ResearchBib Share Your Research, Maximize Your Social Impacts
Sign for Notice Everyday Sign up >> Login

Multi-layer high-precision image classification technology embedded in SE modules

Journal: International Journal of Advances in Computer Science and Technology (IJACST) (Vol.11, No. 8)

Publication Date:

Authors : ;

Page : 29-35

Keywords : ;

Source : Downloadexternal Find it from : Google Scholarexternal

Abstract

Due to the problems of model overfitting and gradient changes to reduce model performance in deep networks, the operation of improving the accuracy of image classification models by superimposing the number of layers of the network cannot be applied to all models. The Squeeze-and-Excitation (SE) module is a plug-and-play attention module in the field of computer vision that focuses on channel relationships. Experiments show that embedding SE modules in ResNet models of different scales brings much higher test accuracy improvement than increasing the depth of the original model; SE modules are extremely generalizable, and their embedding is universal to greatly improve the accuracy of different original models. Experimental results on the CIFAR-10 and Dogs-vs-Cats datasets show that the larger the amount of data, the more it can avoid the overfitting phenomenon of the model. A comparison experiment with the GoogLeNet model showed SENet being superior. According to the published research data, the application of SE modules accounts for 57.59% of the top 30 disciplines such as medical health, automation technology, telecommunications technology, electric power, light industry, automobiles, and transportation

Last modified: 2022-08-21 14:07:38