多层次细粒度特征三分支网络行人重识别方法  

Person Re-identification Method Based on Three-branch Network with Multi-level Fine-grained Feature

作　　者：贺南南 张荣国[1] 胡静[1] 李建伟[1] 李晓波 HE Nan-nan;ZHANG Rong-guo;HU jing;LI Jian-wei;LI Xiao-bo(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China)

机构地区：[1]太原科技大学计算机科学与技术学院,太原030024

出　　处：《太原科技大学学报》2021年第5期348-354,共7页Journal of Taiyuan University of Science and Technology

基　　金：国家自然基金(51375132);山西省自然科学基金(201801D121134);太原科技大学博士科研启动基金(20202057)。

摘　　要：针对行人重识别中信息丢失导致判别性信息缺失的问题,提出了一种多层次细粒度特征三分支网络行人重识别方法。首先,在ResNet50网络上构建中层全局特征分支、多层次全局特征分支和局部特征分支,全局分支提供更加全面的特征表示,局部特征分支提供细粒度的特征表示;其次,在三分支网络上改进了损失函数,使用权重向量和特征向量归一化以消除向量模的影响,通过构建难样本三元组损失以解决类间相似、类内差异分类问题;最后,在Market-1501和DukeMTMC-reid两个数据集上进行实验,rank-1达到了94.0%和87.4%,mAP达到了85.7%和75.5%.和现有的八种方法进行对比实验,结果表明本文方法在行人重识别中具有更好的准确率和精度。In order to solve the problem of lack of discriminative information caused by information loss in person re-identification,a person re-identification method on three-branch network with multi-level fine-grained feature is proposed.Firstly,the middle-level global feature branch,multi-level global feature branch and local feature branch are constructed on resnet50 network.The global branch provides more comprehensive feature representation,and the local feature branch provides fine-grained feature representation.Secondly,the loss function is improved on the three-branch network using weight vector and eigenvector normalization to eliminate the influence of vector norm.Finally,the experiments on market-1501 and dukemtmc Reid data sets show that rank-1 reaches 94.0%and 87.4%,and map reaches 85.7%and 75.5%.Compared with the existing eight methods,the results show that the proposed method has better accuracy and accuracy in pedestrian recognition.

关 键 词：行人重识别 多层次细粒度 特征提取 深度学习 分支网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]