基于区域特征补全和细粒度特征注意力的行人重识别方法  

Person Re-Identification Method Based on Region Feature Completion and Fine-Grained Feature Attention

作　　者：孙志伟 吴广群 马永军[1] SUN Zhiwei;WU Guangqun;MA Yongjun(College of Artificial Intelligence,Tianjin University of Science&Technology,Tianjin 300457,China)

机构地区：[1]天津科技大学人工智能学院,天津300457

出　　处：《天津科技大学学报》2024年第5期63-71,共9页Journal of Tianjin University of Science & Technology

基　　金：国家自然科学基金资助项目(61976156);天津市自然科学基金资助项目(18JCQNJC69500)。

摘　　要：针对现有行人重识别方法对行人图像中遮挡区域特征利用率较低和行人图像中提取的特征细粒度不足的问题,提出了一种基于区域特征补全和细粒度特征注意力的行人重识别方法。首先,提出细粒度特征注意力模块,通过对输入特征进行分割加权重组,并添加注意力机制,以获得细粒度特征;其次,引入区域特征补全模块,通过将输入特征分块聚类,使遮挡区域特征能够通过相同聚类的特征恢复;最后,使用身份损失、加权正则化三元组损失和中心损失对模型进行训练优化。在公开数据集Market-1501、DukeMTMC-reID上进行实验,结果表明模型识别效果有所提升。To address the issues of low utilization of occluded region features and insufficient granularity of extracted fea-tures in existing person re-identification methods,a person re-identification method based on region feature completion and fine-grained feature attention is proposed in this article.Firstly,a fine-grained feature attention module is proposed to segment,weight,and recombine the input features,with the addition of attention mechanisms to obtain fine-grained features.Secondly,a region feature completion module is introduced,which clusters the input features into blocks to enable the recovery of occluded region features through features in the same cluster.Finally,the model is optimized with the use of identity loss,weighted regularized triplet loss,and center loss.Experimental results on publicly available datasets Market-1501 and DukeMTMC-reID demonstrated that the proposed model achieved improved recognition performance.

关 键 词：行人重识别 区域特征补全 细粒度 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]