• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 图形与图像 • 上一篇    下一篇

基于CNN的改进行人重识别技术

熊炜1,2,冯川1,熊子婕1,王娟1,2,刘敏1,2,曾春艳1,2   

  1. (1.湖北工业大学电气与电子工程学院,湖北 武汉 430068;
    2.湖北工业大学太阳能高效利用湖北省协同创新中心,湖北 武汉 430068)
  • 收稿日期:2018-06-05 修回日期:2018-09-06 出版日期:2019-04-25 发布日期:2019-04-25
  • 基金资助:

    国家自然科学基金(61501178,61571182,61601177);湖北省教育厅科学技术研究计划重点项目(D20161404);太阳能高效利用湖北省协同创新中心开放基金(HBSKFZD201411)

Improved pedestrian re-identification based on CNN

XIONG Wei1,2,FENG Chuan1,XIONG Zijie1,WANG Juan1,2,LIU Min1,2,ZENG Chunyan1,2   

  1. (1.School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068;
    2.Hubei Collaborative Innovation Center for HighEfficiency Utilization of Solar Energy,
    Hubei University of Technology,Wuhan 430068,China)
     
  • Received:2018-06-05 Revised:2018-09-06 Online:2019-04-25 Published:2019-04-25

摘要:

针对行人重识别研究中训练样本的不足,为提高识别精度及泛化能力,提出一种基于卷积神经网络的改进行人重识别方法。首先对训练数据集进行扩充,使用生成对抗网络无监督学习方法生成无标签图像;然后与原数据集联合作半监督卷积神经网络训练,通过构建一个Siamese网络,结合分类模型和验证模型的特点进行训练;最后加入无标签图像类别分布方法,计算交叉熵损失来进行相似度量。实验结果表明,在Market-1501、CUHK03和DukeMTMCreID数据集上,该方法相比原有的Siamese方法在Rank-1和mAP等性能指标上有近3~5个百分点的提升。当样本较少时,该方法具有一定应用价值。
 

关键词: 行人重识别, 卷积神经网络, 生成对抗网络, 交叉熵, Siamese

Abstract:

For the lack of training samples in pedestrian reidentification (reID) research, we propose a pedestrian re-ID method based on convolutional neural network (CNN) to improve the recognition accuracy and generalization ability. Firstly, we employ the unsupervised learning method for the generative adversarial network  to generate unlabeled images, so the training data set is expanded. Secondly, the original data set is collaborated to perform semi-supervised CNN training, and a Siamese network is constructed to perform training according to the features of the identification model and the verification model. Finally, the unlabeled image category distribution method is introduced, and the cross entropy loss is calculated to perform similarity measurement. Experiments on the Market1501, CUHK03, and DukeMTMC-reID datasets show that the proposed method has a nearly 3% to 5% improvement in performance indicators such as rank1 and mAP in comparison with the original Siamese method. The proposed method has certain application value in small sample scenarios.pedestrian re-identification;convolutional neural network (CNN);generative adversarial network (GAN);cross entropy;Siamese

Key words: pedestrian re-identification, convolutional neural network (CNN), generative adversarial network (GAN), cross entropy, Siamese