Pest-YOLO: A YOLOv5-Based Lightweight Crop Pest Detection Algorithm

Abstract

Traditional crop pest detection methods face the challenge of numerous parameters and computations, making it difficult to deploy on embedded devices with limited resources. Consequently, a lightweight network is an effective solution to this issue. Based on you only look once (YOLO)v5, this paper aims to design and validate a lightweight and effective pest detector called pest-YOLO. First, a random background augmentation method is proposed to reduce the prediction error rate. Furthermore, a MobileNetV3-light backbone replaces the YOLOv5n backbone to reduce parameters and computations. Finally, the Convolutional Block Attention Module (CBAM) is integrated into the new network to compensate for the reduction in accuracy. Compared to the YOLOv5n model, the pest-YOLO model’s Parameters and Giga Floating Point Operations (GFLOPs) decrease by about 33% and 52.5% significantly, and the Frames per Second (FPS) increase by approximately 11.1%. In contrast, the Mean Average Precision (mAP50) slightly declines by 2.4%, from 92.7% to 90.3%.