Abstract: In view of the difficulty in accurately identifying tomato early blight spot in the complex greenhouse environment, a ResUnet model of integrated channel attention mechanism was proposed. The data set of tomato early blight under complex greenhouse environment was constructed. Lesions were segmented by ResUnet model integrating channel attention mechanism, in which ResUnet network could learn the importance of different depth features, and embedding channel attention mechanism made the improved model pay more attention to the location features of lesions. The accuracy of this model was 97%, 1.99% and 2.97% higher than that of Unet and Resnet101 models, respectively. The parameters and weights of backbone network layer obtained from tomato early blight spot segmentation model were transferred to the single background spot segmentation model of pepper scab, apple gray spot, grape black rot, etc., and improved and the parameters were fine‑tuned to achieve accurate spot segmentation. On the basis of the research algorithm, the intelligent diagnosis system is designed, which can quickly and accurately diagnose the studied crop diseases and provide basis for timely prevention and control.

Key words:  leaf disease, tomato early blight disease, ResUnet, attention mechanism, transfer learning, intelligent diagnose

摘要： 针对温室环境复杂背景下番茄早疫病病斑难以准确识别的问题，提出一种融合通道注意力机制的ResUnet模型。构建温室环境复杂背景下的番茄早疫病数据集；通过融合通道注意力机制的ResUnet模型进行病斑分割，其中ResUnet网络能够学习不同深度特征的重要性，嵌入通道注意力机制使改进的模型更加关注病斑的位置特征。该模型对番茄叶部早疫病病斑分割的准确率为97%，比Unet和Resnet101模型分别提高1.99%和2.97%。将番茄早疫病病斑分割模型得到的骨干网络层参数和权重迁移到单一背景的辣椒结痂、苹果灰斑病、葡萄黑腐病等病斑分割模型，进行改进及参数的微调处理，均能实现病斑的准确分割。在研究算法基础上，设计智能诊断系统，可对作物病害进行快速准确诊断，为及时防控提供依据。

关键词: 作物病斑, 番茄早疫病, ResUnet, 注意力机制, 迁移学习, 智能诊断