A robust real-timed recognizer of printed chinese characters

摘要

A Bayes recognizer supporting 3770 printed (non-simplified) Chinese characters with sizes no smaller than 2.5 mm2 in various fonts has been built with an overall recognition rate of 96% on characters found in general printed materials. These characters, according to their usage frequencies, are divided into two sets and two decision trees, one for each set, which are used to reduce the recognition time by factors of 21 and 3, respectively. These two sets together include over 99.9% of the characters one can find in any piece of text. Storage requirement for the entire recognition system amounts to 4.2 MB at any one time while the system throughput is 20 characters per second running on an IBM RS 6000/320 workstation. Such a performance compares favorably against any printed Chinese character recognizer known today and the key to success lies in the partitioning of the characters according to their usage frequencies and a branch and bound tree search algorithm with likelihood metric.