基于Cu-ViT深度学习的烟草气候斑病害分级识别模型的开发应用

打开文本图片集

中图分类号：S435.72；TP181文献标识码：A文章编号：2097-1354（2025）06-0065-12

Abstract： Accurate grading and identification of tobacco climate spot disease hold multidimensional value in agricultural production， disease control，and environmental protection. Manual identification suffers from issues such as high costs， strong subjectivity，and low efficiency， which can be addressed through image recognition technology. This study， based on the Vision Transformer （ViT） framework，replaces patch embedding with compression units to propose the Cu-ViT model，systematically enhancing the ViT model's capability in image capture and recognition.In simulated tests，the Cu-ViT model achieved an accuracy of 91.23% ，with its F1 score，precision，and recall all surpassing those of ViT as well as advanced recognition models such as ResNet152，InceptionResNetV2，Swin Transformer （SwinT） and VGGNet19. The average recognition time per image was lO4.23 milliseconds. Furthermore，the Cu-ViT model's accuracy， validated in real production environments，outperformed manual identification （ p< 0.01）.These results indicate that the Cu-ViT model is capable of grading and identifying tobacco climate spot disease.

Key words： tobacco climate spot disease; disease grading;image recognition; deep learning

目前，全球有100多个国家和地区种植烟草，烟草种植行业已经成为一些国家的经济支柱[]。（剩余12068字）