基于强区分性特征挖掘的无监督行人重识别  

Unsupervised Person Re-Identification Based on Discriminative Feature Mining

作　　者：徐锟 王晓红[1] 赵徐 徐世豪 Kun Xu;Xiaohong Wang;Xu Zhao;Shihao Xu(College of Communication and Art Design,University of Shanghai for Science and Technology,Shanghai)

机构地区：[1]上海理工大学出版印刷与艺术设计学院,上海

出　　处：《建模与仿真》2024年第3期2011-2022,共12页Modeling and Simulation

摘　　要：目前,基于深度学习的无监督行人重识别正在通过改进聚类方法来提高生成伪标签的准确性和鲁棒性。然而,行人的固有物理特征,如肢体和身体,尚未充分利用。此外,遮挡和姿势导致了在行人图像中准确匹配和识别局部区域的困难,而行人服装颜色引起的噪声干扰也对行人重识别准确性产生了重要影响。为了解决这些问题,本文提出了一种判别性特征挖掘网络(DFMN),通过引入Transformer注意力机制来突出显示行人的有效物理特征;为了深入挖掘不同样本之间的细粒度局部特征,本文采用了基于最短路径的部分对齐分割机制;同时,针对行人服装颜色引起的冗余信息,本文融合了一个alpha通道,可以有效消除噪声干扰。实验结果表明,本文所提出的方法在Market1501数据集和MSMT17数据集上的map指标上分别实现了4.0%和6.4%的提升。At present,Unsupervised Person Re-Identification based on deep learning promotes the accuracy and robustness of generating pseudo labels mainly by improving the clustering method.However,for inherent physical features of pedestrians,such as limbs,and bodies,which are not fully utilized,occlusion,posture pose difficulties in accurately matching and recognizing local areas in pedestrian images,and noise interference caused by pedestrian clothing color have essential impact on person re-identification accuracy.In this paper,we propose a Discriminative Feature Minning Network(DFMN)by introducing a Transformer attention mechanism to highlight the pedestrians’effective physical features.To deeply mining the fine-grained local features among different samples,a part-aligned segmentation mechanism based on the shortest path is used.At the same time,as for the redundant information resulting from the color of pedestrian clothing,we incorporate an alpha channel,which can effectively eliminate noise interference.The experimental results show that the proposed method has achieved improvements of 4.0%and 6.4%on the Market1501 dataset and the MSMT17 dataset for the map metric.

关 键 词：特征挖掘 无监督学习 伪标签 聚类 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]