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 ›› 2022, Vol. 48 ›› Issue (4): 269-275,283. doi: 10.19678/j.issn.1000-3428.0060811

• 开发研究与工程应用 • 上一篇    下一篇

行人再识别中基于无监督学习的粗细粒度特征提取

唐佳敏, 韩华, 黄丽   

  1. 上海工程技术大学 电子电气工程学院, 上海 201620
  • 收稿日期:2021-02-04 修回日期:2021-03-16 发布日期:2022-04-14
  • 作者简介:唐佳敏(1995—),女,硕士研究生,主研方向为目标识别与跟踪、行人再识别;韩华(通信作者),副教授、博士;黄丽,讲师、博士。
  • 基金资助:
    国家自然科学基金(61305014);上海市教育委员会和上海市教育发展基金会“晨光计划”(13CG60)。

Coarse-grained and Fine-grained Features Extraction Based on Unsupervised Learning in Pedestrian Re-identification

TANG Jiamin, HAN Hua, HUANG Li   

  1. School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China
  • Received:2021-02-04 Revised:2021-03-16 Published:2022-04-14

摘要: 行人再识别研究中存在特征判别信息不够丰富的情况,并且遮挡、光照等因素会干扰有效特征的准确提取,对后续相似性度量、度量结果排序等工作都有较大影响。此外,监督学习需要使用标签信息,在面对大型数据集时工作量很大。通过引入无监督学习框架,提出一种粗细粒度判别性特征提取方法。构建基于细粒度和粗粒度特征学习的模型框架,其中包含局部和全局2个分支。在局部分支中,对图像学习到的特征映射提取补丁块,并在未标记数据集上学习不同位置的细粒度补丁特征;在全局分支中,使用无标注数据集的相似度和多样性作为信息来学习粗粒度特征。在此基础上,利用相吸和相斥2个损失函数分别增加类别内相似度和类别间多样性,并结合最小距离准则计算特征之间的相似度,进行无监督的聚类合并。在Market-1501和DukeMTMC-reID数据集上的实验结果表明,该方法对于完成行人再识别任务具有较好的判别性能和鲁棒性,相比所有对比方法的最优结果,其Rank-1指标分别提高5.76%和5.07%,平均精度均值分别提高3.2%和5.6%。

关键词: 计算机视觉, 行人再识别, 无监督学习, 特征学习, 损失函数, 最小距离准则

Abstract: In the research of pedestrian re-identification, there is not enough feature discrimination information, and factors such as occlusion and illumination will interfere with the accurate extraction of effective features, having a decisive impact on the subsequent similarity measurement and ranking of measurement results.In addition, supervised learning models need label information, imposing a heavy workload for large datasets.In view of this, a coarse-grained and fine-grained discriminant feature extraction method is proposed based on unsupervised learning.A model framework is built based on fine-grained and coarse-grained feature learning for local and global feature extraction.For local features, patch blocks are extracted from the feature map learned from the image, and fine-grained patch features at different locations are learned from the unlabeled data set.For global features, the similarity and diversity of unlabeled data sets provide the information for coarse-grained feature learning.On this basis, the two loss functions of attraction and repulsion are used to increase the similarity within categories and the diversity between categories respectively, and the similarity between features is calculated combined with the minimum distance criterion for unsupervised clustering.The experimental results on Market-1501 and DukeMTMC-reID datasets show that this method has good discrimination performance and robustness for solving the pedestrian re-indentification task, whereby the Rank-1 index is improved by 5.76% and 5.07%, respectively, and the mean Average Precision(mAP) is improved by 3.2% and 5.6%, respectively compared with the optimal results of all comparison methods.

Key words: computer vision, pedestrian re-identification, unsupervised learning, feature learning, loss function, minimum distance criterion

中图分类号: