Grape leaves grown in natural environments often exhibit boundary blur and pose challenges for feature extraction during disease identification due to factors such as dense adhesion, irregular morphology, and varying light conditions, significantly reducing the recognition accuracy of existing algorithms. The essence lies in the mismatch between the rigid sampling grid of universe models and the irregular boundary morphology of leaves; as well as the difficulty for general attention mechanisms to spotlight the discriminative area of tiny lesions. To address such root problem, this study proposes a lightweight detect model YOLO-Grape improved based on YOLOv11, whose core innovation is the proposal of an optimization scheme of "boundary priority \(\rightarrow \) lesion enhancement \(\rightarrow \) feature refinement". Through a systematic collaborative scheme, it fundamentally enhances the model’s competency to discriminate boundaries and lesions. This scheme is instantiated into three collaborative modules: first, introduce pre-screening network (DeShuffleNet), which conforms to the leaf outline through deformable convolution to effectively capture the irregular edge information of leaves; second, design EViM_MD module, which spotlights multi-scale lesion areas through multi-dimensional syndication attention; finally, the SDI module refines cross-tier features to guide detailed location with semantic information, achieving optimal fusion of boundary and lesion information while suppressing background noise interference. These designs enable YOLO-Grape to maintain a lightweight size (14.38MB) while demonstrating stronger robustness and higher identification accuracy in complex scenarios such as boundary blur and dense adhesion. In comprehensive evaluations, YOLO-Grape achieves a mAP50–95 of 95.2%, representing a 4.6% improvement over the baseline YOLOv11. It significantly outperforms existing mainstream lightweight models and possesses excellent real-time field detection capability and deployment feasibility, providing a reliable solution for intelligent agricultural disease identification.
