中国烟草学报2026，Vol.32Issue(2)：135-145,11.DOI:10.16472/j.chinatobacco.2025.T0334
基于改进Densenet的烟草叶片病害识别方法
Tobacco leaf disease recognition method based on improved DenseNet
徐天然 1王瑞 1韩清林 1张斌 1赵乾 2韩宜彤 3许向阳3
作者信息
- 1. 山东中烟工业有限责任公司滕州卷烟厂,山东省滕州市鲁班大道3001号 277599
- 2. 枣庄科技职业学院,山东省滕州市学院东路888号 277599
- 3. 北京理工大学自动化学院,北京市海淀区中关村南大街5号 100081
- 折叠
摘要
Abstract
[Background]This study aims to achieve real-time and accurate identification of tobacco leaf diseases using an improved DenseNet model,thereby ensuring the quality and yield of tobacco leaves.[Methods]A high-precision classification method for tobacco leaf diseases was developed using an improved DenseNet model based on multi-scale feature fusion and an attention mechanism.The SE-MSFF-Deep Block,consisting of SEBlock and multi-scale feature fusion,was employed to extract multi-scale features.A Vision Transformer(ViT)-style attention structure was added before the classification layer to integrate scattered disease feature regions.The Focal Loss function was introduced to enhance the recognition accuracy of hard-to-classify and minority samples.Transfer learning was applied to mitigate overfitting on small datasets.[Results]The SMVF-DenseNet model achieved an accuracy of 97.58％for tobacco leaf diseases with an FPS of 23.69.Compared to other mainstream models,it significantly reduced inter-class confusion for difficult-to-class and minority samples.Moreover,the SMVF-DenseNet model maintained competitive classification accuracy on cross-crop plant disease datasets,demonstrating strong generalization capabilities.[Conclusion]The tobacco leaf disease recognition method based on multi-scale feature fusion and the attention mechanism effectively captures both detailed and overall features of disease patches,improving classification accuracy for difficult categories.Validation on additional plant disease datasets shows that the method excels in tobacco disease recognition and demonstrates strong cross-crop generalization,providing a feasible solution for broadly applicable plant disease identification models.关键词
烟草病害识别/DenseNet/多尺度特征融合/注意力机制Key words
tobacco disease recognition/DenseNet/multi-scale feature fusion/attention mechanism引用本文复制引用
徐天然,王瑞,韩清林,张斌,赵乾,韩宜彤,许向阳..基于改进Densenet的烟草叶片病害识别方法[J].中国烟草学报,2026,32(2):135-145,11.
