A Study On The Methods Of Handwritten Numeral Recongnition Based On Deep Learning

With the continuous development of technology,people are increasingly relying on computers in their life.Handwritten numeral recongnition is becoming more and more important.High accuracy and high efficiency are issues that users urgently need to improve.With the continuous development of deep learning,the use of convolutional neural networks has made a breakthrough in a series of computer fields.However,due to its own characteristics and complexity,hand-written digital recognition still has a lot of room for development in both technology and application fields.In recent years,with the rapid change of computer hardware and software technology,the time for deep learning network training has been greatly reduced,and deep learning has been deeply concerned by scholars of the current year.As a direction of deep learning,convolutional neural networks have also achieved fairly rapid development.This paper first studies the methods and theories of traditional convolutional neural networks,including local perception,weight sharing,downsampling,convolution,batch normalization,activation function,pooling,Softmax regression and so on.Based on the in-depth study of the convolutional neural network theory,the traditional convolutional neural network is used to deal with the disadvantages exposed in the handwriting recognition process:(1)The gradient disappears and the gradient drops.The sigmoid activation function is usually used in neural networks to map the negative infinity to the positive infinity to the value range from 0 to 1,but when the network level is very deep,the deviation of the output layer will become smaller and smaller until it becomes 0,The gradient explosion problem is similar to the gradient disappearance.As the gradient propagates,the final parameter gradient will increase exponentially and a gradient explosion will occur.These two problems will lead to an increase in training model error and a decrease in accuracy.(2)As the network level increases,the amount of calculations increases dramatically.As the network level increases,the convolutional convolution operation makes the computational volume of the entire network approximately grow exponentially.This article I proposed two improved solutions:(1)Based on the residual network and recursive neural network handwritten digit recognition,the method proposed in this paper effectively avoids the gradient disappearance problem by adding residual network,and increases the number of trainable parameters,making the network more robust.By adjusting parameters and depth,the experimental results of this paper have significantly improved the recognition rate compared with the traditional convolutional neural network method.(2)Handwriting digital recognition based on Inception network model.In this method,Inception network model is introduced in this paper.By decomposing large volume integrals into small convolution sums,the amount of network computation is greatly reduced,and the entire network is improved greatly.The experimental results show that the model has better robustness.