Efficient field-scale detection system of intensive rice lesions on edge devices

• An efficient Field-Scale intensive rice lesions detection model (EFR-YOLO) was proposed. • The LFPN structure and DES-Head modules were shown to exhibit exceptional accuracy in detecting small and densely distributed rice diseases lesions. • The TRNet module, which replaced the original C2f module, significantly reduced the model’s parameter size while improving detection speed • EFR-YOLO achieves 41.9 FPS on edge computing platforms and is deployed on an inspection robot for real-world applications. Accurate and rapid detection of dense rice lesions under complex field conditions is critical for effective early disease management. Nevertheless, achieving this goal remains significantly challenging due to various factors including fluctuating lighting conditions, the diminutive and dense nature of initial lesions, and limited computational resources. To address this, we developed a rice disease dataset (BBD-Rice) comprising 7840 images and 47,313 annotated lesion instances for rice blast and brown spot diseases. An efficient detection algorithm named EFR-YOLO was proposed based on YOLOv8s. It introduces a lightweight feature pyramid network, referred to as LFPN, to enhance morphological diversity detection in lesions. To bolster the identification of minuscule and clustered pathological spots, the EFR-YOLO architecture integrates a streamlined TRNet backbone alongside a specialized DES-head specifically engineered for multi-scale feature capture. By transitioning to a WIoU-based loss mechanism, our training protocol exhibits accelerated convergence while significantly refining bounding box localization. Experimental benchmarks indicate that EFR-YOLO achieves a competitive mAP@0.5 of 91.6%, despite a substantial 49.11% reduction in the global parameter count compared to standard YOLOv8s. This efficiency translates to a 10.3% boost in processing velocity, a fact corroborated by Grad-CAM heatmaps showing superior focus on high-density lesion clusters. To assess its applicability in real deployment scenarios, the model was integrated into a mobile inspection system and tested on a Jetson Orin Nano, achieving a processing speed of 41.9 FPS. The results indicate that EFR-YOLO is capable of performing stable and efficient rice disease detection in real field environments, supporting autonomous inspection under edge computing constraints.