A Digital Simulation of an Aided Adaptive Character Reading Machine

Abstract

A simulation, on an IBM 709 computer, of a pattern-recognition system using an initial man-machine learning phase. Transformations on a deformed set of 48 samples of each of ten numerals are used to form separation filters, while a second set of 480 similarly varied numerals serve as the "unknown" characters that are examined. Measured probability density distributions of the inked areas of all characters are established, and a weighted stencil or filter is created to distinguish each character relative to the possible set of characters. This experiment demonstrates the extent to which the actual value of the best "score of match" between the unknown and each character in the set provides confidence in recognition. Whenever the best score is too low, it is possible to call for more complex processes to aid recognition permitting the construction of recognition systems of greater accuracy than the basic reading mechanism.