计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (20): 157-164.DOI: 10.3778/j.issn.1002-8331.2103-0311

• 模式识别与人工智能 • 上一篇    下一篇

融合注意力机制与权重聚类学习的行人再识别

孙姣,杨有龙,车金星   

  1. 1.西安电子科技大学 数学与统计学院,西安 710126
    2.南昌工程学院 理学院,南昌 330099
  • 出版日期:2022-10-15 发布日期:2022-10-15

Person Re-Identification Combining Attention Mechanism and Weight Clustering Learning

SUN Jiao, YANG Youlong, CHE Jinxing   

  1. 1.School of Mathematics and Statistics, Xidian University, Xi’an 710126, China
    2.School of Science, Nanchang Institute of Technology, Nanchang 330099, China
  • Online:2022-10-15 Published:2022-10-15

摘要: 行人图像在行人再识别中常通过行人检测器自动检测获得,不仅包含行人主体,还包含一些干扰信息(比如,背景、遮挡等)。在基于注意力机制的行人再识别中,增强了对具有显著性特征行人部件的关注,削弱了对带有干扰信息部件的关注,有利于提取更具辨别力的行人特征表示。在深度学习中,卷积神经网络通过对特征映射重新赋权值,得到注意力特征,提出了一种新颖的基于聚类的全局注意力模块(cluster-based global attention module,CGAM)。在CGAM中,将注意力权重学习过程重新考虑为聚类中心学习过程,将特征映射中的空间位置点视为特征节点,通过聚类算法得到每个特征节点的重要分数并进行归一化后作为注意力权重。利用改进的Resnet50作为基本框架,嵌入注意力模块,得到注意力网络,仅使用了全局分支,具有简单高效特点。综上,基于聚类的注意力设计不仅充分利用了特征节点之间的成对相关性,而且挖掘了丰富的全局结构信息,得到一组更可信的注意力权重。实验结果表明,提出的行人再识别算法在Market-1501和DukeMTMC-reID两个流行数据集上均有显著的效果。

关键词: 行人再识别, 深度学习, 注意力网络, 注意力权重, 聚类算法

Abstract: Pedestrian images in person re-identification(re-ID) are often obtained by automatic detection of pedestrian detectors, which contain the human body and some interference information (e.g. background, occlusion, etc.). For re-ID based on the attention mechanism, the attention to the person parts with salient features is increased, and the attention to the parts with interference information is weakened, which is helpful to extracting more discriminative person feature representations. In deep learning, by re-weighting the feature map to obtain attentional feature, a novel cluster-based global attention module(CGAM) is proposed, the attention weight learning process is reconsidered as the clustering center learning process, the spatial location points in the feature map are regarded as feature nodes, and the important score of each feature node is obtained through the clustering algorithm and normalized as the attention weight. It uses the improved Resnet50 as the backbone network and embeds the attention module to get the attention network. Only the global branch is used, which is simple and efficient. In summary, the cluster-based attention design not only makes full use of the pairwise correlation between feature nodes, but also mines rich global structural information to obtain a set of more credible attention weights. The experimental results show that the proposed method has significant effects on the two popular datasets Market-1501 and DukeMTMC-reID.

Key words: person re-identification, deep learning, attentional network, attentional weight, clustering algorithm