Evaluation of HMM based on-line handwritten character recognition with subcharacter model units

摘要

This paper describes a new on-line handwritten character recognition scheme based on subcharacter HMM units and we report base accuracies of proposed method by using simple stroke units that minimize a memory requirement for models and dictionaries. For evaluation, we collected two kinds of handwritten character databases (α set and β set) from students. The α set is a fixed stroke order Kanji database with 1016 character categories of an old and new educational Kanji from 108 people. The other β set consists of about 1200 characters per person by free stroke order from 80 people and includes whole 6353 Kanji of JIS 1st level and 2nd level sets. In experiments, we achieved correct recognition rate of above 94% by using the a set database and achieved about 80% correct rate in 6535 categories recognition by using the β set database that includes 5337 untrained characters. We also show an improvement of recognition accuracy by using writer adaptation technique.