Off-line Handwritten Digit Recognition Based On BP Artificial Neural Network

Off-line handwritten digit recognition is the current hot spots of OCR (Optical Character Recognition) technology research, and it is also one of the very difficult tasks of the computer character recognition as the reasons of less information, similar shape, difficult to obtain the information of stroke order, large deformation and so on. It can play a great role in the areas of large-scale data statistic, financial sector, automatic mail sort, automatic input of handwritten manuscripts and so on. The research on off-line handwritten digit recognition is also quite significant for the automatic processing of digit information and the development of intelligent input of the computer. If there is any connection with digit recognition, it is need that the classifier must have a high reliability and rate, especially if it has some relationship of financial. So, one of the key links of the design of the process system for such problems is to make out the handwritten digit recognition method, which has not only a high-reliability, but also high rate. Handwritten digit recognition is a very complicated multi-pattern recognition issue, the researchers have spent many years on the methods of handwritten digit recognition, and they have point out many ways of this problem. However, there is still no way can reach the ideal result.In this paper, the object of study is mainly about more than 60,000 handwritten digits, which are written by teachers and students of colleges, and 10000 of them are used for training, 4000 of them are used for testing. The other parts of them are used for latter. The main job of this paper is about new feature extraction and the advantage of BP artificial neural network. The process of feature extraction and BP artificial neural network is achieved by C++Builder.(1) Feature extraction:As each feature has its own strength and weakness, it is difficult for one feature to reach the ideal result. So this paper decides to use the way of Multi-feature combinations, which means Multi-feature integration on the basis of other researchers' research. This is one trend of handwritten digit recognition. In this paper, a new method of the combination of the features is put forward, in the condition of the advantages and disadvantages of various features and the principle of feature selection. First, it describes the features of digits in the condition of the global and local conditions. It means the directional element feature and the contour feature are composed to be a new feature. Second, it complements the case of broken pen, which can not be deal by both of the directional element feature and the contour feature, so the real line feature is pulled in. At last, the effectiveness of the new method is proved after various features being compared. This new method has little operation about thinning of the samples, so it can not only decrease the works of pretreatment, de-noising and pre-processing operation, etc, but also reduce the error caused by thinning.(2) BP artificial neural network:This paper improved the adaptive factor of gradient to improve the transfer function on the base of the means of momentum, aimed at the training difficult to escape the flat area of error. This method had an application on off-line handwritten recognition. The results showed that this method could improve the convergent speed of it.The recognition rate of training samples is 95.58%, and the recognition rate of testing samples is 84.6%.