Abstract: Rapid localization and accurate identification of tomato leaf diseases can help in the rational use of pesticides, thereby ensuring the quality and yield of tomatoes. In order to address the problem of poor performance of existing detection methods for tomato leaf disease, a selfsupervised detection method  for small sample tomato leaf disease was proposed. Firstly, a set of shared weight backbone networks were used to extract semantic features of tomato leaves in the visual space. Then, the visual semantic features were input into a deep autoencoder network, and the feature encoding network was optimized by calculating the contrast loss between the encoded and original features. Finally, the encoded features were used to guide the localization and identification of unknown tomato leaf diseases. In addition, a double loss optimization strategy was designed to obtain more robust guiding feature sets. Through testing experiments on a selfbuilt tomato disease leaf dataset and an opensource dataset, the proposed model achieved recognition accuracies of 0.946 2 and 0.963 9 on the selfbuilt and opensource datasets, respectively, which were superior to current stateoftheart object detection methods.

Key words: tomato leaf disease detection, selfsupervised learning, autoencoder network, dual loss, semantic feature

摘要： 番茄叶片病害的快速定位与精准识别有助于合理使用杀虫剂，进而保障番茄的质量与产量。针对现有番茄叶片病害检测方法检测性能不佳的问题，提出一种自监督下的小样本番茄叶片病害检测方法。首先，利用一组共享权重的主干网络提取番茄叶片在视觉空间中的语义特征；然后，将视觉语义特征作为深度自编码网络的输入，通过计算编码压缩后的特征与原始特征间的对比损失优化特征编码网络；最后，利用编码压缩后的特征指导番茄叶片的未知病害定位与识别。此外，为获得更鲁棒的指导特征集，设计一种双损失的优化策略。通过在自建的番茄病害叶片数据集和开源数据集上进行测试试验，所提出模型分别在自建和开源数据集上实现0.946 2和0.963 9的识别精准率，优于当前经典的目标检测方法。

关键词: 番茄叶片病害检测, 自监督学习, 自编码网络, 双损失, 语义特征