数据与计算发展前沿 ›› 2021, Vol. 3 ›› Issue (5): 118-129.

doi: 10.11871/jfdc.issn.2096-742X.2021.05.009

• 技术与应用 • 上一篇    下一篇

基于高分辨率网络和注意力机制的真伪卷烟包装鉴别

肖楠1,2(),周明珠3(),邢军3(),罗泽1(),李晓辉3,*()   

  1. 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
  • 收稿日期:2021-03-12 出版日期:2021-10-20 发布日期:2021-11-24
  • 通讯作者: 李晓辉
  • 作者简介:肖楠,中国科学院计算机网络信息中心,中国科学院大学,硕士研究生,主要研究方向为机器学习与图像处理。
    本文承担工作为：实验设计、实现与论文撰写。
    XIAO Nan is a student at Computer Network Information Center of the Chinese Academy of Sciences / University of the Chinese Academy of Sciences. Her main research directions are machine learning and image processing.
    In this paper, she undertakes the work of experimental design, implementation, and paper writing.
    E-mail: xiaonan@cnic.cn|周明珠,国家烟草质量监督检验中心, 高级工程师,主要研究方向为烟草物理性能检测技术与标准研究。
    本文承担工作为：方法优化。
    ZHOU Mingzhu is a senior engineer at China National Tobacco Quality Supervision Test Center. Her research directions are on testing technologies and standards of tobacco physical properties.
    In this paper, she undertakes the work of method optimization.
    E-mail: zhoumz@aztri.com.cn|邢军,国家烟草质量监督检验中心,研究员,主要研究方向为检测技术与标准研究。
    本文承担工作为：现状分析。
    XING Jun is a research fellow at China National Tobacco Quality Supervision Test Center. Her research directions are on testing technologies and standards.
    In this paper, she undertakes the work of the analysis of current studying status.
    E-mail: xingj@ztri.com.cn|罗泽,中国科学院计算机网络信息中心,研究员,博士生导师,主要研究方向为海量数据分布处理理论和方法,数据挖掘和机器学习理论、方法和应用。
    本文承担工作为：总结与优化指导。
    LUO Ze, Ph.D. Supervisor, is a research fellow at Computer Network Information Center of the Chinese Academy of Sciences / University of the Chinese Academy of Sciences. His research directions are massive data distribution processing theory and method, data mining and machine learning theory, method and application.
    In this paper, he undertakes the work of summary and optimiz-ation guidance.
    E-mail: luoze@cnic.cn|李晓辉,国家烟草质量监督检验中心,高级工程师,主要研究方向为烟草物理性能检测技术与标准研究。
    本文承担工作为：流程设计和应用分析。
    LI Xiaohui is a senior engineer at China National Tobacco Quality Supervision Test Center. Her research directions are on testing technologies and standards of tobacco physical properties.
    In this paper, she undertakes the work of process design and application analysis.
    E-mail: lixh@aztri.com.cn
  • 基金资助:
    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”(110201901026SJ-05)

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

XIAO Nan1,2(),ZHOU Mingzhu3(),XING Jun3(),LUO Ze1(),LI Xiaohui3,*()   

  1. 1. Computer Network Information Center, Chinese Academy of Sciences, Beijing 100190, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
    3. China National Tobacco Quality Supervision & Test Center, Zhengzhou, Henan 450001, China
  • Received:2021-03-12 Online:2021-10-20 Published:2021-11-24
  • Contact: LI Xiaohui

RichHTML

18

PDF

82

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。

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

Abstract:

[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.

Key words: cigarette packet, authentication, convolutional neural networks, high-resolution networks, attention mechanism