Lightweight detection system for tea plant diseases and pests based on improved YOLOv8s

Abstract

To address the practical issues of significant target scale variations in tea pest and disease images, difficulty in distinguishing diseased spots from backgrounds, and the need for detection models to adapt to resource-constrained edge devices, this paper proposes the YOLOv8s-CFW lightweight detection model. In terms of lightweight design, leveraging the prior knowledge that tea pest and disease image edges and textures primarily rely on pointwise channel combinations and exhibit low sensitivity to spatial positions, the model replaces high-resolution convolutions in shallow layers of the backbone network with depthwise separable convolutions, compressing computational load to one-ninth of standard convolutions. To compensate for the insufficient channel interaction caused by depthwise separable convolutions, a CBAM module is serially embedded at the end of the feature extraction network. Through two-phase compensation—channel reweighting and spatial position calibration—the model enhances disease feature identification. Differing from global lightweight strategies like GhostNet and MobileNet, this model forms a collaborative architecture combining local lightweight design with terminal refinement. Experimental results show that YOLOv8s-CFW achieves an mAP of 98.2% on the tea pest and disease dataset, a 3.7% improvement over the original model, with a 45.2% reduction in parameters and model size compressed to 12.4MB. CPU single-image detection takes 180ms, achieving 19.2 FPS on Jetson Nano edge devices. The Flask and Neo4j-based detection system supports 20 concurrent users with an average response time below 220ms, meeting real-time detection requirements for tea garden edge devices.