高级搜索

留言板

尊敬的读者、作者、审稿人, 关于本刊的投稿、审稿、编辑和出版的任何问题, 您可以本页添加留言。我们将尽快给您答复。谢谢您的支持!

姓名
邮箱
手机号码
标题
留言内容
验证码

双域滤波三元组度量学习的行人再识别

肖进胜 郭浩文 张舒豪 邹文涛 王元方 谢红刚

肖进胜, 郭浩文, 张舒豪, 邹文涛, 王元方, 谢红刚. 双域滤波三元组度量学习的行人再识别[J]. 电子与信息学报, 2022, 44(11): 3931-3940. doi: 10.11999/JEIT210385
引用本文: 肖进胜, 郭浩文, 张舒豪, 邹文涛, 王元方, 谢红刚. 双域滤波三元组度量学习的行人再识别[J]. 电子与信息学报, 2022, 44(11): 3931-3940. doi: 10.11999/JEIT210385
XIAO Jinsheng, GUO Haowen, ZHANG Shuhao, ZOU Wentao, WANG Yuanfang, XIE Honggang. Pedestrian Re-IDentification Algorithm Based on Dual-domain Filtering and Triple Metric Learning[J]. Journal of Electronics & Information Technology, 2022, 44(11): 3931-3940. doi: 10.11999/JEIT210385
Citation: XIAO Jinsheng, GUO Haowen, ZHANG Shuhao, ZOU Wentao, WANG Yuanfang, XIE Honggang. Pedestrian Re-IDentification Algorithm Based on Dual-domain Filtering and Triple Metric Learning[J]. Journal of Electronics & Information Technology, 2022, 44(11): 3931-3940. doi: 10.11999/JEIT210385

双域滤波三元组度量学习的行人再识别

doi: 10.11999/JEIT210385
基金项目: 国家自然科学基金(42101448)
详细信息
    作者简介:

    肖进胜:男,博士,副教授,硕士生导师,研究方向为图像与视频处理

    郭浩文:男,硕士生,研究方向为图像处理与行人重识别

    张舒豪:男,硕士生,研究方向为图像与计算机视觉

    邹文涛:男,硕士生,研究方向为图像与视频处理

    王元方:男,硕士生,研究方向为图像与视频处理

    谢红刚:男,博士,副教授,硕士生导师,研究方向为图像与机器视觉

    通讯作者:

    谢红刚 xiehg@hbut.edu.cn

  • 中图分类号: TN911.73; TP391.41

Pedestrian Re-IDentification Algorithm Based on Dual-domain Filtering and Triple Metric Learning

Funds: The National Natural Science Foundation of China (42101448)
  • 摘要: 在图像的捕获、传输或者处理过程中都有可能产生噪声,当图像被大量噪声影响时,许多行人再识别(ReID)方法将很难提取具有足够表达能力的行人特征,表现出较差的鲁棒性。该文主要针对低质图像的行人再识别问题,提出双域滤波分解构建3元组,用于训练度量学习模型。所提方法主要分为两个部分,首先分析了监控视频中不同图像噪声的分布特性,通过双域滤波进行图像增强。然后基于双域滤波分解对图像噪声具有很好的分离作用,该文提出一种新的3元组构建方式。在训练阶段,将双域滤波生成的低频原始图像和高频噪声图像,与原图一起作为输入3元组,网络可以进一步抑制噪声分量。同时优化了损失函数,将3元组损失和对比损失组合使用。最后利用re-ranking扩充排序表,提高识别的准确率。在加噪Market-1501和CUHK03数据集上的平均Rank-1为78.3%和21.7%,平均准确率均值(mAP)为66.9%和20.5%。加噪前后的Rank-1精度损失只有1.9%和7.8%,表明该文模型在含噪情况表现出较强的鲁棒性。
  • 图  1  3元组度量模型结构图

    图  2  双域滤波流程图

    图  3  模型训练整体结构图

    图  4  测试示意图

    图  5  Market-1501数据集上各方法CMC曲线

    图  6  CUHK03数据集上各方法CMC曲线

    表  1  不同图像增强方法在Market-1501和CUHK03数据集上的性能对比(%)

    指标无处理[12]K-SVD[13]Huang等人[14]Kang等人[15]Luo等人[16]Son等人[17]双域滤波[11]
    高斯噪声Rank-174.0/18.472.3/19.375.3/20.173.5/19.775.0/20.274.8/19.975.5/20.5
    mAP49.2/17.246.7/18.151.4/18.847.8/18.651.2/19.151.0/18.751.9/19.2
    椒盐噪声Rank-164.8/18.565.6/19.166.6/20.966.3/19.666.4/20.166.2/19.666.4/20.3
    mAP40.3/17.243.0/17.942.7/19.242.2/18.242.4/18.842.1/18.142.2/18.9
    雨噪声Rank-175.5/15.974.9/16.575.9/16.676.2/17.576.6/17.875.9/16.876.8/18.1
    mAP51.4/14.651.3/15.451.7/15.552.0/16.052.1/16.451.8/15.763.4/16.8
    无噪声Rank-178.9/22.276.2/22.379.1/22.477.7/22.379.0/22.478.9/22.379.2/22.5
    mAP55.0/21.052.4/21.865.1/22.553.8/22.064.7/22.664.4/22.266.8/22.9
    下载: 导出CSV

    表  2  不同图像增强方法的指标平均增益(%)

    数据集指标增益K-SVD[13]Huang等人[14]Kang等人[15]Luo等人[16]Son等人[17]双域滤波[11]
    Market-1501/
    CUHK03
    Rank-1增益–0.5/0.71.1/1.60.6/1.31.2/1.80.9/1.21.4/2.0
    mAP增益–0.02/0.81.6/1.50.3/1.31.6/1.81.3/1.25.5/2.0
    下载: 导出CSV

    表  3  消融实验结果

    3元组网络Triplet lossContrastive lossrerankingRank-1 (%)mAP (%)
    常规改进
    76.765.3
    74.953.2
    76.554.6
    78.767.8
    下载: 导出CSV

    表  4  Market-1501数据集各方法的Rank-1和mAP(%)

    方法原始图像高斯噪声椒盐噪声雨噪声
    Rank-1mAPRank-1mAPRank-1mAPRank-1mAP
    LOMO+XQDA[2]43.822.737.519.53921.635.720.3
    SpindleNet[7]76.9/72.4/74.4/73.7/
    IDE_ResNet_50[12]78.95574.249.475.551.975.150.6
    SVDNet[20]82.362.178.658.480.359.278.157.5
    APR[21]84.364.779.260.281.461.980.860.8
    PIE[6]79.356.074.850.576.153.077.154.2
    本文模型79.868.377.565.878.667.178.767.8
    下载: 导出CSV

    表  5  Market-1501数据集各方法Rank-1和mAP的平均下降率(%)

    LOMO+XQDA[2]SpindleNet[7]IDE_ResNet_50[12]SVDNet[20]APR[21]PIE[6]本文模型
    Rank-1下降14.64.45.04.04.54.21.9
    mAP下降9.8/7.96.05.86.12.0
    下载: 导出CSV

    表  6  CUHK03数据集各方法的Rank-1和mAP(%)

    方法原始图像高斯噪声椒盐噪声雨噪声
    Rank-1mAPRank-1mAPRank-1mAPRank-1mAP
    LOMO+XQDA[2]14.813.68.29.512.710.810.510.1
    SpindleNet[7]33.8/31.7/32.2/31.5/
    IDE_ResNet_50[12]22.221.018.417.619.518.818.917.8
    SVDNet[20]40.937.837.228.838.431.636.427.1
    APR[21]45.746.842.441.643.342.742.842.1
    PIE[6]34.231.131.425.833.129.629.422.6
    本文模型23.522.721.119.821.420.222.521.5
    下载: 导出CSV

    表  7  CUHK03数据集上各方法Rank-1和mAP的平均下降率(%)

    LOMO +XQDA[2]SpindleNet[7]IDE_ResNet_50[12]SVDNet[20]APR[21]PIE[6]本文模型
    Rank-1下降29.35.914.78.76.38.57.8
    mAP下降25.5/14.022.810.016.49.7
    下载: 导出CSV
  • [1] 何果财, 刘峡壁. 基于图像三元组挖掘的无监督视觉表示学习[J]. 计算机学报, 2018, 42(12): 2787–2803. doi: 10.11897/SP.J.1016.2018.02787

    HE Guocai and LIU Xiabi. Unsupervised visual representation learning with image triplets mining[J]. Chinese Journal of Computers, 2018, 42(12): 2787–2803. doi: 10.11897/SP.J.1016.2018.02787
    [2] LIAO Shengcai, HU Yang, ZHU Xiangyu, et al. Person re-identification by local maximal occurrence representation and metric learning[C]. 2015 IEEE Conference on Computer Vision and Pattern Recognition, Boston, USA, 2015: 2197–2206.
    [3] 桑海峰, 王传正, 吕应宇, 等. 基于多信息流动卷积神经网络的行人再识别[J]. 电子学报, 2019, 47(2): 351–357. doi: 10.3969/j.issn.0372-2112.2019.02.014

    SANG Haifeng, WANG Chuanzheng, LÜ Yingyu, et al. Person re-identification based on multi-information flow convolutional neural network[J]. Acta Electronica Sinica, 2019, 47(2): 351–357. doi: 10.3969/j.issn.0372-2112.2019.02.014
    [4] LUO Hao, JIANG Wei, ZHANG Xuan, et al. AlignedReID++: Dynamically matching local information for person re-identification[J]. Pattern Recognition, 2019, 94: 53–61. doi: 10.1016/j.patcog.2019.05.028
    [5] 刘一敏, 蒋建国, 齐美彬, 等. 融合生成对抗网络和姿态估计的视频行人再识别方法[J]. 自动化学报, 2020, 46(3): 576–584. doi: 10.16383/j.aas.c180054

    LIU Yimin, JIANG Jianguo, QI Meibin, et al. Video-based person re-identification method based on GAN and pose estimation[J]. Acta Automatica Sinica, 2020, 46(3): 576–584. doi: 10.16383/j.aas.c180054
    [6] ZHENG Liang, HUANG Yujia, LU Huchuan, et al. Pose-invariant embedding for deep person re-identification[J]. IEEE Transactions on Image Processing, 2019, 28(9): 4500–4509. doi: 10.1109/TIP.2019.2910414
    [7] ZHAO Haiyu, TIAN Maoqing, SUN Shuyang, et al. Spindle net: Person re-identification with human body region guided feature decomposition and fusion[C]. The 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 907–915.
    [8] KÖSTINGER M, HIRZER M, WOHLHART P, et al. Large scale metric learning from equivalence constraints[C]. 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, USA, 2012: 2288–2295.
    [9] HERMANS A, BEYER L, and LEIBE B. In defense of the triplet loss for person re-identification[J]. arXiv preprint arXiv: 1703.07737. 2017.
    [10] CHEN Weihua, CHEN Xiaotang, ZHANG Jianguo, et al. Beyond triplet loss: A deep quadruplet network for person re-identification[C]. 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 1320–1329.
    [11] XIAO Jinsheng, ZOU Wentao, CHEN Yunhua, et al. Single image rain removal based on depth of field and sparse coding[J]. Pattern Recognition Letters, 2018, 116: 212–217. doi: 10.1016/j.patrec.2018.10.006
    [12] ZHONG Zhun, ZHENG Liang, CAO Donglin, et al. Re-ranking person re-identification with k-reciprocal encoding[C]. 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 3652–3661.
    [13] AHARON M, ELAD M, and BRUCKSTEIN A. K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation[J]. IEEE Transactions on Signal Processing, 2006, 54(11): 4311–4322. doi: 10.1109/TSP.2006.881199
    [14] HUANG Dean, KANG Liwei, YANG Minchun, et al. Context-aware single image rain removal[C]. 2012 IEEE International Conference on Multimedia and Expo, Melbourne, Australia, 2012: 164–169.
    [15] KANG Liwei, LIN C W, and FU Y H. Automatic single-image-based rain streaks removal via image decomposition[J]. IEEE Transactions on Image Processing, 2012, 21(4): 1742–1755. doi: 10.1109/TIP.2011.2179057
    [16] LUO Yu, XU Yong, and JI Hui. Removing rain from a single image via discriminative sparse coding[C]. 2015 IEEE International Conference on Computer Vision, Santiago, Chile, 2015: 3397–3405.
    [17] SON C H and ZHANG Xioaping. Rain detection and removal via shrinkage-based sparse coding and learned rain dictionary[J]. Journal of Imaging Science and Technology, 2020, 64(3): 30501. doi: 10.2352/J.ImagingSci.Technol.2020.64.3.030501
    [18] 肖进胜, 李文昊, 姜红, 等. 基于双域滤波的三维块匹配视频去噪算法[J]. 通信学报, 2015, 36(9): 91–97. doi: 10.11959/j.issn.1000-436x.2015245

    XIAO Jinsheng, LI Wenhao, JIANG Hong, et al. Three dimensional block-matching video denoising algorithm based on dual-domain filtering[J]. Journal on Communications, 2015, 36(9): 91–97. doi: 10.11959/j.issn.1000-436x.2015245
    [19] 陈巧媛, 陈莹. 基于困难样本三元组损失的多任务行人再识别[J]. 计算机辅助设计与图形学学报, 2019, 31(7): 1156–1165. doi: 10.3724/SP.J.1089.2019.17463

    CHEN Qiaoyuan and CHEN Ying. TriHard loss based multi-task person re-identification[J]. Journal of Computer-Aided Design &Computer Graphics, 2019, 31(7): 1156–1165. doi: 10.3724/SP.J.1089.2019.17463
    [20] SUN Yifan, ZHENG Liang, DENG Weijian, et al. SVDNet for pedestrian retrieval[C]. 2017 IEEE International Conference on Computer Vision, Venice, Italy, 2017: 3820–3828.
    [21] LIN Yutian, ZHENG Liang, ZHENG Zhedong, et al. Improving person re-identification by attribute and identity learning[J]. Pattern Recognition, 2019, 95: 151–161. doi: 10.1016/j.patcog.2019.06.006
  • 加载中
图(6) / 表(7)
计量
  • 文章访问数:  88
  • HTML全文浏览量:  25
  • PDF下载量:  23
  • 被引次数: 0
出版历程
  • 收稿日期:  2021-05-07
  • 修回日期:  2022-09-02
  • 网络出版日期:  2022-09-03
  • 刊出日期:  2022-11-14

目录

    /

    返回文章
    返回