Attention Embedding ResNet for Pest Classification

Abstract

Agriculture drives the development of a country’s economic system. Nowadays, due to population growth and continuous growth in demand for food, agriculture and food industry have become indispensable activities. However, pest has always been considered as a serious challenge affecting crop production. The main hazard of pest is the reduction of crop yields, which reduces food product and even causes famine in some areas. Therefore, the detection and classification of pest and their prevention play a vital role in agriculture. With the advancement of computer technology, accurate and rapid identification of pest can help avoid economic losses caused by pest. This article will carry out the task of pest classification based on the basic deep learning model. The main contributions are as follows: Based on the ResNet-50, different attention modules are introduced for different purposes designs, namely Efficient Channel Attention and Coordinate Attention. The former is an improved SE module, which can more effectively integrate the information between image channels and enhance the network’s learning ability; the latter embeds location information into channel attention. It can obtain information in a larger area without introducing large overheads. Experimental results show that embedding the ECA attention module and CCO module improve network prediction accuracy by about 2%. And these two attention modules do not cause a substantial increase in the amount of calculation, so embedding the attention module is very useful for improving the network’s learning ability.