基于改进自蒸馏EMA一DeiT与迁移学习的番茄叶片病害识别方法

打开文本图片集

中图分类号：S436.412 文献标识码：A 文章编号：2095-5553（2025）10-0192-12

Abstract：At present，the existing tomato leaf disease recognitionmodelhas the problems poor generalization low recognition accuracy.This paper proposesa tomato leaf disease recognition model basedon improvedEMA—iT that addresses theisues.The method uses the mainstream model Data-eficient image Transformers（iT）as the benchmark model，combines theExponential Moving Averagealgorithm with the iTmodel，adoptsself-knowledgedistilation method，enhances thetomatoleaf diseaserecognitionalgorithm through transfer learning toimproveitsrecognition accuracy．The model iscompared with VGG16，nsNet121，EficientNet，ResNet50，Vision Transformer，Swin Transformer， DCNN models，showing an accuracy improvement 1.1% to 14% .The improved EMA—iT model achieves an average accuracy 99.6% in the PlantVillage ten-class tomato leaf disease dataset. To evaluate the recognition performance tomato leaf diseases incomplex natural environments，this model wastested onthe Dataset

Tomato Leaves six-class tomato leafdiseasedataset，the PlantDoc dataset， the Tomato-Village dataset，achieving an average recognition rate 98.2% ， 97.1% ， 97.6% ，respectively.Compared with other models，the model demonstrated superior performance.The improved EMA—iT model proposed in thispaper has high recognition accuracyfor tomatoleaf diseasesincomplexnatural environmentscanefectivelyaid decision-making for tomato disease recognition systems in agricultural production

Keywords：tomato disease recognition；transfer learning；self-distilation；exponential moving average

0 引言

目前，中国位居亚洲番茄产量第一，同时也是全球最大的番茄生产国[1，截至2024年，中国的番茄产量种植面积约 120khm2 ，超过了美国加州，成为世界第一大加工番茄生产基地[2]。（剩余19025字）