基于全局滤波池化多关系Transformer网络的行人重识别  

Person Re-identification Based on Global Filter Pooling Multi Relation Transformer Network

作　　者：焦传扬 丁学明[1] JIAO Chuanyang;DING Xueming(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《控制工程》2024年第5期912-919,共8页Control Engineering of China

基　　金：国家高技术研究发展计划资助项目(61673277)。

摘　　要：行人重识别(Re-identification,ReID)的关键挑战之一是提取关键且鲁棒的特征,近年来,Transformer网络不断展现其在该问题上具有强大的特征提取和表达能力。针对传统Transformer网络局部信息获取不如卷积神经网络的问题,提出一个基于ReID的全局滤波池化多关系Transformer(Trans-global filter pooling multi relationship-ReID,TFMR)网络新型框架,解决了Transformer网络局部关系建模不够丰富的问题。多关系(multi relation,MR)网络考虑身体多个部位间的关系,使特征包含局部信息之间的联系,增强特征中行人生理结构的关联。同时设计了全局滤波池化(global filter pooling,GFP)模块,将其嵌入到Transformer网络中,降低图片中噪点的干扰并减少视图变化造成的特征偏差,从而获取人物图像中更清晰的全局特征,提升识别准确率。实验表明,模型在区分行人信息问题上具有高效性,在Market-1501、DukeMTMC-ReID和MSMT17数据集中优于其他模型。One of the key challenges of person re-identification is to extract key and robust features.In recent years,Transformer network continues to show its strong ability of feature extraction and expression.For the problem that the local information acquisition of traditional Transformer network is not as good as convolutional neural network,a new framework of global filter pooling multi relational network(TFMR)is proposed,which solves the problem that the local relational modeling of Transformer network is not rich enough.Multi relational(MR)network considers the relationship between multiple parts of the body,makes the features contain the connection between local information,and enhances the correlation of pedestrian physiological structure in the features.At the same time,the global filter pooling(GFP)module is designed and embedded into the Transformer network to reduce the interference of noise in the picture and the deviation of features caused by view changes,so as to obtain clearer global features in the character image and improve the recognition accuracy.Experiments show that this model is efficient in distinguishing pedestrian information,and is superior to other models in Market-1501,DukeMTMC-ReID and MSMT17 datasets.

关 键 词：行人重识别 TRANSFORMER 全局滤波池化 多关系网络 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]