A task-adapted structurally reparameterized diverse-branch YOLO network for mulberry leaf disease detection

Mulberry leaf diseases reduce leaf yield and quality, with direct consequences for sericulture. Detecting symptoms in field imagery is challenging because lesions are often small and visually similar across disease types. In this study, a task-adapted YOLOv10-based detector is proposed that integrates a modified diverse-branch feature extraction module (C2fDBN) and maintains efficient inference through structural reparameterization. The detection target is defined as the whole-leaf category rather than individual lesion spots to align with practical classification requirements. Experiments were conducted on an in-house field-collected mulberry leaf dataset and an independent public dataset, both annotated for healthy leaves and disease classes. On both datasets, the proposed detector achieves competitive performance and frequently surpasses or attains performance comparable to strong object detection baselines (YOLOv8, YOLOv9, YOLOv10, YOLOv11, YOLOv26, RT-DETR, Faster R-CNN) in terms of mAP, while maintaining a comparable inference-efficiency profile. These results indicate that the proposed model can support leaf-level field scouting and agricultural decision-making within the evaluated datasets.