《计算机应用》唯一官方网站 ›› 2022, Vol. 42 ›› Issue (11): 3573-3579.DOI: 10.11772/j.issn.1001-9081.2021122124

• 第二十一届中国虚拟现实大会 • 上一篇    

基于多粒度特征生成对抗网络的跨分辨率行人重识别

耿艳兵, 廉永健()   

  1. 中北大学 大数据学院,太原 030051
  • 收稿日期:2021-12-17 修回日期:2022-02-28 接受日期:2022-03-07 发布日期:2022-05-17 出版日期:2022-11-10
  • 通讯作者: 廉永健
  • 作者简介:耿艳兵(1980—),女,河南漯河人,讲师,博士,CCF会员,主要研究方向:图像处理、模式识别、人工智能
    廉永健(1975—),男,山西平遥人,讲师,博士,主要研究方向:图像处理、模式识别、人工智能、虚拟现实。lyj@nuc.edu.cn
  • 基金资助:
    山西省自然科学基金资助项目(201901D111154)

Cross‑resolution person re‑identification by generative adversarial network based on multi‑granularity features

Yanbing GENG, Yongjian LIAN()   

  1. School of Data Science and Technology,North University of China,Taiyuan Shanxi 030051,China
  • Received:2021-12-17 Revised:2022-02-28 Accepted:2022-03-07 Online:2022-05-17 Published:2022-11-10
  • Contact: Yongjian LIAN
  • About author:GENG Yanbing, born in 1980, Ph. D., lecturer. Her research interests include image processing, pattern recognition, artificial intelligence.
    LIAN Yongjian, born in 1975, Ph. D., lecturer. His research interests include image processing, pattern recognition, artificial intelligence, virtual reality.
  • Supported by:
    Natural Science Foundation of Shanxi Province(201901D111154)

摘要:

现有基于生成对抗网络(GAN)的超分辨率(SR)重建方法用于跨分辨率行人重识别(ReID)时,重建图像在纹理结构内容的恢复和特征一致性保持方面均存在不足。针对上述问题,提出基于多粒度信息生成网络的跨分辨率行人ReID方法。首先,在生成器的多层网络上均引入自注意力机制,聚焦多粒度稳定的结构关联区域,重点恢复低分辨率(LR)行人图像的纹理结构信息;同时,在生成器后增加一个识别器,在训练过程中最小化生成图像与真实图像在不同粒度特征上的损失,提升生成图像与真实图像在特征上的一致性。然后,联合自注意力生成器和识别器,与判别器交替优化,在内容和特征上改进生成图像。最后,联合改进的GAN和行人ReID网络交替训练优化网络的模型参数,直至模型收敛。在多个跨分辨率行人数据集上的实验结果表明,所提算法的累计匹配曲线(CMC)在其首选识别率(rank?1)上的准确率较现有同类算法平均提升10个百分点,在提升SR图像内容一致性和特征表达一致性方面均表现更优。

关键词: 跨分辨率, 行人重识别, 生成对抗网络, 自注意力机制, 多粒度特征

Abstract:

Existing Super Resolution (SR) reconstruction methods based on Generative Adversarial Network (GAN) for cross?resolution person Re?IDentification (ReID) suffer from deficiencies in both texture structure content recovery and feature consistency maintenance of the reconstructed images. To solve these problems, a cross?resolution pedestrian re?identification method based on multi?granularity information generation network was proposed. Firstly, a self?attention mechanism was introduced into multiple layers of generator to focus on multi?granularity stable regions with structural correlation, focusing on recovering the texture and structure information of the Low Resolution (LR) person image. At the same time, an identifier was added at the end of the generator to minimize the loss in different granularity features between the generated image and the real image during the training process, improving the feature consistency between the generated image and the real image in terms of features. Secondly, the self?attention generator and identifier were jointed, then they were optimized alternately with the discriminator to improve the generated image on content and features. Finally, the improved GAN and person re?identification network were combined to train the model parameters of the optimized network alternately until the model converged. Comparison Experimental results on several cross?resolution person re?identification datasets show that the proposed algorithm improves rank?1 accuracy on Cumulative Match Characteristic(CMC) by 10 percentage points on average, and has better performance in enhancing both content consistency and feature expression consistency of SR images.

Key words: cross?resolution, person Re?IDentification (ReID), Generative Adversarial Network (GAN), self?attention mechanism, multi?granularity feature

中图分类号: