Abstract: Early diagnosis and treatment of tomato diseases can help to improve the yield of tomatoes. The combination of artificial intelligence and agricultural production can achieve realtime nondestructive detection of tomato diseases. In this study, a research method for tomato leaf disease classification and recognition based on deep learning is proposed. Five kinds of common diseases of tomato leaves are selected for experimentation. Improvements are made to the MobileNetV3 model, and the effects of different learning methods, activation functions, and optimization algorithms on the accuracy of the model are analyzed. The model is compared with MobileNetV3, VGG16, ResNet50, and InceptionV3, and the robustness of the model is evaluated by tenfold crossvalidation. The research shows that the model has good classification performance, achieving an average recognition accuracy of 97.29% for common tomato leaf disease images. The model is superior to other models in terms of model size, running time, and classification accuracy, providing a reference for the recognition of common tomato leaf diseases.

Key words: tomato diseases, multilayer perceptron, dilated convolution, focal loss, identification and classification

摘要： 对番茄病害进行及早的诊断与治疗有助于提升番茄的产量，将人工智能与农业生产相结合可以对番茄病害进行快速地无损伤检测。基于此提出一种基于深度学习的番茄叶片病害分类识别研究方法，选取番茄叶片的5类常见病害进行试验，以MobileNetV3为基础模型进行改进，分析不同学习方式、激活函数及优化算法对该模型准确性的影响。并将该模型与MobileNetV3、VGG16、ResNet50和InceptionV3作对比，同时采用十折交叉验证对模型的鲁棒性进行评估。研究表明，该模型分类性能良好，对常见的番茄叶片病害图像的平均识别准确率可达97.29%，无论模型大小、运行时间还是分类精度上都优于其他几个模型，为番茄叶片常见病害识别提供一定的可参考性。

关键词: 番茄病害, 多层感知机, 空洞卷积, 损失函数, 识别分类