Domain-specific image classification based on improved residual networks

Abstract

Aiming at the problems of high target similarity and strong camouflage in the domain-specific, and the traditional image classification technology is difficult to achieve accurate classification, it is proposed to use the deep neural network Resnet50 as the feature extraction network, and combine the attention mechanism and improve it, which can improve the ability of learning effective features; and use depthwise separable convolution to replace standard convolution, which can reduce computational parameters and save computational space. It is verified by experiments that the accuracy of the improved algorithm in this paper is 0.71% higher than that of the Resnet50 prototype, and 0.39 % higher than that of the Resnet50 +SE algorithm model in the image classification of the domain-specific.