An improved DenseNet model to classify the damage caused by cotton aphid

Abstract

Accurate and timely detection and classification of cotton aphid damage are essential for the control of cotton aphids, a major pest in cotton in China and many other countries. However, use of existing convolutional neural networks (CNN) to classify the levels of damage caused by the pest is undesirable because of their low accuracy caused by complex field backgrounds and different lighting conditions. In this study, a lightweight classification network, CA_DenseNet_BC_40, with improved DenseNet was proposed by introducing the network architecture of DenseNet and Coordinate Attention module for classifying the levels of damage caused by cotton aphids in a natural field environment. The results of analyses show that the CA_DenseNet_BC_40 network outperformed the existing networks ResNet50, ShuffleNet, Ghost, MobileNetv3, and DenseNet on the accuracy of classification for cotton aphid damages. The classification accuracy of the proposed network reached as high as 97.3 % and the size of parameters was only 0.18 M that was smaller than those of the lightweight convolutional neural network models such as Mobinenet and GhostNet. The proposed model can be used to automatically detect and classify the levels of damage caused by cotton aphids in natural field conditions with a high accuracy.