Research On Handwritten Numeral Recognition Algorithm And Application Based On Convolutional Neural Network

The handwritten numeral recognition belongs to the basic image classification problems,which has high practical application value and has broad application prospects in the fields of cloud computing?finance?postal and so on.Due to the unrestricted problems in the handwritten numeral writing,it is relatively difficult to achieve fast and effective recognition.With the advent of deep learning related algorithms and the rapid development of computer hardware technology,image classification methods based on convolutional neural networks have gradually become a research hotspot.Because the convolutional neural network has strong function representation ability and network generalization ability,the recognition accuracy can often exceed the traditional image classification method.Therefore,it is of great singnificance to study the application of convolutional neural network to realize the recognition of handwritten numerals.In this paper,the existing problems in the handwritten numeralal recognition method of convolutional neural networks are studied and analyzed.The specific contents include:1?On the issues about the slow convergence speed and low identification rate in handwritten numeral recognition,based on convolutional neural networks in which the convolution kernels are initialized randomly,an improved algorithm in which the convolution kernels are initialized by principal component analysis is proposed.All the experimental results on the MNIST database demonstrate that this improved algorithm has faster convergence speed,higher recognition rate under the iterations limited condition,and it exhibits a batter performance than the convolution kernels initialized randomly method.2?In order to solve the problems of low efficiency in the back propagation from SPP layer to convolution layer during the training process of spatial pyramid pooling network,a back propagation method based on maximum position mapping is proposed.During the back propagation from the spatial pyramid pooling layer to the convolution layer,the fixed gradient vectors of the spatial pyramid pooling network are mapped to the gradient image of the original feature map.The back propagation is realized finally experimental results demonstrate that the mean square error curves converge when the MNIST database is used in single-scale and multi-scale training process,which updated all network layer parameters and achieved parameter sharing between different scale images for each training,in the case of single-scale training,the network has higher recognition rate and batter convergence compared with traditional convolutional neural networks.Under multi-scale conditions,recognition rate has a little change.3?Taking the test paper score as the research object,based on Matlab machine vision toolbox and Caffe convolutional neural network framework,a test paper score recognition application system based on convolutional neural network is designed,and the weight of the system is combined with the actual situation.Fine tune.The experimental results show that the system can realize the input of images of different scales,and it has a good recognition effect for non-adhesive handwritten numerals,and verifies the feasibility of the system.