基于风格转换的无监督聚类行人重识别

PDF下载HTML

Title:

Clustering approach based on style transfer for unsupervised person re-identification

作者:

张智¹, 毕晓君²

1. 哈尔滨工程大学 信息与通信工程学院，黑龙江 哈尔滨 150001;
2. 中央民族大学 信息工程学院，北京 100081

Author(s):

ZHANG Zhi¹, BI Xiaojun²

1. College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China;
2. School of Information Engineering, Minzu University of China, Beijing 100081, China

关键词:

机器视觉; 行人重识别; 无监督; 聚类; 风格转换; 生成对抗网络; 残差块; 跨域

Keywords:

machine vision; pedestrian re-identification; unsupervised; clustering; style transformation; generative adversarial networks; residual block; cross domain

分类号:

TP391

DOI:

10.11992/tis.202012014

摘要:

无监督行人重识别中源域与目标域间的巨大差异性是影响模型性能的最关键因素。基于聚类的无监督行人重识别方法挖掘目标域数据间的相似性，以此缓解该问题，但仍未消除域间差异性。本文提出一种基于风格转换的无监督聚类行人重识别方法。首先，针对基于聚类方法的模型存在受域间差异性影响的问题，将一种基于生成对抗网络的风格转换方法引入到聚类方法模型中，将源域数据转换为目标域风格数据，直接减小域间差异性，提升模型的识别性能。其次，针对风格转换模型的生成器存在转换尺度单一以及特征信息传递效率低的问题，使用一种新型残差块替换原始残差块并将其引入到生成器上采样和下采样中，形成多特征尺度转换以及信息传递效率高的生成器，提升风格转换效果，降低域间差异性，进一步提升整体模型的识别效果。在Market1501以及Duke-MTMC-reID数据集上对所提的算法进行实验，结果表明改进方法取得了更好的识别效果。

Abstract:

The substantial difference between the source and target domains is the most crucial factor affecting the performance of unsupervised person re-identification models. The clustering-based unsupervised person re-identification method alleviates the problem to a certain extent by mining the similarity between the target domain, but it does not fundamentally eliminate the discrepancy between the domains. This paper proposes a clustering approach based on cross-domain style transfer for unsupervised pedestrian re-identification. First, to avoid the difference between domains in clustering-based unsupervised person re-identification models, the across-domain style transfer method based on a generative adversarial network is introduced into the clustering process. It transfers the source domain data to the target domain style data, which directly reduces the difference between domains and improves the recognition effect of the model. Second, the generator of cross-domain style transfer model has a single transfer scale and low efficiency of characteristics information transfer. A new type of residual block is proposed to replace the original residual block; then, it is inserted into the generator to achieve up-sampling and down-sampling. The specific generator has more characteristics of the scale transfer, and it transmits information more effectively. The cross-domain style transfer model can better transfer the style of the source and target domains, further reduce the difference between the two domains, and improve the recognition effect of the overall model. Extensive experiments were implemented on Market1501 and Duke-MTMC-Reid datasets to examine the proposed method, and the results showed that the proposed improved method achieved a better recognition effect.