Abstract
Abstract To address the challenges of high computational cost, strong background interference, and scale variation in tea pest and disease detection under complex field conditions, this paper proposes MSDR-Net, a lightweight YOLOv8n-based detection method. MSDR-Net is designed as a modified YOLOv8n variant rather than a detector developed from scratch. The model integrates depthwise separable convolution, parallel multi-scale feature extraction, and a SimpleRouter adaptive gated feature-fusion module to improve pest and lesion representation while controlling model complexity. Experimental results on the constructed tea pest and disease dataset show that MSDR-Net achieves an mAP@0.5 of 0.980. Compared with YOLOv8n, the parameter count, GFLOPs, and model size are reduced by 37.1%, 13.6%, and 34.6%, respectively. In addition, a desktop monitoring prototype was implemented to verify the basic functional feasibility of the proposed method. These results suggest that MSDR-Net provides a favorable balance between detection accuracy and model efficiency under the current experimental setting, while broader cross-domain validation and edge-device deployment tests remain necessary in future work.