基于改进ConvNeXt模型的真假卷烟烟丝识别方法  

Method for identifying cut tobacco from genuine and fake cigarettes based on improved ConvNeXt model

作　　者：王树才[1] 黄开虎 丁美宙 纪晓楠 陶栩 WANG Shucai;HUANG Kaihu;DING Meizhou;JI Xiaonan;TAO Xu(College of Engineering,Huazhong Agricultural University,Wuhan 430070,China;Technology Center,China Tobacco Henan Industrial Co.,Ltd.,Zhengzhou 450000,China)

机构地区：[1]华中农业大学工学院,武汉市430070 [2]河南中烟工业有限责任公司技术中心,郑州市450000

出　　处：《烟草科技》2024年第5期103-112,共10页Tobacco Science & Technology

基　　金：烟草行业烟草加工形态研究重点实验室项目“烟草物料二维形态特征与成丝特性研究”(A202019)。

摘　　要：为实现真假卷烟的快速识别,基于卷积神经网络ConvNeXt模型提出一种不同品牌卷烟烟丝分类识别方法。采集4种品牌卷烟的真假烟丝图像,制作深度学习数据集。基于ConvNeXt模型,引入卷积注意力模块(Convolutional Block Attention Module,CBAM),提高模型的特征提取能力;搭建特征金字塔结构,实现不同尺度的特征融合,增强模型对烟丝图像的特征表达能力;在多尺度融合结构中引入GhostNetV2卷积,降低模型复杂度和计算量。将改进后ConvNeXt_CM模型以及常用的图像分类模型ResNet50、DensNet121、EfficientNetV2进行对比测试,结果表明:①相比原始ConvNeXt-模型,ConvNeXt_CM模型的宏F1分数达到95.46%,平均精度值达到87.13%,宏精确度与宏召回率分别提升6.08、6.13百分点,模型大小为27.31 M,识别单张图像平均用时0.024 s/张。②与ResNet50、DensNet121、EfficientNetV2模型相比,ConvNeXt_CM模型图像识别性能更加优异,宏F1分数分别提升21.94、20.19、31.05百分点。该方法可为提升模型的图像识别能力、完成烟丝分类识别任务提供支持。In order to quickly identify genuine and fake cigarettes,a method for classification and identification of cut tobacco from cigarettes of different brands based on ConvNeXt convolutional neural network model was proposed.The images of cut tobacco from genuine and fake cigarettes of four brands were collected to create a deep learning dataset.Based on the ConvNeXt model,a Convolutional Block Attention Module(CBAM)was introduced to improve the feature extraction capability of the model.A feature pyramid structure was developed to achieve feature fusion of different scales and improve the feature expression ability of the model.GhostNetV2 convolution was introduced into the multi-scale fusion structure to reduce model complexity and computational work.The improved ConvNeXt_CM model and the commonly used image classification models including ResNet50,DensNet121,and EfficientNetV2 were comparatively tested.The results showed that:1)Compared with the original ConvNeXt model,the macro F1 score and average accuracy of the ConvNeXt_CM model reached 95.46%and 87.13%,respectively;its macro precision and macro recall rates elevated by 6.08 and 6.13 percentage points,respectively.The size of the model was 27.31 M,and the time needed for identifying an image averaged 0.024 s.2)The ConvNeXt_CM model advantaged over the ResNet50,DensNet121,and EfficientNetV2 models in image recognition efficiency,and its macro F1 score elevated by 21.94,20.19,and 31.05 percentage points,respectively.The proposed method helps improve the image recognition capabilities of the models and cigarette authentication.

关 键 词：卷烟 烟丝图像 ConvNeXt 真假识别 卷积神经网络 

分 类 号：TS46[农业科学—烟草工业] TS439