Abstract:  In order to accurately detect grape diseases under complex environmental factors in orchards, a grape disease detection model Swin-TDL based on Swin Transformer is proposed. Kmeans++ clustering algorithm is used to calculate the anchor frames of the model input images to improve the detection accuracy. The Swin Transformer network is used as the backbone network of Swin-TDL for more accurate acquisition of target feature information. Feature pyramid networks and path aggregation networks are used to fuse information from feature layers of different depths in the backbone network to enhance the semantic and localization information of detection targets.  The SIoU loss function is used as a boundary regression prediction loss function for improving the speed of training and the accuracy of model inference.  Soft-NMS is used to post‑process the target bounding boxes to improve the detection rate of occluded and overlapped targets. Finally, model training and performance testing were carried out in the field grape disease dataset. The experimental results showed that the average  accuracy of the Swin-TDL model was 92.7% and the average  detection time was 15.3 ms. The comprehensive performance of Swin-TDL was better than the comparative detection algorithm, which could provide a reference for the research of grape plant protection equipment.

Key words: grape disease detection, orchard complex environment, image enhancement, deep learning, Swin Transformer

摘要： 为在果园复杂环境因素下准确检测葡萄病害，提出一种基于Swin Transformer的葡萄病害检测模型Swin-TDL。采用Kmeans++聚类算法计算模型输入图像的锚框以提高检测精度；以Swin Transformer网络作为Swin-TDL的骨干网络更准确地获取目标特征信息；特征金字塔网络和路径聚合网络用于融合骨干网络中不同深度的特征层信息以增强检测目标的语义信息和定位信息；使用SIoU损失函数作为边界回归预测损失函数用于提高训练的速度和模型推理的准确性；使用Soft-NMS对目标边界框后处理以提高遮挡及重叠目标的检出率。在田间葡萄病害数据集中进行模型训练和性能测试，结果表明，Swin-TDL模型的平均精度均值为92.7%，平均检测时间为15.3 ms，综合性能优于对比检测算法，可以为葡萄植保装备研究提供参考。

关键词: 葡萄病害检测, 果园复杂环境, 图像增强, 深度学习, Swin Transformer