In this paper, we propose a novel stroke segmentation method and two modified relaxation methods with which to recognize handwritten Chinese characters. The new stroke segmentation method is based on thinning and quadratic equation fit. After stroke segmentation and size normalization, redundant strokes are detected and deleted. With the deletion of redundant strokes, only prominent strokes that must appear in the writing of a character are stored in the database, greatly decreasing the matching time and complexity. The features of each stroke are then extracted. Using the features of a character, two modified relaxation matching methods are introduced to recognize the unknown input character. After relaxation matching, certain rules are devised to detect certain distorted characters, and also to resolve ambiguity when it happens. Some of these distorted characters will be reconstructed and their feature vector lists will be modified to reflect the change. It is then sent back to the feedback relaxation matching process once again. Experiments are conducted on 300 constrained handwritten characters written by two people; 13 similar characters and 42 distorted characters are also tested to verify the validity of the approach. The success rate for stroke segmentation is 93.3%. Two proposed matching methods are also efficient and suitable for stroke-based Chinese character recognition. The overall recognition rate is 98.7%, without using Rules 2-5 and without considering the abnormal connection and intersection of strokes.
