A Multi-Scale Perception-Enhanced Lightweight Network with Knowledge Distillation for Rice Leaf Disease Detection

Rice is a critical crop for global food security and economic stability. However, various diseases, including rice blast and bacterial leaf blight, pose significant threats to rice cultivation. Existing methods for detecting rice leaf diseases suffer from low efficiency and limited generalization capability. These methods are incapable of capturing variations of disease characteristics across different growth cycles. Therefore, a lightweight detection model named lightweight knowledge distillation YOLO (LWKD-YOLO) is proposed. The convolutional layers in the YOLOv8 network are replaced with the ADown module. This change significantly reduces computational load while improving detection accuracy. A lightweight detection head, termed the lightweight shared re-parameterizable convolutional detection head (LSRP-Head), was designed. It incorporates group normalization RepConv, further reducing computational complexity while enhancing multi-scale perception capabilities. Furthermore, based on the improved ADown module and LSRP-Head, the YOLOv8x model is employed as a teacher model for inter-channel correlation knowledge distillation. This effectively enhances the ability to learn complex rice leaf disease features. The effectiveness of the proposed method was verified through ablation and comparative experiments on the constructed rice leaf disease dataset. Compared with the baseline model, LWKD-YOLO increases mAP@50 by 1.4%, reduces the number of parameters by 1.3M, and lowers FLOPs by 3.1G. As a result, the proposed model enables efficient rice leaf disease detection in complex environments, demonstrating notable economic and practical significance.