Online hand-written character recognition based on stroke HMMs

摘要

We propose a new approach of on-line handwritten character recognition that fully utilize continuous speech recognition technology. Basic idea of this approach is that we treat a Kanji character as a sequence of strokes, and this is similar to a relation of sentence speech and phonemes in continuous speech recognition. Therefore, recognition system basically consists of input features, HMMs, dictionaries and decoder. In our method, velocity vectors of pen movement are input features, the pen movements in the 8 directions are modeled by 25 kinds of substroke HMM, Kanji dictionary is hierarchically defined by substroke unit, and recognition results are obtained as maximum likelihood path searched by Viterbi decoder. The advantages of this approach are the following: The memory requirement for the dictionary and models is very small. All characters defined in the dictionary can be recognized by stroke HMMs trained by small amount of data. The writer adaptation of HMM is easily performed by writing a few sample characters. Furthermore, it can recognize characters written with no sight and therefore it will be possible to use it as a new modality of human-interface.