Abstract: In order to meet the needs of orchard plant protection equipment for real-time disease detection model, recognition accuracy and lightweight, a grape leaf disease detection model YOLO-SL based on the improved YOLOv3 model is proposed. First, the component modules of the lightweight network ShuffleNetv2 are introduced to optimize the original feature extraction network of YOLOv3 to reduce the network model parameters, then the CBAM attention mechanism is incorporated in the optimized feature extraction network, and a small target feature detection layer is added to the feature pyramid structure of the YOLOv3 network model to improve the detection model recognition accuracy. Finally, comparative experiments with different detection models were conducted on the data-enhanced dataset. The experimental results showed that the YOLO-SL network model could reach an average detection accuracy of 90.4%, and the average detection time was reduced to 32.2 ms, and the weight size was reduced to 18.3% of the original YOLOv3 model, and this method could be used for the detection technology of grape leaf diseases in the actual working environment application.