电子学报 ›› 2019, Vol. 47 ›› Issue (2): 351-357.DOI: 10.3969/j.issn.0372-2112.2019.02.014

• 学术论文 • 上一篇    下一篇

基于多信息流动卷积神经网络的行人再识别

桑海峰1, 王传正1, 吕应宇1, 何大阔2, 刘晴2   

  1. 1. 沈阳工业大学信息科学与工程学院, 辽宁沈阳 110870;
    2. 东北大学信息科学与工程学院, 辽宁沈阳 110819
  • 收稿日期:2018-03-28 修回日期:2018-05-22 出版日期:2019-02-25
    • 通讯作者:
    • 王传正
    • 作者简介:
    • 桑海峰 男,1978年生于辽宁沈阳,博士,沈阳工业大学教授,主要研究方向为视觉检测技术与图像处理.E-mail:sanghaif@163.com
    • 基金资助:
    • 国家自然科学基金 (No.61773105); 辽宁省自然科学基金 (No.20170540675); 辽宁省教育厅科研项目 (No.LQGD2017023)

Person Re-identification Based on Multi-information Flow Convolutional Neural Network

SANG Hai-feng1, WANG Chuan-zheng1, LÜ Ying-yu1, HE Da-kuo2, LIU Qing2   

  1. 1. School of Information Science & Engineering, Shenyang University of Technology, Shenyang, Liaoning 110870, China;
    2. College of Information Science & Engineering, Northeastern University, Shenyang, Liaoning 110819, China
  • Received:2018-03-28 Revised:2018-05-22 Online:2019-02-25 Published:2019-02-25

摘要: 行人再识别问题中,由于视角、光照和行人姿态等因素的变化,导致难以提取有效的行人特征,降低识别精度.而深度神经网络在训练样本较少的情况下较难训练,易出现过拟合现象.针对上述问题,本文提出一种多信息流动卷积神经网络(Multi-information Flow Convolutional Neural Network,MiF-CNN)模型,模型中包含一个特殊的卷积结构,该结构中每层卷积层提取到的特征与后续所有卷积层的输入相连接,增强了网络的特征信息流动性和梯度的反向传播效率,使得模型提取到的行人特征更具判别力.采用多损失函数组合方式训练网络模型,更好的区分行人类别.最后利用欧氏距离对行人特征相似性进行排序.在标准行人再识别数据集VIPeR和CUHK01上的实验表明,本文方法进一步提高了行人再识别精度,并有效改善了深度神经网络的过拟合现象.

关键词: 行人再识别, 多信息流动, 特征提取, 卷积神经网络

Abstract: It is challenging to learn efficient features in person re-identification task due to complex variations of viewpoints,illumination,pose etc.In addition,deep neural network still suffers from overfitting with a small training set.To solve these problems,a Multi-information Flow Convolutional Neural Network(MiF-CNN)is designed for person re-identification which contains a specific convolutional architecture.In this architecture,features from each convolution layer are concatenated with all subsequent convolution layers so that it can improve the flowability of feature information and the efficiency of backpropagation gradient.In this way,the proposed network can learn more discriminative features.Moreover,a combination of three loss functions is used to train proposed network.Finally,the identification accuracy is obtained by ranking the similarity of extracted features measured by Euclidean distance.Experimental results on the VIPeR and CUHK01 datasets demonstrate that MiF-CNN outperform most of existing methods of person re-identification and reduce overfitting more effectively.

Key words: person re-identification, multi-information flow, feature extraction, convolutional neural network

中图分类号: