A Multi-stage Neural Network System For Handwritten Character Recognition

Handwritten character recognition is a very important and active research in pattern recognition. Theoretically, it is not an isolated technique. It concerns with the problem that all the other areas of pattern recognition must confronted; Practically, being a kind of information processing measure, character recognition has a very broad application background and vast need of market. Thus, it is of both theoretical and practical significance. Artificial neural network recognition method is a new method of the research field in recent years, and this method has some merit that traditional technique do not have; good tolerance for error, strong sorting ability, strong parallel handling ability and strong self-learning ability as well as its off-line training and on-line recognizing. All these merits contribute its perfect performance in handling vast data set and handling in timely manner. The characteristics of handwritten character set(especially the numeral set) demand the recognition system take account of both high accuracy and reliability. It is very difficult to give considerations to both of them. To deal with this problem, three complement classifiers are given and their recognition results are combined by the proposed integration method. In this system, different network is combined to different feature extraction. Multiple neural networks collaborate perfectly to make full use of features so that give us satisfying results. This paper mainly involves preconditioning and feature extraction of handwritten character picture, the main theory of neural network, and training of recognition system of neural network.