基于全局多粒度池化的可见光红外行人重识别  

Visible infrared person re-identification based on global multi-granularity pooling

作　　者：周航 黄春光[1] 程海[1] Zhou Hang;Huang Chunguang;Cheng Hai(School of Electronic Engineering Institute,Heilongjiang University,Harbin 150000,China)

机构地区：[1]黑龙江大学电子工程学院,哈尔滨150000

出　　处：《电子测量技术》2022年第1期122-128,共7页Electronic Measurement Technology

基　　金：国家自然科学基金青年基金(51607059);黑龙江大学基础科学研究项目(KJCX201904,2020-KYYWF-1001)资助。

摘　　要：可见光红外行人重新识别是一种跨模态检索的问题。由于可见光和红外图像模态差异较大,能够精确的匹配行人仍然具有很大的挑战。最近的研究表明,利用池化描述身体部位的局部特征以及人图像本身的全局特征,即使在身体部位缺失的情况下,也能给出鲁棒的特征表示,但是简单的全局平均池化很难获取行人的细节特征。针对这个问题,提出一种新的全局多粒度池化的方法,利用全局平均池化和全局最大池化结合的方法,提取行人更多的背景和纹理信息。此外,传统的三元组损失在跨模态行人重识别上效果并不好。设计了一种新的跨模态三元损失,以优化类内和类间距离,并监督网络学习有区别的特征表示。通过实验证明了所提方法的有效性,并在RegDB和SYSU-MM01数据集上分别取得了88.01%Rank-1,79.26%mAP,和60.24%Rank-1,57.50%mAP的结果。Visible infrared person re-identification is a cross-modal retrieval problem.Being able to accurately match pedestrians remains challenging due to the large modal differences between visible and infrared images.Recent research has shown that using pooling to describe local features of body parts as well as global features of the human image itself can give a robust feature representation even when body parts are missing,but simple global average pooling is difficult to obtain detailed features of pedestrians.To address this problem,this paper proposes a new global multi-granularity pooling approach that uses a combination of global average pooling(GAP)and global maximum pooling(GMP)to extract more background and texture information of person.In addition,the traditional triplet loss does not work well for cross-modal person re-identification.We design a new cross-modal triplet loss to optimise intra-class and inter-class distances and supervise the network to learn differentiated feature representations.We experimentally demonstrate the effectiveness of the proposed method and achieves 88.01% Rank-1,79.26% mAP,and 60.24% Rank-1,57.50% mAP on the RegDB and SYSU-MM01 datasets,respectively.

关 键 词：全局多粒度池化 可见光红外行人重识别 困难跨模态三元损失 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]