通信学报 ›› 2022, Vol. 43 ›› Issue (7): 153-162.doi: 10.11959/j.issn.1000-436x.2022136

• 学术论文 • 上一篇    下一篇

基于语义融合的域内相似性分组行人重识别

寇旗旗1, 黄绩2, 程德强2, 李云龙2, 张剑英2   

  1. 1 中国矿业大学计算机科学与技术学院,江苏 徐州 221116
    2 中国矿业大学信息与控制工程学院,江苏 徐州 221116
  • 修回日期:2022-06-15 出版日期:2022-07-25 发布日期:2022-06-01
  • 作者简介:寇旗旗(1988- ),男,河南襄城人,博士,中国矿业大学讲师,主要研究方向为智能信息处理、计算机视觉、模式识别
    黄绩(1995-),男,山西忻州人,中国矿业大学硕士生,主要研究方向为无监督行人重识别
    程德强(1979- ),男,河南洛阳人,博士,中国矿业大学教授、博士生导师,主要研究方向为机器视觉与模式识别、图像处理与视频编码、图像智能检测与信息处理
    李云龙(1997- ),男,河南开封人,中国矿业大学硕士生,主要研究方向为无监督行人重识别
    张剑英(1964- ),女,江苏徐州人,博士,中国矿业大学教授,主要研究方向为信息处理、电磁场理论及应用等
  • 基金资助:
    中央高校基本科研业务费专项资金资助项目(2020QN49)

Person re-identification with intra-domain similarity grouping based on semantic fusion

Qiqi KOU1, Ji HUANG2, Deqiang CHENG2, Yunlong LI2, Jianying ZHANG2   

  1. 1 School of Computer Science and Technology, China University of Mining and Technology, Xuzhou 221116, China
    2 School of Information and Control Engineering, China University of Mining and Technology, Xuzhou 221116, China
  • Revised:2022-06-15 Online:2022-07-25 Published:2022-06-01
  • Supported by:
    The Fundamental Research Funds for the Central Universities(2020QN49)

摘要:

无监督跨域行人重识别旨在使有标签源域数据集上训练的模型适应目标域数据集。然而,基于聚类的无监督跨域行人重识别算法在网络特征学习过程中常因输入行人图片情况各异而产生噪声,从而影响聚类效果。针对这一问题,提出一种基于语义融合的域内相似性分组行人重识别网络,首先在Baseline网络的基础上添加语义融合层,依次从空间和通道2个方面对中间特征图进行相似特征的语义融合,从而提升网络的自适应感知能力。此外,通过充分利用域内相似性细粒度信息,进而提高网络对全局和局部特征的聚类精准度。通过在DukeMTMC-ReID、Market1501和MSMT17这3个公开数据集上进行实验,结果表明,所提算法的均值平均精度(mAP)和Rank识别准确率与近年无监督跨域行人重识别算法相比有显著提升。

关键词: 无监督跨域, 行人重识别, 语义融合, 自适应感知, 细粒度信息

Abstract:

Unsupervised cross-domain person re-identification aims to adapt a model trained on a labeled source-domain dataset to a target-domain dataset.However, the cluster-based unsupervised cross-domain pedestrian re-identification algorithm often generates noise due to the different input pedestrian pictures during the network feature learning process, which affects the clustering results.To solve this problem, An intra-domain similarity grouping pedestrian re-identification network based on semantic fusion was proposed.Firstly, a semantic fusion layer was added on the basis of the Baseline network, and the semantic fusion of similar features was performed on the intermediate feature maps from the two aspects of space and channel in turn, so as to improve the adaptive perception ability of the network.In addition, by making full use of the fine-grained information of intra-domain similarity, the network’s clustering accuracy of global and local features was improved.Experiments were carried out on three public datasets, DukeMTMC-ReID, Market1501, MSMT17, and the results demonstrate that the mAP and Rank recognition accuracy are significantly improved compared with recent unsupervised cross-domain person re-identification algorithms.

Key words: unsupervised cross domain, person re-identification, semantic fusion, adaptive perception, fine-grained information

中图分类号: 

No Suggested Reading articles found!