计算机应用 ›› 2019, Vol. 39 ›› Issue (4): 977-983.DOI: 10.11772/j.issn.1001-9081.2018091889

• 人工智能 • 上一篇    下一篇

基于孪生网络和双向最大边界排序损失的行人再识别

祁子梁1, 曲寒冰1,2, 赵传虎1, 董良2, 李博昭2, 王长生2   

  1. 1. 河北工业大学 人工智能与数据科学学院, 天津 300401;
    2. 北京市科学技术研究院 北京市新技术应用研究所, 北京 100035
  • 收稿日期:2018-09-10 修回日期:2018-10-29 发布日期:2019-04-10 出版日期:2019-04-10
  • 通讯作者: 曲寒冰
  • 作者简介:祁子梁(1993-),男,河北邯郸人,硕士研究生,主要研究方向:计算机视觉、行人再识别;曲寒冰(1977-),男,黑龙江哈尔滨人,副研究员,博士,CCF会员,主要研究方向:机器学习、计算机视觉、生物识别、图像处理;赵传虎(1993-),男,河南平顶山人,硕士,主要研究方向:机器学习、数据挖掘;董良(1990-),男,河北邢台人,硕士,主要研究方向:数据挖掘、知识发现、机器学习、时空模式、社会网络;李博昭(1993-),女,河北邢台人,硕士,主要研究方向:机器学习、图像处理、模式识别;王长生(1989-),男,山东潍坊人,硕士研究生,主要研究方向:数据挖掘、机器学习。
  • 基金资助:
    国家重点研发计划项目(2018YFC08097000,2018YFC0704800,2018YFF0301000);国家自然科学基金资助项目(91746207);北京市科学技术研究院萌芽计划项目(GS201817)。

Person re-identification based on Siamese network and bidirectional max margin ranking loss

QI Ziliang1, QU Hanbing1,2, ZHAO Chuanhu1, DONG Liang2, LI Bozhao2, WANG changsheng2   

  1. 1. School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China;
    2. Beijing Institute of New Technology Applications, Beijing Academy of Science and Technology, Beijing 100035, China
  • Received:2018-09-10 Revised:2018-10-29 Online:2019-04-10 Published:2019-04-10
  • Supported by:
    This work is partially supported by the National Key R&D Program of China (2018YFC08097000, 2018YFC0704800, 2018YFF0301000), the National Natural Science Foundation of China (91746207), the Beijing Academy of Science and Technology Budding Plan (GS201817).

摘要: 针对在实际场景中存在的不同行人图像之间比相同行人图像之间更相似所造成的行人再识别准确率较低的问题,提出一种基于孪生网络并结合识别损失和双向最大边界排序损失的行人再识别方法。首先,对在超大数据集上预训练过的神经网络模型进行结构改造,主要是对最后的全连接层进行改造,使模型可以在行人再识别数据集上进行识别判断;其次,联合识别损失和排序损失监督网络在训练集上的训练,并通过正样本对的相似度值减去负样本对的相似度值大于预定阈值这一判定条件,来使得负例图像对之间的距离大于正例图像对之间的距离;最后,使用训练好的神经网络模型在测试集上测试,提取特征并比对特征之间的余弦相似度。在公开数据集Market-1501、CUHK03和DukeMTMC-reID上进行的实验结果表明,所提方法分别取得了89.4%、86.7%、77.2%的rank-1识别率,高于其他典型的行人再识别方法,并且该方法在基准网络结构下最高达到了10.04%的rank-1识别率提升。

关键词: 行人再识别, 孪生网络, 双向最大边界, 排序损失, 卷积神经网络

Abstract: Focusing on the low accuracy of person re-identification caused by that the similarity between different pedestrians' images is more than that between the same pedestrians' images in reality, a person re-identification method based on Siamese network combined with identification loss and bidirectional max margin ranking loss was proposed. Firstly, a neural network model which was pre-trained on a huge dataset, especially its final full-connected layer was structurally modified so that it can output correct results on the person re-identification dataset. Secondly, training of the network on the training set was supervised by the combination of identification loss and ranking loss. And according to that the difference between the similarity of the positive and negative sample pairs is greater than the predetermined value, the distance between negative sample pair was made to be larger than that of positive sample pair. Finally, a trained neural network model was used to test on the test set, extracting features and comparing the cosine similarity between the features. Experimental result on the open datasets Market-1501, CUHK03 and DukeMTMC-reID show that rank-1 recognition rates of the proposed method reach 89.4%, 86.7%, and 77.2% respectively, which are higher than those of other classical methods. Moreover, the proposed method can achieve a rank-1 rate improvement of up to 10.04% under baseline network structure.

Key words: person re-identification, Siamese network, bidirectional max margin, ranking loss, Convolutional Neural Network (CNN)

中图分类号: