结合多尺度特征与混淆学习的跨模态行人重识别  

Cross-modal person re-identification combining multi-scale features and confusion learning

作　　者：王路遥 王凤随 闫涛[1,2,3] 陈元妹 WANG Luyao;WANG Fengsui;YAN Tao;CHEN Yuanmei(School of Electrical Engineering,Anhui Polytechnic University,Wuhu 241000,China;Anhui Key Laboratory of Detection Technology and Energy Saving Devices,Anhui Polytechnic University,Wuhu 241000,China;Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment,Ministry of Education,Anhui Polytechnic University,Wuhu 241000,China)

机构地区：[1]安徽工程大学电气工程学院,安徽芜湖241000 [2]安徽工程大学检测技术与节能装置安徽省重点实验室,安徽芜湖241000 [3]安徽工程大学高端装备先进感知与智能控制教育部重点实验室,安徽芜湖241000

出　　处：《智能系统学报》2024年第4期898-908,共11页CAAI Transactions on Intelligent Systems

基　　金：安徽省自然科学基金项目(2108085MF197);安徽高校省级自然科学研究重点项目(KJ2019A0162);安徽工程大学国家自然科学基金预研项目(Xjky2022040).

摘　　要：跨模态行人重识别研究的重难点主要来自于行人图像之间巨大的模态差异和模态内差异。针对这些问题,提出一种结合多尺度特征与混淆学习的网络结构。为实现高效的特征提取、缩小模态内差异,将网络设计为多尺度特征互补的形式,分别学习行人的局部细化特征与全局粗糙特征,从细粒度和粗粒度两方面来增强网络的特征表达能力。利用混淆学习策略,模糊网络的模态识别反馈,挖掘稳定且有效的模态无关属性应对模态差异,来提高特征对模态变化的鲁棒性。在大规模数据集SYSU-MM01的全搜索模式下该算法首位击中率和平均精度(mean average precision,mAP)的结果分别为76.69%和72.45%,在RegDB数据集的可见光到红外模式下该算法首位击中率和mAP的结果分别为94.62%和94.60%,优于现有的主要方法,验证了所提方法的有效性。The difficulties of cross-modal person re-identification research mainly come from the huge modal differences and intra-modal differences between pedestrian images.To address these issues,a network structure combining multi-scale features with obfuscation learning is proposed.In order to achieve high-efficiency feature extraction and reduce intra-modal differences,the network is designed as a complementary form of multi-scale features to learn local refinement features and global rough features of pedestrians respectively.The feature expression ability of the network is enhanced from fine-grained and coarse-grained aspects.Confusion learning strategy is used to fuzzy the modal identification feedback of the network,and mine the stable and effective modal-independent attributes to cope with modal differences,so as to improve the robustness of features to modal changes.In the all-search mode of the large-scale data set SYSU-MM01,the results of the first hit rate and mean average precision(mAP)of the algorithm are 76.69%and 72.45%,respectively.In the Visible to Infrared mode of the RegDB data set,the results of the first hit rate and mAP of the algorithm are 94.62%and 94.60%,respectively,which are better than the main existing methods,verifying effectiveness of the proposed method.

关 键 词：机器视觉 行人重识别 跨模态 多尺度特征 粗粒度 细粒度 混淆学习 模态无关属性 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]