基于高分辨率网络和注意力机制的真伪卷烟包装鉴别  

Authenticity Identification of Cigarettes Based on Attention Mechanism and High-resolution Network

作　　者：肖楠 周明珠[3] 邢军[3] 罗泽[1] 李晓辉[3] XIAO Nan;ZHOU Mingzhu;XING Jun;LUO Ze;LI Xiaohui(Computer Network Information Center,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100049,China;China National Tobacco Quality Supervision&Test Center,Zhengzhou,Henan 450001,China)

机构地区：[1]中国科学院计算机网络信息中心,北京100190 [2]中国科学院大学,北京100049 [3]国家烟草质量监督检验中心,河南郑州450001

出　　处：《数据与计算发展前沿》2021年第5期118-129,共12页Frontiers of Data & Computing

基　　金：中国烟草总公司科技重大专项项目“卷烟产品鉴别大数据构建及应用研究”(110201901026(SJ-05))。

摘　　要：【目的】针对真伪卷烟包装鉴别任务对类别精度要求高、深度残差等网络不能提取出更具判别力特征的问题,本文从保证图像高分辨率表征的角度出发,提出了结合高分辨率网络(High-Resolution Network,HRNet)和注意力机制的方法,以得到更具表现力的特征,从而达到提高真伪卷烟包装鉴别准确度的目的。【方法】以具备并行子网结构的高分辨率网络为骨干网络,通过多分辨率特征融合方法获得鉴别卷烟真伪的高质量特征,并在此网络基础上嵌入了高效通道注意力(Efficient Channel Attention,ECA)模块,有效地增强了通道之间的信息交互。【结果】经过实验验证,本文提出的方法不仅可以学习到更好的特征表示,而且准确率可达到97.21%。【局限】模型着重关注了通道维度的相关性,忽略了特征空间位置信息,还有改进空间。【结论】通过将高分辨率网络和注意力机制相结合,可以有效地提高卷烟真伪鉴别的准确度,并为相关研究提供了一种新的研究思路。[Objective]Authenticity identification of cigarettes requires high classification accuracy,and classic convolutional neural networks such as deep residual networks cannot extract sufficient discriminative features.Therefore,we propose a method combining high-resolution network and attention mechanism to obtain more expressive features.This method helps us achieve the purpose of improving the accuracy of authenticity cigarette packaging identification.[Methods]We take the high-resolution network,with parallel subnet structure,as the backbone network,and the high-quality features for identifying the authenticity of cigarettes are obtained through the multi-resolution feature fusion method.What’s more,we embedded the efficient channel attention(ECA)module into this network,which effectively enhances information exchange between different channels.[Results]Experimental results show that the method proposed in this paper can not only learn better feature representations but also achieve an accuracy of 97.21%.[Limitations]The model focuses on the correlation of channel dimensions,ignoring the location information of the feature space which may help to improve model performance.[Conclusions]By combining the high-resolution network and the attention mechanism,the accuracy of cigarette authenticity identification can be effectively improved,and a new research idea can be provided for related research.

关 键 词：卷烟包装 真伪鉴别 卷积神经网络 高分辨率网络 注意力机制 

分 类 号：TS47[农业科学—烟草工业]