This thesis presents a method of applying structural and global shape knowledge to hidden Markov model (HMM) based Korean handwriting recognition. Although successfully used in the Korean handwriting recognition area, HMM needs structural and global shape knowledge because of two problems: weak discrimination power of statistical method and insufficient information of local features.
To apply structural shape knowledge, graphemes are divided into basic strokes of lines and circles by analyzing relationship between HMM self transition and a character input. Then, geometric features such as distance and angle between basic strokes are checked to verify graphemes. Position relation between graphemes is also checked to eliminate unusual segmentation of a character.
To apply global shape knowledge, three features such as accumulated angle change, duration information of basic strokes and pen-up movement ratio are used. The accumulated angle change is summation of angle change of pen movement in the grapheme. The duration information of basic strokes is related with the length ratio between basic strokes. The pen-up movement ratio is used for discrimination between graphemes and ligatures.
All the knowledge is merged into HMM recognition system in the probabilistic framework. The HMM grapheme model generates grapheme boundary hypothesis which is verified by the structural and global knowledge. In the experiment, we use three kinds of data sets amount to 15,250,~ 3,127, ~ 16,427 characters respectively. The result shows improved average recognition rate from 94.8% to 95.6% with reducing 14.5% error.