Abstract: Plant diseases can have a disastrous impact on food safety. They can directly lead to a significant decrease in the quality and yield of crops. Therefore, early identification of plant diseases is very important. Traditional crop disease diagnosis requires very high professional knowledge, not only timeconsuming and laborious, but also inefficient. In this paper, we use the method of deep learning to develop the Faster R-CNN network model based on TensorFlow, and taking potato leaves as research samples. We expanded the image of potato leaves with early blight, late blight and health by local enhancement method. The COCO initial weight was applied for transfer learning, and the influence of data categories on model detection results was explored. The results showed that with the increase of training data categories, the performance of the model would be slightly reduced. At the same time, we also trained YOLOv3 and YOLOv4 network to compare with this model. The test results showed that the performance of our proposed Fater R-CNN model was better than other network model, and the best accuracy of the model reached 99.5% after testing. This study can provide technical support for potato disease detection.

Key words: plant disease, deep learning, Faster R-CNN, transfer learning, disease detection

摘要： 植物病害对食品安全具有灾难性的影响，它可以直接导致农作物的质量和产量显著下降，因此对植物病害的早期鉴定非常重要。传统的农作物病害诊断需要非常高的专业知识，不仅费时费力，还效率低下。针对这些问题，利用深度学习的方法，以马铃薯叶片为研究样本，基于TensorFlow开发Faster R-CNN网络模型。采用本地增强的方式对带有早疫病、晚疫病和健康的马铃薯叶片进行图像扩充，应用COCO初始权重进行迁移学习，探究了数据类别对模型检测结果的影响。结果表明，随着训练数据类别的增多模型性能会有略微的降低。同时还训练YOLOv3，YOLOv4网络与该模型进行对比，测试结果表明，所提出的Fater R-CNN模型优于其他网络模型。经检测该模型最佳精度达到99.5%，该研究为马铃薯病害检测提供了技术支持。

关键词: 植物病害, 深度学习, Faster R-CNN, 迁移学习, 病害检测