Off-line Handwritten Digits Recognition Based On BP Neural Network

BP neural network is widely used in nonlinear modeling pattern recognition etc. In this paper we give a kind of handwritten digits recognition method based on BP neural network. An artificial neural networts is an information processing system that has certain performance characteristics in common with biological neural networks.Neural networks possesses many of advantages: simple structure, easy implementation in hardware, the basic parallel computational architecture, and the model has features of study, memory, self adapting and diversity.This paper uses the MATLAB imread.m function to read the user-defined handwritten digital picture, and uses the normalization method to preprocess images.Then we extract features pixel by pixel. After pre-processing and feature extraction of the self collective hand-written digital samples, we gain the input vectors.By distilling the characteristic vector from samples, selecting enough stored samples to train the BP Neural Network, and putting the samples standing for the hand-written digits into trained Neural Network, the distinguished digital scripts are obtained according to the output of the neural network.Some modified BP algorithms are discussed in order to eliminate the disadvantages of standard back-propagation algorithm, such as alowly convergence speed and great computational complexity. Useing Matlab neural network toolbox of the lates vertion, the training speed of several typical BP training algorithms is compared. The simulation experiment on digital samples shows that the BP neural network model used for the problem of the hand-written digits recognition is capable of recognition to certain extent, and the algorithm is operated easily. If the relative variables can be choused properly, better recognition effect of the system can be getting. In this paper, the correct rate is about 65.33%.