Semi-Automatic Tibetan Component Annotation from Online Handwritten Tibetan Character Database by Optimizing Segmentation Hypotheses

One of important steps in hybrid statistical-structural recognition method for handwritten characters is to label primitives for classifier training and label structural position information for structural recognition. In this paper, we propose a semi-automatic component (primitive) annotation method for online handwritten Tibetan character database. All samples of each character class are over-segmented into sub-structure block sequences. We select correct segmentation points from one of segmented character samples and get component templates of this character class. Other samples of the same character class with sub-structure block sequences are matched with the component templates by optimizing segmentation hypotheses strategy. Character samples segmented by error are re-annotated with minimal human effort at semi-automatic re-annotation module. At last we measure the performance of our component-based recognition method on the character database with component annotation for reference.