Performance evaluation of a DCD-YOLO model in grape disease detection

In response to the severe impact of common diseases such as measles, wilt disease, and rotten holes on grape yield and quality in grape cultivation, as well as the limitations of existing detection models in identifying subtle lesions and locating irregular lesions, this study aims to propose an improved YOLOv11n-based detection model for grape leaf diseases. The goal is to enhance the accuracy and robustness of grape leaf disease detection, thereby providing technical support for the early diagnosis, prevention, and control of grape diseases. An improved model was constructed based on the YOLOv11n algorithm, with key improvements including: embedding the DPPA dual-channel attention module in the backbone network to fuse multi-scale features through dual pathways of local detail perception and global context capture, thereby enhancing the discriminative ability for heterogeneous lesions; replacing the original C3K2 structure with the C2f-DCN module, which introduces deformable convolution to adaptively capture the morphology of irregular lesions and improve localization accuracy. A grape leaf disease dataset (GLDDD) containing 5,000 images was used, divided into training, validation, and test sets in an 8:1:1 ratio. Combined with data augmentation techniques such as Mosaic-9 and Randaugment, the model was trained with 250 training iterations. The improvement effects were verified through ablation experiments and comparative experiments. The following were observed: (1) The performance of the improved model was significantly enhanced, with an mAP of 95.7%, showing excellent performance among similar models. (2) Ablation experiments indicated that the C2f-DCN module had a prominent effect in improving precision and mAP, with minimal impact on parameter count and model size; the DPPA module could improve precision to a certain extent, and their combination effectively optimized model performance. However, although the DFF module improved recall, it significantly increased the parameter count and model size while leading to a decrease in precision. (3) Comparative experiments showed that the improved DCD-YOLO model achieved a 4.9% higher detection precision for measles disease than YOLOv11n and a 5.4% higher precision than YOLOv11n-DPPA. The average detection accuracy for the three diseases was significantly higher than that of other comparative models. (4) The model achieved the optimal performance balance at 250 training epochs, with a precision of 92.3%, a recall of 90.1%, and an mAP of 95.7%, demonstrating the best performance in detection accuracy and coverage capability. The improved YOLOv11n model proposed in this study effectively addresses the issues of difficulty in identifying subtle grape leaf lesions and low localization accuracy of irregular lesions by introducing the DPPA and C2f-DCN modules. It significantly enhances the performance of grape leaf disease detection, meets the practical needs of grape disease detection, and provides a feasible technical solution for the automated detection of crop diseases in the agricultural field.