摘要
现有无监督行人重识别算法使用残差网络,仅能提取粗略的全局特征,对细微的局部特征反映不足,且聚类方法生成的伪标签会引入噪声,影响特征判别。针对上述问题,提出一种双分支引导对比学习的方法。首先,引入一种有效的特征提取方式,将提取的特征分为全局分支和局部分支,提高对局部信息的利用;其次,通过全局特征和局部特征之间的一致性细化全局特征预测的伪标签,充分利用局部特征和整体特征之间的互补关系,有效降低伪标签聚类产生的噪声;最后,引入对比学习模块,将细化的标签进行对比学习,提高模型的鲁棒性。在Market1501、DukeMTMC-ReID以及MSMT17数据集上的实验结果验证了所提方法的有效性及高性能。
The current unsupervised pedestrian re-identification algorithms using residual networks can only extract rough global features,but it can’t adequately reflect subtle local features.In addition,the pseudo labels generated by clustering methods introduce noise,which will affect the performance of feature discrimination.A dual branch guided contrastive learning method was proposed.Firstly,an effective feature extraction method was introduced,which divided the extracted features into global branches and local branches to improve the utilization of local information.Secondly,the consistency between global and local features was proposed to refine the pseudo labels for global feature prediction,utilizing the complementary relationship between local and global features,thereby effectively reducing the noise generated by pseudo label clustering.Finally,a contrastive learning module was proposed to perform contrastive learning on refined labels and improve the robustness of the model.The experimental results on the Market1501,DukeMTMC-ReID,and MSMT17 datasets validate the effectiveness of the proposed method.
作者
任航佳
梁凤梅
REN Hangjia;LIANG Fengmei(College of Electronic Information Engineering,Taiyuan University of Technology,Jinzhong 030600,China)
出处
《电信科学》
北大核心
2025年第6期92-102,共11页
Telecommunications Science
基金
虚拟现实技术与系统全国重点实验室开放课题基金项目(No.VRLAB2023A06)
山西省科技合作交流专项(No.202104041101030)。
关键词
无监督行人重识别
全局特征
局部特征
标签细化
对比学习
unsupervised pedestrian re-identification
global feature
local feature
label refinement
contrastive learning
