Though hand-written Hangul recognition has been studied more than 20 years, a practical system is still far away to be realized. The difficulty is mainly due to infinite number of variations on writing style. Unless recognition algorithm possesses capabilities of adaptation on writing style, hand-written recognition may not be achieved.
In this thesis, stroke matching is used as the basic mechanism of Hangul recognition. based on the strokes, several easily extractable features are computed, such as the number of strokes per character, relative length of a stroke, and stroke slant. major issues are how to build character models with these features and how to select best matching character model for an input character image. In the learning phase, character models are adjusted through updating mean and standard deviation of each feature value. The weight of each feature, which reflects the importance of the feature for computing matching score, is also adjusted in the learning phase.
Experiments were performed with 522 classes of Hangul being used most frequently. After trainning with 6 sets of 522 Hangul, the system achieved 68% recognition rate for 4 sets of Hangul written by the same writers, while the recognition rate for the same trainning data was 81%. Recognition rate will be improved after sufficient trainning with a large number of data.
