finally a more sophisticated system is described for recognizing typed letters, and the form in which bidirectional flow of information was used in the study is considered in more general terms.