吴彦丞,,
李邵梅,
高超
国家数字交换系统工程技术研究中心 ??郑州 ??450002
基金项目:国家自然科学基金(61601513)
详细信息
作者简介:陈鸿昶：男，1964年生，教授，博士生导师，研究方向为通信与信息系统、计算机视觉
吴彦丞：男，1994年生，硕士生，研究方向为计算机视觉
李邵梅：女，1982年生，博士，讲师，研究方向为通信与信息系统、计算机视觉
高超：男，1982年生，博士，讲师，研究方向为通信与信息系统、计算机视觉
通讯作者:吴彦丞 wuyc1994@163.com
中图分类号:TP391.41计量
文章访问数:3097
HTML全文浏览量:1793
PDF下载量:95
被引次数:0
出版历程
收稿日期:2018-07-20
修回日期:2019-03-03
网络出版日期:2019-04-17
刊出日期:2019-09-10
Person Re-identification Based on Attribute Hierarchy Recognition
Hongchang CHEN,Yancheng WU,,
Shaomei LI,
Chao GAO
China National Digital Switching System Engineering & Technological R&D Center, Zhengzhou 450002, China
Funds:The National Natural Science Foundation of China (61601513)
摘要
摘要:为了提高行人再识别算法的识别效果，该文提出一种基于注意力模型的行人属性分级识别神经网络模型，相对于现有算法，该模型有以下3大优点：一是在网络的特征提取部分，设计用于识别行人属性的注意力模型，提取行人属性信息和显著性程度；二是在网络的特征识别部分，针对行人属性的显著性程度和包含的信息量大小，利用注意力模型对属性进行分级识别；三是分析属性之间的相关性，根据上一级的识别结果，调整下一级的识别策略，从而提高小目标属性的识别准确率，进而提高行人再识别的准确率。实验结果表明，该文提出的模型相较于现有方法，有效提高了行人再识别的首位准确率，其中，Market1501数据集上，首位准确率达到了93.1%，在DukeMTMC数据集上，首位准确率达到了81.7%。
关键词:行人再识别/
注意力模型/
深度学习/
显著性/
属性分级
Abstract:In order to improve the accuracy rate of person re-identification, a pedestrian attribute hierarchy recognition neural network is proposed based on attention model. Compared with the existing algorithms, the model has the following three advantages. Firstly, the attention model is used in this paper to identify the pedestrian attributes, and to extract of pedestrian attribute information and degree of significance. Secondly, the attention model in used in this paper to classify the attributes according to the significance of the pedestrian attributes and the amount of informationcontained. Thirdly, this paper analyzes the correlation between attributes, and adjusts the next level identification strategy according to the recognition results of the upper level. It can improve the recognition accuracy of small target attributes, and the accuracy of pedestrian recognition is improved. The experimental results show that the proposed model can effectively improve the first accuracy rate (rank-1) of person re-identification compared with the existing methods. On the Market1501 dataset, the first accuracy rate is 93.1%, and the first accuracy rate is 81.7% on the DukeMTMC dataset.
Key words:Person re-identification/
Attention model/
Deep learning/
Saliency/
Hierarchy
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=139168b5-a745-4ffd-94cd-dd9e7efbefe7
