An Improved CNN-ResNet Deep Learning Neural Network And Its Applications

Residual neural network(ResNet)is a deep neural network that is currently used more often.Compared with the widely used convolutional neural networks,residual networks solve the problems of network degradation and gradient disappearance that occur with the increase of network depth,therefore widely used.Improvement is made on the basis of residual neural network,and the purpose is to improve the accuracy by deepening the network structure.This paper makes structural improvements based on the traditional residual network: the basic structure of a convolutional neural network is added to the residual neural network,and a batch of batches is added while ensuring the o riginal eigenvalues of the input data of the residual network.The new eigenval ues generated by the convolutional neural network are continuously trained and self-learned by the parameters in the network structure,and finally the purpos e of improving the accuracy of the network is achieved.In the added convolut ional neural network structure,regarding the size of the convolution kernel,the appropriate size and structure of the convolution kernel are selected: in the co nvolution layer,the multi-layer convolution form of the small convolution kern el is compared with the traditional large convolution kernel.The multi-layer of the small convolution kernel greatly reduces the number of parameters in the network structure,effectively reducing the operation time and improving the op eration efficiency.In addition,different from the traditional residual network,af ter applying the batch standardized structure to each part of the network struct ure,this article puts the batch standardized structure into the residual module o f the residual network,which reduces a certain number of calculations and red uces the calculation time of the data.This article improved the activation function based on the traditional ReLu activation function: Since the ReLu function has a gradient of 0 on the negati ve semi-axis,neuron necrosis is prone to occur after updating the parameters when data with a large gradient is entered,and the gradient will not change a nymore.The input data will be processed again.The improved activation functi on is infinitely close to 0 on the negative semi-axis,which solves the problem of neuron necrosis.Compared with other commonly used activation functions,the new activation function has a gradient value at positive infinity to solve t he phenomenon of gradient disappearance and gradient explosion which are eas y to appear.The training resulted on multiple data sets show that the method has been greatly improved the detection accuracy and precision,the generalization abilit y of traditional residual networks,and it also reduce the number of parameters in the network,and effectively reduced the training time.