基于语义伪标签和双重特征存储库的无监督跨模态行人重识别  

Unsupervised Cross-Modality Person Re-identification Based on Semantic Pseudo-Label and Dual Feature Memory Banks

作　　者：孙锐[1,2] 余益衡 张磊[1,2] 张旭东[1] SUN Rui;YU Yiheng;ZHANG Lei;ZHANG Xudong(School of Computer Science and Information Engineering,Hefei University of Technology,Hefei 230601;Anhui Key Laboratory of Industry Safety and Emergency Technology,Hefei University of Technology,Hefei 230009)

机构地区：[1]合肥工业大学计算机与信息学院,合肥230601 [2]合肥工业大学工业安全与应急技术安徽省重点实验室,合肥230009

出　　处：《模式识别与人工智能》2022年第10期904-914,共11页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金面上项目(No.61876057);安徽省自然科学基金项目(No.2208085MF158);安徽省重点研发计划-科技强警专项项目(No.202004d07020012)。

摘　　要：现有的有监督可见光-近红外行人重识别方法需要大量人力资源去除手工标注数据,容易受到标注数据场景的限制,难以满足真实多变应用场景的泛化性。因此,文中提出基于语义伪标签和双重特征存储库的无监督跨模态行人重识别方法。首先,提出基于对比学习框架的预训练方法,利用可见光行人图像和其生成的辅助灰度图像进行训练。利用该预训练方法获取对颜色变化具有鲁棒性的语义特征提取网络。然后,使用DBSCAN(Density-Based Spatial Clustering of Applications with Noise)聚类方法生成语义伪标签。相比现有的伪标签生成方法,文中提出的语义伪标签在生成过程中充分利用跨模态数据之间的结构信息,减少跨模态数据颜色变化带来的模态差异。此外,文中还构建实例级困难样本特征存储库和中心级聚类特征存储库,充分利用困难样本特征和聚类特征,让模型对噪声伪标签具有更强的鲁棒性。在SYSU-MM01、RegDB两个跨模态数据集上的实验验证文中方法的有效性。The existing supervised visible infrared person re-identification methods require a lot of human resources to manually label the data and they fail to adapt to the generalization of real and changeable application scenes due to the limitation by the labeled data scene.In this paper,an unsupervised cross-modality person re-identification method based on semantic pseudo-label and dual feature memory banks is proposed.Firstly,a pre-training method based on the contrast learning framework is proposed,using the visible image and its generated auxiliary gray image for training.The pre-training method is employed to obtain the semantic feature extraction network that is robust to color changes.Then,semantic pseudo-label is generated by density based spatial clustering of applications with noise(DBSCAN)clustering method.Compared with the existing pseudo-label generation methods,the proposed method makes full use of the structural information between the cross-modality data in the generation process,and thus the modality discrepancy caused by the color change of the cross-modality data is reduced.In addition,an instance-level hard sample feature memory bank and a centroid-level clustering feature memory bank are constructed to make the model more robust to noise pseudo-label by hard sample features and clustering features.Experimental results obtained on two cross-modality datasets,SYSU-MM01 and RegDB,demonstrate the effectiveness of the proposed method.

关 键 词：无监督跨模态行人重识别 语义伪标签 双重特征存储库 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]