Tea pests and diseases detection method based on multi scale dynamic routing network

Abstract To address the challenges of high computational cost and strong background interference in automatic recognition of tea pests and diseases in complex tea garden scenarios, this paper proposes a detection method named MSDR-Net (Multi-Scale Dynamic Routing Network). First, to significantly reduce the model parameter count and computational complexity, a lightweight backbone is constructed using depthwise separable convolutions. Second, a parallel multi-scale feature extraction structure is designed to capture both the contours and details of small pests and large disease spots through differentiated branches. Finally, to suppress background interference and improve feature fusion efficiency, a SimpleRouter dynamic routing mechanism is introduced to enable adaptive filtering and weighted fusion of key features. Experimental results on a self-built real-world tea pest and disease dataset show that the model achieves a mean average precision ([email protected]) of 98.0\\%, which is 0.3 percentage points higher than the baseline model YOLOv8n. Meanwhile, the parameter count, computational complexity, and model size are reduced to 1.89M, 7.0 GFLOPs, and 3.91 MB, representing reductions of 37.1\\%, 13.6\\%, and 34.6\\%, respectively, compared to the baseline model. Furthermore, an intelligent monitoring system developed based on this model verifies its effectiveness and usability in practical applications.