行人重识别中基于多分支的鲁棒性特征挖掘网络设计  

Robust Feature Mining Network Design Based on Multi-Branch in Person Re-Identification

作　　者：刘润 汪淑娟[1] LIU Run;WANG Shujuan(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500

出　　处：《电视技术》2022年第5期61-66,共6页Video Engineering

摘　　要：目前人员再识别方法大多都是使用深度卷积神经网络(Convolution Neural Network,CNN)作为骨干网络,尽管深度卷积网络在分类或对象检测等任务中都很有效,但它们往往只关注对象最具鉴别性的部分,并没有检索所有相关特征。在没有提取到对象所有特征的情况下,会导致深度卷积神经网络在重识别任务中的性能表现不佳。为了让网络可以挖掘更具鲁棒性的特征,同时在执行行人重识别任务时可以有效地应对行人图片中更精细的变化,设计了一个多分支的鲁棒性特征挖掘网络。该网络结合了擦除操作与注意力机制,在不同数据集上均有较好的性能。Most of the current person re-identification methods use a deep Convolutional Neural Network(CNN)as the backbone network.Although deep convolutional networks are effective in tasks such as classification or object detection,they often only focus on the most important objects.The discriminative part does not retrieve all relevant features.In the case of not extracting all the features of the object,the performance of deep convolutional neural networks in re-identification tasks will be poor.In order to allow the network to mine more robust features and to effectively deal with finer changes in pedestrian images when performing person re-identification tasks,a multi-branch robust feature mining network is designed.The network combines erasure operation and attention mechanism,and has good performance on different datasets.

关 键 词：行人重识别 鲁棒性特征 深度卷积神经网络 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]