Enhanced YOLOv8 algorithm for leaf disease detection with lightweight GOCR-ELAN module and loss function: WSIoU

Leaf disease detection holds significant application value in the agricultural domain, as timely and accurate detection of crop leaf disease targets is crucial for improving crop yield and quality. To handle varying crop leaf disease target sizes, occlusion issues, and detection errors in complex environments, the YOLOv8 structure has been enhanced. Firstly, to tackle the issues of target diversity and loss of image features, this paper designs the GOCR-ELAN lightweight module to replace some of the C2f modules in the Backbone, thereby reducing the parameters in the model and enhancing the network's feature extraction capability. Secondly, replacing the CBS convolution in the network with the ADown downsampling module effectively addresses issues such as feature selection and preservation in occluded scenes, further reducing the algorithm's parameter count. Finally, to tackle missed detections and false alarms in complex environments, this paper introduces the WSIoU loss function optimization algorithm to enhance both convergence speed and localization accuracy. Averaged experimental results suggest that, relative to the original YOLOv8 algorithm, parameters have been reduced by 28.7 %, the GFLOPs metric has decreased by 43.2 %, MAP50 has increased from 86 % to 87.7 %, and MAP50-95 has risen from 67 % to 68.9 %, achieving both lightweight model construction and improved detection performance. The trained model is just 4.55 MB, smaller than the lightest YOLOv5 model, and remains highly competitive in detection accuracy compared to larger models.