基于伪标签正则化损失的无监督行人重识别  

Unsupervised Person Re-Identification with Pseudo Label Regularization Loss

作　　者：贾洁茹 张硕蕊 钱宇华 阮秋琦[4] JIA Jie-ru;ZHANG Shuo-rui;QIAN Yu-hua;RUAN Qiu-qi(Institute of Big Data Science and Industry,Shanxi University,Taiyuan,Shanxi 030006,China;School of Information and Technology,Shanxi University,Taiyuan,Shanxi 030006,China;Engineering Research Center for Machine Vision and Data Mining of Shanxi,Province,Taiyuan,Shanxi 030006,China;Institute of Information Science,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]山西大学大数据科学与产业研究院,山西太原030006 [2]山西大学计算机与信息技术学院,山西太原030006 [3]山西省机器视觉与数据挖掘工程研究中心,山西太原030006 [4]北京交通大学信息科学研究所,北京100044

出　　处：《电子学报》2024年第5期1743-1758,共16页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62106133)。

摘　　要：无监督行人重识别旨在不需要行人身份标签的情况下,将查询的行人图像与候选集中的行人图像相匹配.目前主流的无监督行人重识别方法通常先利用聚类算法生成伪标签,然后利用伪标签训练深度神经网络.然而由于模型初始表征能力不足和聚类算法的局限性等,伪标签中会引入大量噪声,严重误导模型优化过程,导致模型性能退化.为了减轻伪标签噪声的影响,本文提出了一种新的伪标签正则化损失函数,用伪标签的置信度分数和样本相似度对伪标签噪声进行约束.具体来说,本文首先提出了一种聚类引导的注意力机制,根据伪标签与聚类中心的语义相关程度来估计伪标签的置信度,以此来识别噪声标签并给正确标签分配更多的权重,有效降低伪标签噪声在总体损失函数中的作用.同时,为了充分利用伪标签的判别能力,本文利用伪标签进行在线软样本挖掘,构建mini-batch中的正负样本对并为每个正负样本对计算一个连续的权重分数.通过将以上两种权重引入到对比损失中,本文提出的伪标签正则化损失函数可以有效抑制伪标签噪声的影响,减轻标签噪声对训练过程的影响,提高模型的准确性和鲁棒性.在多个公开行人数据集上的实验结果验证了本文方法的有效性,在Market1501、DukeMTMC-reID和MSMT17数据集上mAP分别达到了85.9%、75.1%和29.3%.Unsupervised person re-identification aims to match query pedestrian images with images in the gallery without the need for identity labels.Currently,mainstream unsupervised person re-identification methods typically utilize clustering algorithms to generate pseudo-labels,which are subsequently exploited to train deep neural networks.However,due to the model’s inferior representation ability at early stages and the limitations of the clustering algorithms,a vast of noise is inevitably introduced into the pseudo-labels,which seriously misleads the training process and impedes the model performance.In this paper,we propose a novel pseudo-label regularization loss(PLRL)to remedy the detrimental effect of pseudo-label noises.Concretely,firstly,this paper proposes a clustering-guided attention mechanism(CGA)to estimate the confidence of pseudo-labels based on the semantic relevance between pseudo-labels and clustering centers.The CGA score is able to identify noisy labels and assign more weight to correct labels,which effectively reduces the influence of pseudo-label noise in the overall loss function.Meanwhile,for the sake of fully utilizing the discriminative power of pseudo-labels,this paper performs soft sample mining using pseudo-labels,which constructs positive and negative sample pairs in minibatches and calculates a continuous weight score for each pair.By incorporating the confidence of pseudo-labels and the similarity of samples into the contrastive loss,the newly designed pseudo-label regularization loss can effectively alleviate the influence of pseudo-label noise in the training process,thereby improving the accuracy and robustness of the model.Experiments and ablation studies on multiple public datasets demonstrate its effectiveness and superiority,with the mAP on Market1501,DukeMTMC-reID,and MSMT17 datasets reaching 85.9%,75.1%,and 29.3%,respectively.

关 键 词：行人重识别 无监督学习 伪标签噪声 对比学习 聚类优化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]