基于迁移学习的MobileViT-CBAM鲜烟叶成熟度识别模型研究  

MobileViT-CBAM Model for Fresh Tobacco Leaf Maturity Recognition Based on Transfer Learning

作　　者：赵泮真 王松峰[1] 齐飞 胡强 王爱华[1] 李亚纯 孟令峰 尹东 段史江 王志生 ZHAO Panzhen;WANG Songfeng;QI Fei;HU Qiang;WANG Aihua;LI Yachun;MENG Lingfeng;YIN Dong;DUAN Shijiang;WANG Zhisheng(Institute of Tobacco Research of Chinese Academy of Agricultural Sciences/Key Laboratory of Tobacco Biology and Processing,Ministry of Agriculture and Rural Affairs,Qingdao 266101,China;Graduate School of Chinese Academy of Agricultural Sciences,Beijing 100081,China;Fuzhou Branch of Jiangxi Provincial Tobacco Company,Fuzhou 344000,Jiangxi,China;Ji’an Branch of Jiangxi Provincial Tobacco Company,Ji’an 343009,Jiangxi,China)

机构地区：[1]中国农业科学院烟草研究所/农业农村部烟草生物学与加工重点实验室,青岛266101 [2]中国农业科学院研究生院,北京100081 [3]江西省烟草公司抚州市公司,江西抚州344000 [4]江西省烟草公司吉安市公司,江西吉安343009

出　　处：《中国烟草科学》2025年第2期93-100,共8页Chinese Tobacco Science

基　　金：中国烟草总公司江西省烟草公司科技项目(202201011);中国烟草总公司科技重点项目(110202102007);中国农业科学院科技创新工程(ASTIP-TRIC03)。

摘　　要：为建立更加经济高效的烟叶成熟度无损智能识别技术,本研究构建了一种可以部署到移动设备的轻量级网络模型MobileViT-CBAM。首先采集云烟87中、上部不同成熟度烟叶图像构建数据集,将CBAM注意力机制模块引入到MobileViT结构中增强鲜烟叶成熟度图像特征的表达能力,其次将原有激活函数Swish函数替换为更平滑的SMU函数,帮助模型更快收敛,最后使用迁移学习提高模型的训练效率和泛化能力,以实现复杂大田环境下鲜烟叶成熟度分类。结果表明,MobileViT-CBAM在鲜烟叶成熟度分类任务中准确率达92.81%,较VGG16、ResNet34、Vision Transformer、Swin Transformer、MobileNetV2和MobileViT等模型具有显著的性能提升。所提出的MobileViT-CBAM模型能够有效识别烟叶成熟程度,为烟叶智能采集装备视觉系统提供技术支撑。To establish more economical and efficient non-destructive intelligent recognition technology for tobacco leaf maturity,a lightweight network model MobileViT-CBAM on mobile devices was constructed.Firstly,a dataset was built by collecting images of the middle and upper leaves of‘Yunyan87’with different maturity.The CBAM attention mechanism module was introduced into the MobileViT structure to enhance the feature expression ability of fresh tobacco leaf maturity images.Secondly,the original activation function Swish was replaced with the smoother SMU function to help the model converge faster.Finally,transfer learning was employed to improve the training efficiency and generalization ability of the model and achieve the classification of fresh tobacco leaf maturity in complex field environment.Results showed that MobileViT-CBAM exhibited an accuracy of 92.81%in maturity classification of fresh tobacco leaves,which is significantly superior to the models of VGG16,ResNet34,Vision Transformer,Swin Transformer,MobileNetV2,and MobileViT.The proposed MobileViT-CBAM model can effectively identify the maturity degree of tobacco leaves,providing technical support for the visual system of intelligent tobacco harvesting equipment.

关 键 词：鲜烟叶成熟度 轻量级 分类模型 注意力机制 

分 类 号：S572[农业科学—烟草工业] S126[农业科学—作物学]