【摘要】 针对番茄种植中病害程度识别难度大、检测成本高等问题，以及传统方法主观性强、时效性差等缺陷，提出了一种新的Tomato Vit模型，包含3个核心模块:(1)基于Res Net-50和Vision Transformer的混合主干网络，用于提取病害特征;(2)特征自验证模块(FSV)，通过token替换和预测机制增强模型对病害特征细微变化的识别能力;(3)多尺度局部全局注意力模块(MSGL)，结合全局与局部注意力机制，更好地捕获病害的微小变化特征。该方法将番茄病害检测从单一分类扩展至病害程度评估，将番茄的8种常见病害细分为健康、早期和严重3个等级，实现对病害发展的不同程度监测。在包含18个类别的番茄病害程度数据集上进行实验，模型达到89.79%的准确率和0.90的F1-score，优于现有方法。更多还原

【Abstract】 To address the challenges of identifying diseases in tomato cultivation,such as detection costs,strong subjectivity and poor timeliness,this paper proposes a tomato disease classification and grading method based on Vision Transformer. A novel model,TomatoVit,is introduced,which consists of three core modules:( 1) a hybrid backbone network based on ResNet-50 and Vision Transformer for extracting disease features;( 2) a Feature Self-Validation( FSV) module that enhances the model ’s ability to recognize subtle variations in disease features through token replacement and prediction mechanisms; and( 3) a Multi-Scale Global-Local Attention( MSGL) module,which combines global and local attention mechanisms to better capture the fine-grained features of disease variations. This approach transforms tomato disease detection from simple classification to severity evaluation,subdividing eight common tomato diseases into three stages: healthy,early,and severe stage. It enables monitoring of disease progression.Experiments conducted on a tomato disease dataset comprising 18 categories demonstrate the proposed model achieves an accuracy of 89. 79% and an F1-score of 0. 90,outperforming other methods.更多还原