欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2022, Vol. 43 ›› Issue (6): 1193-1200.DOI: 10.11996/JG.j.2095-302X.2022061193

• 图像处理与计算机视觉 • 上一篇    下一篇

基于视觉信息积累的行人重识别网络

  

  1. 北京工业大学人工智能与自动化学院，北京 100124
  • 出版日期:2022-12-30 发布日期:2023-01-11
  • 基金资助:
    第7批全国博士后创新人才支持计划项目(BX20220025)；第70批全国博士后面上基金项目(2021M700303) 

Visual information accumulation network for person re-identification

  1. School of Artificial Intelligence and Automation, Beijing University of Technology, Beijing 100124, China
  • Online:2022-12-30 Published:2023-01-11
  • Supported by:
    The 7th National Postdoctoral Innovative Talent Support Program (BX20220025); The 70th Batch of National Post-Doctoral Fellowships (2021M700303)

PDF (PC)

74

摘要：

在以往的行人重识别方法中，绝大部分的工作集中于图像注意力区域的学习，却忽视了非注意 力区域对最终特征学习的影响，如果在关注图像注意力区域的同时加强非注意力区域的特征学习，可进一步丰 富最终的行人特征，有利于行人身份信息的准确识别。基于此，提出了视觉信息积累网络(VIA Net)，该网络整 体采用两分支结构，一个分支倾向于学习图像的全局特征，另一个分支则拓展为多分支结构，通过结合注意力 区域和非注意力区域的特征逐步加强局部特征的学习，实现视觉信息的积累，进一步丰富特征信息。实验结果 表明，在 Market-1501 等行人重识别数据集上，所提出的 VIA Net 网络达到了较高的实验性能；同时，在 In-Shop Clothes Retrieval 数据集上的实验证明：该网络也适用于一般的图像检索任务，具有一定的通用性。

关键词: 行人重识别, 视觉信息, 注意力区域, 非注意力区域, 度量学习

Abstract:

The preceding person re-identification methods were mostly focused on the learning of the image attention region, but ignored the impact of the non-attention region on the final feature learning. If the feature learning of image non-attention regions is enhanced while focusing on attention regions, the final person features can be further enriched, which is beneficial to the accurate identification of person identity information. Based on this, this paper proposed a visual information accumulation network (VIA Net), adopting two branches. One branch tended to learn the global features of the image, and the other branch was expanded into a multi-branch structure. By combining the features of the attention and non-attention regions, the learning of local features could be gradually strengthened, thus realizing the accumulation of visual information and further enriching the feature information. The experimental results show that the proposed VIA Net could attain high experimental performance in terms of person re-identification datasets such as Market-1501. At the same time, the experiment on the In-Shop Clothes Retrieval dataset shows that the network could also be applicable to general image retrieval tasks and possess certain universality. 

Key words:  , person re-identification, visual information, attention region, non-attention region, metric learning

中图分类号: