The Research On Neural Network For Off-line Handwritten Chinese Characters Recognition

Chinese characters, the symbol of the Chinese nation and the natives character that are thousands of years old, have many features such as a wide range of sorts, complex segmentation and a lot of similar characters. Thus, that computers recognize handwritten Chinese characters represents the highest level of pattern recognition. Chinese characters recognition plays a great part in machine translation, office automation, press and publication, and so on. Neural network is a kind of information processing models, used for simulating biologic neural systems. Because of the high ability in the adaptive drill and parallel processing,neural network has the huge potential in pattern recognition problems of large scale like off-line recognition for handwritten Chinese characters. Thus, researching off-line handwritten Chinese characters recognition has the theory and practical value. There is no more complicated pattern recognition than off-line handwritten Chinese characters. The characters recognition consists of several steps, including preprocessing, feature extraction, classification and postprocessing. Preprocessing that is made up of a whole array of operations, for example, image denoising, thinning, binarization, and normalization, deals with handwritten Chinese characters images produced by the scanisters. Feature extraction that serves as the important basis of recognition extracts features of handwritten Chinese characters that are very stable and can describe the shape of characters. And the fuzzy directional line element feature extraction method, one of extracting feature approaches, is used widely. Classification is very crucial for off-line handwritten Chinese characters. The distance classifier, the neural work classifier and the support vector classifier can be employed for the classification frequently. The task of postprocessing is that Chinese characters misidentified by the system will be rectified according to the context to improve the recognition rate. Sphere neighborhood model, a geometrical representation of MP neural model, is the model of neural network by which the training problem of neural network may be transformed into the geometrical covering problem of a point set based on n +1?dimensional hypersphere. In this paper a modification approach for neural network is introduced according to the analysis of the neural network separate boundary faces and the original algorithm. Moreover, it is very limited for single neural network to do recognition and classification. Thus the combination of several neural networks is considered and neural work ensemble is applied. Neural network ensemble can markedly improve the generalization ability of the system through training many individual networks and combining their results. The method of neural network ensemble based on the modification approach is also presented in this paper. The better experimental results show that the improved algorithm and the method of ensemble are competent in off-line handwritten Chinese characters.