In this thesis, several attempts are made to improve the performance of a handwritten cursive Hangul(Korean) character recognizer which has been developed by interconnecting hidden Markov models(HMMs).
Variations in handwritten Hangul phoneme is due to not only writers' styles but also nature of Hangul character structure itself. Therefore, instead of one HMM model per every phoneme, we have proposed multiple phoneme models cosidering locations of the phoneme in the character. We believe it would absorb the variations in phoneme writing better. We also proposed using multiple features to increase discrimination power. In order to utilize multiple features, we provided two modelling methods. One is discrete HMM with multiple codebooks and the other is continuous HMM which can model continuous observations directly. With such provisions, input pattern is represented by directional and positional features.
Experiments were conducted with 8,141 Hangul characters written by 10 persons. The experiment results show that the proposed method reduces error rate by 14.94% compared with that of single-model and single-feature method.
