Study On Technology Of Handwritten Digits Recognition

Handwritten digit recognition is widely used, and requires a lower error rate. The traditional method of handwritten digit recognition is extracting high dimensional feature set of handwritten digits and using feature selection method to filter out a lower dimensional collection, and then training the neural network for digits's classifiers. The approaches often fails to meet requirements of high-precision of recognition. Taking high performance hand-written digital systems as a starting point, the paper abandoned the method.The neural network designed in the paper put all pixels grayscale of character image as inputs,which ensures integrity of characteristics information of digital character. In the neural network design, convolution network is applying Against characteristics of the digital images. Convolutional network is specifically designed for image recognition whose weight sharing can reduce the training parameters of the network and make the neural network structure simple and more adaptable.This convolution network Includes 5 layers:an input layer, three middle layer and an output layer, where the middle layer consists of two convolution layers.The neuron in each convolution layer compose a number of feather maps,each of which is result of convolution between previous layer and a specific convolution kernel which is a sharing weights array of the neurons in the feather map.Learning algorithm is still BP algorithm.The paper introduced second method to the traditional BP algorithm, which weights update is no longer a constant learning rate, but adaptive variables related to the Hessian matrix of the output cost function. Second-order method greatly accelerated the rate of decline of the cost function. In addition, the using of a small number of random input sample volume simplifies the error convergence condition, and is easy to implement, and random inputs avoid local minimum, so as to improve training effect. The paper select hyperbolic tangent function as activation function, whose (-1,+1) value range is more suitable for this network output. For the error of the network error is less than 1/10 of the last training phase, the back propagation process was skipped because of small weight correction, which improved the efficiency of training.Before recognition the picture should be done some treatment, such as binary segmentation, removement the border, character segmentation, morphological processing, filtering noise and normalization. Then identify splited by the neural network ordinally. Experimental results show that the neural network classifier used in the paper had high recognition rate for the MNIST library handwriting samples, and the system also made a good recognition performence to the number of characters in frame.