In this thesis, we proposed a new representation scheme for on-line handwritten Hangul recognition, based on hiearchical curve segmentation, which is robust even when there is great variance of writing style.
A character is hierarchically represented by 3 layers-jaso, curve, and arc. A character is composed of jaso-chosung, jungsung, and jongsung. A jaso is composed of curves, and each curve is composed of arcs. Each of these layers also includes relations of its components. Jaso model is constructed by learning features. Each jaso component of a character is matched against jaso models according to jaso combining rules, and the match degrees are combined.
Experiments were conducted with 12 sets of 506 hangul characters written by 8 people. Among the sample data sets, 8 sets were used to construct jaso models. The system achieved 95% recognition accuracy for the training data, and 85% for the test data. When 5 candidate characters were generated, the system got 100% accuracy for the training data and 96% for the test data.
