摘要
针对换装行人重识别任务中由复杂环境和行人服装变化等因素导致的识别精度下降的问题,提出姿态引导的双分支换装行人重识别网络PGNet,该网络采用以外观特征为基础、由姿态特征引导的双分支结构.为了有效去除服装相关信息的干扰,降低其对模型性能的影响,同时保留深度表征特征,设计多层次特征融合模块;构建动作关联和自然拓扑邻接矩阵,组合为双重矩阵后输入图卷积网络,并引入邻接矩阵加权机制以增强模型对姿态特征的捕捉能力;采用双线性多特征池化方法增强姿态与外观特征的互补性,从而提升识别精度.实验结果表明,PGNet在换装数据集PRCC、VC-Clothes、Celeb-reID以及Celeb-reID-light上的mAP指标分别为60.5%、84.7%、15.7%、22.6%,Rank-1指标分别为63.7%、93.3%、59.5%、41.2%,优于SirNet等其他对比方法,验证了所提方法能够有效降低服装变化的影响,并显著提高识别精度.
A pose-guided dual-branch clothing-changing person re-identification network(PGNet)was proposed to address the issue of reduced recognition accuracy in the clothing-changing person re-identification tasks caused by complex environments and clothing variations.The network adopted a dual-branch architecture based on appearance features and guided by pose features.To effectively remove the interference of clothing-related information,reduce its impact on model performance,and preserve the deep representational features,a multi-level feature fusion module was designed.An action-related adjacency matrix and a natural topology adjacency matrix were constructed and combined to form a dual adjacency matrix,which was input into the graph convolutional network.An adjacency matrix weighting mechanism was introduced to enhance the model’s ability to capture pose features.A bilinear multi-feature pooling method was adopted to enhance the complementarity between the pose features and the appearance features,thereby improving the recognition accuracy.Experimental results demonstrated that the PGNet achieved mAP values of 60.5%,84.7%,15.7%,22.6%,and Rank-1 accuracies of 63.7%,93.3%,59.5%,41.2%on the clothing-changing datasets of PRCC,VC-Clothes,Celeb-reID,and Celeb-reID-light,respectively,outperforming other comparative methods such as SirNet.The proposed method can effectively reduce the impact of clothing variations and significantly improve the recognition accuracy.
作者
周思瑶
夏楠
江佳鸿
ZHOU Siyao;XIA Nan;JIANG Jiahong(School of Information Science and Engineering,Dalian Polytechnic University,Dalian 116034,China)
出处
《浙江大学学报(工学版)》
北大核心
2026年第1期71-80,共10页
Journal of Zhejiang University(Engineering Science)
基金
教育部产学合作协同育人资助项目(220603231024713)。
关键词
换装行人重识别
姿态引导
特征融合
图卷积网络
注意力机制
clothing-changing person re-identification
pose guide
feature fusion
graph convolutional network
attention mechanism
