基于多尺度多粒度融合的行人重识别方法  

Person Re-Identification Method Based on Multi-Scale and Multi-Granularity Fusion

作　　者：符进武 范自柱 石林瑞 郭心悦 黄祎婧 FU Jinwu;FAN Zizhu;SHI Linrui;GUO Xinyue;HUANG Yijing(School of Science,East China Jiaotong University,Nanchang 330013,China)

机构地区：[1]华东交通大学理学院,南昌330013

出　　处：《计算机工程》2022年第3期271-279,共9页Computer Engineering

基　　金：国家自然科学基金(61991401,61673097,61702117);江西省自然科学基金重点项目(20192ACBL20010)。

摘　　要：行人重识别是指利用计算机视觉技术在给定监控的图像中识别目标行人,受拍摄场景视角和姿势变化、遮挡等因素的影响,现有基于局部特征的行人重识别方法所提取的特征辨别力差,从而导致重识别精度较低。为有效地利用特征信息,提出一种多尺度多粒度融合的行人重识别方法 MMF-Net。通过多个分支结构学习不同尺度和不同粒度的特征,并利用局部特征学习优化全局特征,以加强全局特征和局部特征的关联性。同时,在网络的低层引入语义监督模块以提取低层特征,并将其作为行人图像相似性度量的补充,实现低层特征和高层特征的优势互补。基于改进的池化层,通过结合最大池化和平均池化的特点获取具有强辨别力的特征。实验结果表明,MMFNet方法在Market-1501数据集上的首位命中率和mAP分别为95.7%和89.1%,相比FPR、MGN、BDB等方法,其具有较优的鲁棒性。Person re-identification refers to the use of computer vision technology to recognize the target person in a given monitored image.Factors,such as the change of shooting scene angle,posture,and occlusion,among others,result in poor feature discrimination extraction with existing person re-identification methods based on local features,associating these methods with low re-identification accuracy. To make effective use of feature information,this paper proposes a person re-identification method,MMF-Net,based on multi-scale and multi-granularity fusion,whereby features with different scales and granularity are learned by multiple branch structures.Local feature learning is used to optimize global features,thereby enhancing the correlation between global features and local features. A semantic supervision module is also introduced into the lower layer of the network to extract low-level features,which are used as a supplement to person image similarity measurement,to benefit from the complementary advantages of low-level and high-level features.On the improved pooling layer,the features with strong discrimination are obtained by combining the features of maximum and average pooling.The experimental results show that the Rank-1 and mAP of MMF-Net method is 95.7% and 89.1%,respectively.Compared with FPR,MGN,BDB method,it has better robustness.

关 键 词：行人重识别 特征学习 局部特征 低层特征 池化 多尺度多粒度融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]