Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (19): 211-219.DOI: 10.3778/j.issn.1002-8331.2205-0475

• Graphics and Image Processing • Previous Articles     Next Articles

Person Re-Identification Driven by Diverse Fine-Grained Features and Relation Network

XU Ruyu, WU Lin, SU Xingwang, HUANG Jinbo, WANG Xiaoming   

  1. School of Computer and Software Engineering, Xihua University, Chengdu 610039, China
  • Online:2023-10-01 Published:2023-10-01

多样细粒度特征与关系网络驱动的行人重识别

许茹玉,吴琳,粟兴旺,黄金玻,王晓明   

  1. 西华大学 计算机与软件工程学院,成都 610039

Abstract: In image person re-identification, the traditional method based on deep network usually adopts simple local blocking method to extract features. This method ignores the correlation between features and restricts the improvement of network identification ability. Aiming at this problem, a network model(DFFRRID) containing diverse features is proposed. The network model is designed with three modules behind the backbone network, which are used to extract rough global features, local features segmented from coarse to fine-grained, and correlation features between local features. The three modules complement each other and finally integrate the extracted global and local features into the classification network. In the loss optimization stage, the combined loss of cross entropy loss and label smoothing cross entropy loss is used for classification learning to prevent network overfitting. Experimental results on Market1501, DukeMTMC-ReID, and CUHK03 datasets verify the effectiveness of the proposed model, The accuracy of Rank-1 of reached 95.3%, 89.3% and 80.5%, respectively, the mAP accuracy are 88.6%, 78.9% and 73.9%. which further improves the accuracy of image pedestrian re-recognition compared with most methods.

Key words: person re-identification, deep learning, fine-grained division, local features, relation network

摘要: 在图像行人重识别(person re-identification)中,基于深度网络的传统方法通常采用了简单局部切块方式进行特征提取。该方式忽略了特征间的关联性,限制了网络鉴别能力的提高。针对该问题,提出了一种包含多样化特征的网络模型(DFFRRID)。该网络模型在主干网络后设计了三个模块,分别用于提取粗糙的全局特征、由粗到细粒度切分的局部特征以及局部特征间的关联性特征。三个模块相互补充,最后将提取的全局与局部特征集成到分类网络。在损失优化阶段,利用交叉熵损失与标签平滑交叉熵损失的联合损失进行分类学习,以防止网络过拟合。在Market1501、DukeMTMC-ReID以及CUHK03数据集上的实验结果验证了所提模型的有效性,DFFRRID的Rank-1准确率分别达到了95.3%、89.3%和80.5%,mAP精度分别为88.6%、78.9%和73.9%,其相对于大多数方法进一步提高了图像行人重识别精度。

关键词: 行人重识别, 深度学习, 细粒度划分, 局部特征, 关系网络