作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2021, Vol. 47 ›› Issue (4): 226-233,240. doi: 10.19678/j.issn.1000-3428.0057115

• 图形图像处理 • 上一篇    下一篇

融合RGB与灰度图像特征的行人再识别方法

姜国权, 肖禛禛, 霍占强   

  1. 河南理工大学 计算机科学与技术学院, 河南 焦作 454000
  • 收稿日期:2020-01-06 修回日期:2020-02-25 发布日期:2020-03-05
  • 作者简介:姜国权(1969-),男,副教授、博士,主研方向为图像处理、模式识别;肖禛禛,硕士;霍占强(通信作者),副教授、博士。
  • 基金资助:
    国家自然科学基金(61572173);河南省高校科技创新团队支持计划(19IRTSTHN012)。

Pedestrian Re-Identification Method Combining RGB and Grayscale Image Features

JIANG Guoquan, XIAO Zhenzhen, HUO Zhanqiang   

  1. School of Computer Science and Technology, Henan Polytechnic University, Jiaozuo, Henan 454000, China
  • Received:2020-01-06 Revised:2020-02-25 Published:2020-03-05

摘要: 针对行人再识别过程中相同身份行人图像颜色不一致,以及不同身份行人图像颜色相近问题,提出一种基于双分支残差网络的行人再识别方法。将RGB图像和灰度图像分别输入预训练的ResNet-50网络,获得RGB图像特征和灰度图像特征并对其进行融合,利用统一水平划分策略学习融合特征,同时将RGB特征、灰度特征和融合特征的拼接结果作为最终特征表示。在Market1501、DukeMTMC-ReID和CUHK03数据集上的实验结果表明,与PCB、Mancs等行人再识别方法相比,该方法的平均精度均值和首位命中率更高,且对于图像颜色变化具有更强的鲁棒性。

关键词: 深度学习, 行人再识别, RGB图像特征, 灰度图像特征, 融合特征

Abstract: To address the problems in pedestrain Re-Identification(ReID),including color inconsistency of pedestrain images with the same identity and color similarity between pedestrain images with different identities,this paper proposes a pedestrain ReID method based on double branch residual network.The RGB images and grayscale images are input into the pre-trained ResNet-50 separately to obtain RGB and grayscale features that are subsequently fused.Then,the fusion features are learned by using a unified horizontal partition strategy.Finally,the RGB,grayscale and fusion features are concatenated to act as the final feature representation.Experimental results on Market1501,DukeMTMC-ReID and CUHK03 datasets show that the proposed method has mean Average Precision(mAP) and Rank-1 accuracy than PCB,Mancs and other pedestrian ReID methods,and has stronger robustness to image color changes.

Key words: deep learning, pedestrian Re-Identification(ReID), RGB image feature, grayscale image feature, fusion feature

中图分类号: