Research On Unconstrained On-line Handwritten Chinese Character Feature Extraction And Recognition Fusion

Handwritten character recognition has been one of the most important topic in the field of pattern recognition for many years. Many researchers have made greate effort for it. The technology of handwritten character recognition has been widely used for human-computer interaction in many handheld devices after more than fourty years of effort by researchers. Handwritten character recognition can be divided into off-line character recognition and onl-ine character recognition. Unconstrained on-line handwritten Chinese character recognition technology can let the user feel freely to input no matter in any style, speed even any angle. However, unconstrained on-line handwritten Chinese character recognition is still a problem because Chinese character set is very large, the structure is complex and the style is randomly.Feature extration is a key step in pattern recognition and the recognition result is greatly dependent on it. It is necessary to do deeply research into feature extration when we research the handwritten character recognition. Many off-line and on-line character recognition methods have been proposed. How to concurrent use off-line and on-line recognition mothods for on-line Chinese character recognition is a problem to be worth studying.In order to solve the above problems, the research has done in this thesis is listed as follows:1. Some mainstream feature extraction methods and classifiers for handwritten character recognition are introduced in this thesis, including Gabor feature, Grad feature, 8 direction feature, direction change feature, distance classifier and MQDF classifier. The performance of these methods have been compared by experiments in large scale Chinese character recognition.2. Imaginary stroke model has been successfully implemented to solve the various writing styles problem due to cursive writing. However, this approach also causes confusions among characters with similar but actually different trajectory. Both the benefit and the defect of the imaginary stroke model have been investigated exhaustively through theoretical analysis and experiments, and then four modification methods are proposed in this thesis. Experiments result indicate that all these proposed methods can give a better recognition accuracy.3. In order to make use of on-line recognition method and off-line recognition method at the same time, three fusion methods based on on-line recognition and off-line recognition are proposed. With their good complementarity, the fusion of these two approaches significantly improve the recognition performance of cursilve on-line handwritten Chinese characters.