基于轻量多分支网络的行人重识别方法  

Person re-identification method based on lightweight multi-branch network

作　　者：罗丽洁 韩华[1] 金婕[1] 黄丽 LUO Lijie;HAN Hua;JIN Jie;HUANG Li(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《智能计算机与应用》2022年第11期103-110,共8页Intelligent Computer and Applications

基　　金：国家自然科学基金(61305014);上海市自然科学基金(22ZR1426200);上海市教育委员会和上海市教育发展基金会“晨光计划”资助项目(13CG60)。

摘　　要：最新的行人重识别方法大都是基于卷积神经网络(CNN),虽然这些网络在分类或目标检测等多项任务中有着不错的表现,但这些方法大都侧重于图像最具辨别力的部分,忽视了其他的一些相关特征,而重识别任务需要更加丰富,具有多样性的特征。本文提出了一种基于OSNet(Omni-scale Network)的多分支网络结构,OSNet是一个轻量级的Re-ID模型,可将标准卷积分解为点卷积和深度卷积以便达到减少参数的目的。网络主干部分加入了注意力模块,可以抑制无用信息;而多分支的结构可以提取到更多样性的特征。该网络主要由全局分支、局部分支、顶部擦除分支和通道分支四个部分构成。全局分支用于提取图像的整体特征;局部分支能学习到细粒度的特征;顶部擦除分支通过擦除高激活性区域,使得网络更加关注于激活性较差的区域;通道分支使网络学习到更多与通道有关的信息。在Market-1501、CUHK03、DukeMTMC-reID三个公开数据集上的实验结果表明,本文提出的方法针对行人重识别问题有着优秀的性能表现。Most of the latest person re-identification(Re-ID) methods are based on convolutional neural networks(CNN). Although these networks have good performance in many tasks, such as classification or object detection, these methods often focus on the most discriminative part of an image and ignore some other relevant features. The Re-ID tasks need more abundant and diverse features. This paper proposes a multi-branch network structure based on OSNet(Omni-scale Network). OSNet is a lightweight Re-ID model, which decomposes the standard convolution into pointwise convolution and depthwise convolution in order to reduce parameters.The attention modules are added to the network backbone, which could inhibit useless information, and the multi-branch structure can extract more diverse features. It is mainly composed of four parts: global branch, local branch, top erased branch and channel branch. The global branch is used to extract the global features of pedestrian images;local branches can learn more fine-grained features;the top erased branch makes the network pay more attention to the low informative regions by erasing the most activated regions;channel branch can make the network learn more channel information. The experimental results on Market-1501、CUHK03 and DukeMTMC-reID show that the proposed method has excellent performance for person re-identification.

关 键 词：行人重识别 神经网络 多分支网络 注意力模块 顶端擦除 

分 类 号：TP931.41[自动化与计算机技术]