Off-line handwritten Chinese character recognition is a typical large number of classes pattern recognition problem. It is considered to be very difficult and regarded as one of the ultimate goals of character recognition research due to the large vocabulary, complex structure, lots of similar characters, and infinite variations of shapes resulting from the writing style. In this dissertation, enlightened by the idea of metasynthesis from qualitative to quantitative approach, a Network Integration method based on Supervised Learning (NISL) is proposed to deal with the character recognition problems. The research work in this dissertation can be described as follows: 1. As a first step, integration methods for free handwritten numeral recognition are studied, four different classifiers used artificial neural networks are proposed. To deal with the reject threshold problem, a group of formulas are also presented. Then the results of the individual classifiers are integrated by NISL. The experimental results reveal that the system performance is greatly improved by the proposed method. 2. In order to deal with the problems of handwritten Chinese character recognition, NISL for numeral recognition is improved. Another version of NISL suitable for large vocabulary classification problems is proposed. 3. Five individual classifiers for handwritten Chinese Character recognition are proposed and their results are integrated by NISL. The experimental results are very exciting. After integration, the recognition rate for l0 sets of samples (3,755 classes, 10 samples per class) is as high as about 90%. It is higher than result of the best individual, classifier by 7.15% and fully demonstrates the effectiveness of the proposed method. 4. An integration handwritten Chinese character recognition system is implemented under Windows environment. The recognition speed is 2.4 character per second. 5. A neural network accelerating card is designed and implemented and thus provides hardware support to the system. In brief, the work in this dissertation is creative. The idea to view integration as classification provides a new way for integration research. The presentation and implementation of NISL for handwritten Chinese character recognition not only improve the Chinese character recognition research, but also provide a practical way for the application of artificial neural networks to large vocabulary classification problems. Finally, it should be pointed out that the work in this dissertation can be directly applied to the other areas of pattern recognition.
修改评论