基于双注意力机制的可见光-红外行人重识别  

Visible-Infrared Person Re-Identification Based on Dual Attention Mechanism

作　　者：魏克铭 韩星宇 王辉[2] 范自柱 Wei Keming;Han Xingyu;Wang Hui;Fan Zizhu(School of Science,East China Jiaotong University,Nanchang 330013,China;Key Laboratory of Advanced Control and Optimization of Jiangxi Province,East China Jiaotong University,Nanchang 330013,China)

机构地区：[1]华东交通大学理学院,江西南昌330013 [2]华东交通大学江西省先进控制与优化重点实验室,江西南昌330013

出　　处：《华东交通大学学报》2024年第2期87-94,共8页Journal of East China Jiaotong University

基　　金：国家自然科学基金项目(61991401);江西省自然科学基金项目(20224BAB212014)。

摘　　要：【目的】由于可见光图像和红外图像之间的巨大模态差异,导致可见光-红外行人重识别是一项非常具有挑战性的图像检索问题。【方法】为了进一步减小两种模态之间的差异,重点关注行人信息,提出一种基于双注意力机制的网络结构用于可见光-红外行人重识别。一方面通过双注意力机制挖掘不同尺度的行人空间信息和增强局部特征的通道交互能力,另一方面利用全局分支和局部分支,学习多粒度的特征信息,使不同粒度信息可以相互补充,形成一个更具辨别性的特征。【结果】在两个公共数据集上的实验结果表明,该方法相较于基线有明显的提升,在RegDB数据集和SYSU-MM01数据集上均表现出理想的性能。【结论】该方法可为以后解决可见光-红外行人重识别的模态差异问题提供有效的参考。【Objective】Visible-infrared person re-identification is a very challenging image retrieval problem due to the huge modal difference between visible and infrared images.【Method】In order to further reduce the difference between the two modalities and focus on pedestrian information,a network structure based on dual attention mechanism is proposed for visible-infrared person re-identification.On the one hand,through the dual attention mechanism to mine personal spatial information of different scales and enhance the channel interaction ability of local features.On the other hand,through learning multi-granular feature information through using global and local branches,different granular information can complement with each other to form a more discriminating feature.【Result】Experimental results on two public datasets show that the proposed method has a significant improvement compared with the baseline,and shows ideal performance on both the RegDB dataset and the SYSU-MM01 dataset.【Conclusion】The proposed method can provide an effective reference for solving the problem of modal difference of visible-infrared person re-identification in the future.

关 键 词：可见光-红外行人重识别 注意力机制 缓解模态差异 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] U-9[自动化与计算机技术—计算机科学与技术]