Algorithm of image classification based on Atrous-CNN

Abstract

An image classification algorithm based on atrous convolution neural network (Atrous-CNN) is proposed in this paper, in which atrous convolution is used to replace the convolution layer in the traditional convolutional neural network. When the receptive field and other network parameters are the same, the training accuracy and testing accuracy of the two algorithms are tested. Mnist handwritten digital recognition data set was adopted as the training set and test set of the network model. The three network models were trained 500,1000,5000, 10000 times respectively, and their accuracy and training duration were recorded. Through experiments, compared with the traditional convolutional neural network, the Atrous-CNN has less time consuming and better performance than the traditional convolutional neural network with the same accuracy.