Abstract:  Tomato is an important vegetable crop, but its yield and quality are often affected by various leaf diseases. To address this problem, computer vision techniques have been widely applied in the automatic disease identification. Existing methods can be divided into two categories: handcrafted feature extraction methods and deep learningbased approaches. While methods using handcrafted features are simple and efficient, they lack robustness. In contrast, deep learning methods can effectively improve recognition accuracy but require largescale annotated data and high computational complexity. To mitigate these issues, this paper proposes a tomato leaf disease recognition framework based on enhanced graph structure （TDR—EGS）. TDR—EGS incorporates topological relationships between samples to enable alternating training of graph learning and singlesample learning, thereby effectively enhancing classification performance without increasing inference complexity. Specifically, it first extracts singlesample features through a convolutional neural network and then uses these features to construct a knearest neighbor graph to uncover structural information between samples. This approach allows graph learning and singlesample learning to collaborate under a shared network structure and external storage. Experiments on 11 tomato diseases show that TDR—EGS significantly improves the performance of various mainstream baseline models without adding inference complexity, achieving recognition accuracy up to 98.61%. Moreover, even with only 60% of the label information, TDR—EGSs performance still approaches or exceeds the fullysupervised baseline models, demonstrating its effectiveness and generalization ability. This study provides an efficient and universal solution for agricultural disease recognition applications.

Key words: tomato leaf, disease recognition, graph learning, knearest neighbor graph, alternating training, deep learning

摘要： 番茄作为重要的蔬菜作物，其产量和质量常受到各类叶部病害的影响。针对此问题，计算机视觉技术已被广泛应用于病害的自动识别中。现有方法主要分为基于手工特征提取与深度学习两大类。基于手工特征的方法虽然简洁高效，但在鲁棒性方面存在限制；而基于深度学习的方法，尽管能有效提升识别准确性，但往往需要较大的数据标注量与较高的计算复杂性。为解决这些问题，提出一种基于图结构增强的番茄叶部病害识别框架（TDR—EGS）。TDR—EGS通过整合样本间的拓扑关系，实现图学习与单样本学习的交替训练，从而在不增加模型推理阶段复杂度的前提下有效提升分类性能。首先通过卷积神经网络提取单样本特征，然后利用这些特征构建k近邻图以挖掘样本间的结构信息。这种方法使得图学习和单样本学习能够在共享的网络结构和外部存储机制的支持下协同工作。在11种番茄病害上的试验结果表明，TDR—EGS能在不增加推理复杂度的前提下有效提升多种主流基准模型的性能，最高达到98.61%的识别精度。此外，即使在仅使用60%标签信息的条件下，TDR—EGS的性能仍可以接近或超过完全监督学习的基准模型，充分证明该框架的有效性和泛化能力，为农业病害识别应用提供一种高效且通用的解决方案。

关键词: 番茄叶部, 病害识别, 图学习, k近邻图, 交替训练, 深度学习