基于孪生并行注意力网络的包装印刷品商标真伪鉴别研究  

Research on Discriminating the Authenticity of Packaging Printed Logos Based on Siamese Parallel Attention Network

作　　者：王晓红[1] 宛东 WANG Xiaohong;WAN Dong(College of Communication and Art Design,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学出版印刷与艺术设计学院,上海200093

出　　处：《包装学报》2023年第1期86-94,共9页Packaging Journal

基　　金：国家新闻出版署智能与绿色柔版印刷重点实验室招标课题资助(ZBKT202108)

摘　　要：为了借助手机精确快速地鉴别微小篡改的包装印刷品商标真伪,提出孪生并行注意力卷积神经网络判别模型。通过孪生网络的共享权重机制降低网络系统的表征偏差,通过并行注意力机制提高对微小篡改变化特征的提取能力,最大程度降低打印拍照引入的噪声对篡改特征提取的影响。在2种印刷纸张、2种拍摄光源和2种拍摄手机组合的8种开放场景中,拍摄多组真伪商标的印刷图像,建立篡改面积0.4%~0.7%的商标真伪数据集。模型在该打印拍照数据集上鉴别准确率为94%以上,在真实山寨商标上的鉴别准确率为100%。本文提出的孪生并行注意力卷积神经网络模型,具有较高的细粒度鉴别精度和较强的泛化能力,能够在开放场景下有效地实现基于图像微小篡改的包装印刷品商标真伪的鉴别。In order to accurately and quickly identify the authenticity of the slightly tampered packaging printed logos with the help of mobile phones,an authenticity discrimination algorithm based on Siamese parallel attention convolutional neural network is proposed.The algorithm can reduce the representation bias of the network system through the shared weight mechanism of the Siamese network,improve the ability to extract minor tampering features through the parallel attention mechanism,and minimize the influence of noise introduced by printing and photographing on tampering feature extraction.The logo authenticity dataset with tampering area of 0.4%to 0.7%was established by photographing multiple forged logo printed images in 8 scenes combining 2 types of printing papers,2 kinds of shooting light sources and 2 mobile devices.The discrimination accuracy of the model on the Print-Photo dataset is more than 94%,and the discrimination accuracy on the real counterfeit trademark is 100%.The experimental results show that the Siamese parallel attention convolutional neural network has high fine-grained discrimination accuracy and strong generalization ability,and can effectively realize the identification of the authenticity of packaging printed logos based on image minor tampering in open scenarios.

关 键 词：孪生卷积神经网络 双注意力机制 微小篡改 商标真伪鉴别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TS851.6[自动化与计算机技术—计算机科学与技术]