Handwritten Digit Recognition Based On BP Neural Network And SVM

Handwritten Digit Recognition is a main branch of Optical Character Recognition. It is one of the most classic subjects in pattern recognition research, and one of the most successful applications. It has been applied in many fields. Digit Recognition has few patterns, which helps to analyze and verify some new theories and methods. Further more, the methods in handwritten digit recognition are easy to generalize in some relative problems. Therefore, it is significant to make deep researches on it.In this paper, a handwritten digit recognition system based on BP neural network and Support Vector Machine (SVM) is built up, which is made up of two parts, including learning and recognition. During the realizing of the system, the following was done in this paper.1. Make some improvements on BP training algorithm to quicken the network constringency speed and avoid fake saturation phenomenon.2. Probability statistics of misclassification patterns is given in the learning phase of BP Neural Networks. The system will set some confidence parameters according to this Probability statistics. In recognition phase, the system will decide if the output of BP neural networks need to pass the SVM according to this confidence parameter.3. Apply SMO algorithm and improve DirectSVM algorithm in the learning phase of SVM. For the SVM spends long times in training the large sample set, we applied DirectSVM algorithm and improved it. And we get an approximate support vector set in much shorter time. Then we can use this vector set to train the SVM by SMO algorithm, avoiding depress of the whole performance of the system. In this way, we get a faster SVM classifier.4. A handwritten digit recognition system of two levels is built up in this paper. The BP Neural Networks serves as the first level classifier, and the SVM serves as the second level classifier. If the recognition output of BP Neural Networks accords with some certain confidence, the system will output the recognition result. If the recognition output of BP Neural Networks doesn't accord with the confidence, the system will select the two maximal patterns to be classified in the SVM classifier. The result of experiment shows that the system has improved the precision of handwritten numeral recognition at a certain extent, and it is worth to be studied later.