An Spodoptera Frugiperda pest detection algorithm based on improved feature fusion

Abstract

Maize is one of the important foods and feed crops with high agricultural value in the agricultural fields of northern China. Accurate detection and treatment of nightshade spodoptera frugiperda in agricultural fields is essential to increase agricultural production. Compared with the traditional manual detection methods, the convolutional neural network-based spodoptera frugiperda detection algorithm has the better recognition performance. However, the method of using convolutional neural networks for the recognition of spodoptera frugiperda still suffers from low accuracy and slow detection speed. To improve the detection accuracy and speed of the spodoptera frugiperda, the algorithm modified the information fusion based on the YOLOv3 model, and used quadruple upsampling fusion instead of the original double upsampling fusion to remove the redundant information brought by repeated fusion. In the experiment, the dataset consists of 4000 training images and 200 test images for training and testing. The experimental data show that the improved YOLOv3 algorithm improves the accuracy by 11.42% and the detection time by 0.0023s compared with that before the improvement.