Recognition of Handprinted Alphanumerics, Symbols and "Katakana" Characters

Abstract

This paper describes a new recognition system for 106 handprinted characters. The physical image of a written character is converted to an original binary pattern. This binary pattern is smoothed and thinned to a skeleton pattern. An input pattern feature string, described by symbols of singular points and direction codes of segments, is extracted from the skeleton pattern. The successful candidate categories are determined by comparing the input pattern feature string to the standard pattern feature strings. The desired category from the candidate categories is selected by the tournament structured decision tree. The system has been developed as the recognition system for a commercial optical character reader, OCR N6370, system. The recognition is carried out by custom ordered LSIs for preprocessing and a micro-computer for feature extraction and decision.