For the character recognition problem, it is known that the prior knowledge about the structure of patterns can be utilized as a guidance to obtain an accurate match efficiently. Naturally, the strokes, which are the primitive components of Chinese characters, play an important role to guide the correct recognition. Based on this high-level structural information, a hierarchical deformation model is proposed to describe the deformation of on-line cursive Chinese characters. The new approach consists of two levels of match processes. First, the attributed string editing algorithm matches two sequences of turn points extracted from the input and the reference characters to determine the stroke matches. Next, the constrained parabola transformation is used to reduce the difference between the matched strokes appropriately. Experimental results show that the hierarchical deformation model is a quite accurate approximation to the deformation of cursive Chinese characters with much lower computational cost. Furthermore, the distance measure between deformable characters derived in this paper is robust enough to greatly improve the performance of practical recognition systems.
ASJC Scopus subject areas
- Signal Processing
- Computer Vision and Pattern Recognition
- Artificial Intelligence