联合损失优化孪生网络的行人重识别

计算机工程与科学

联合损失优化孪生网络的行人重识别

樊琳1,2，张惊雷1,2

（1.天津理工大学电气电子工程学院，天津 300384；2．天津市复杂系统控制理论及应用重点实验室，天津 300384）

收稿日期:2019-07-13 修回日期:2019-08-29 出版日期:2020-02-25 发布日期:2020-02-25

Person re-identification based

on joint loss and siamese network

FAN Lin1,2,ZHANG Jing-lei1,2

(1.School of Electrical and Electronic Engineering,Tianjin University of Technology,Tianjin 300384;

2.Tianjin Key Laboratory for Control Theory and Applications in Complicated Systems,Tianjin 300384,China)

Received:2019-07-13 Revised:2019-08-29 Online:2020-02-25 Published:2020-02-25

摘要/Abstract

摘要：

针对行人重识别应用中行人图像易受到光照、相似着装、拍摄角度影响而出现难分样本对，导致错误匹配的问题，提出一种联合损失结合孪生网络的行人重识别优化算法。首先利用残差卷积神经网络提取图像特征，并以焦点损失（Focal Loss）和交叉熵损失的联合损失对提取的特征进行监督训练，增加模型对难分样本对的关注度；然后采用余弦距离计算图像间的相似度实现行人的重识别；最后加入重排序算法降低误匹配率。采用Market-1501和DukeMTMC-reID数据集进行实验，结果表明，该算法的匹配率分别为91.2%和84.4%，平均精度均值（mAP）分别为85.8%和78.6%。

关键词: 行人重识别, 残差卷积神经网络, 孪生网络, Focal Loss

Abstract:

In person re-identification applications, pedestrian images are susceptible to illumination changing, similar wearing and different shooting angles, so it is difficult to distinguish sample pairs, which leads to incorrect matching. Aiming at this problem, a person re-identification optimization algorithm based on joint loss and siamese network is proposed. Firstly, the residual convolutional neural network is used to extract image features. The joint loss including focal loss and cross entropy loss is used to carry out supervised traning on the extracted features, in order to increase the models attention to the indistinguishable pairs. Then, the cosine distance is used to calculate the similarity between images to realize the pedestrian recognition. Finally, a re-ranking algorithm is adopted to reduce the mismatch rate. Experimental results on open datasets including Market-1501 and Duke MTMC-reID show that the matching rate (Rank-N) is 91.2% and 84.4%, respectively, and the mean Average Precision（mAP）is 85.8% and 78.6%.

Key words: person re-identification, residual convolutional neural network, siamese network, Focal Loss

樊琳1,2，张惊雷1,2. 联合损失优化孪生网络的行人重识别[J]. 计算机工程与科学.

FAN Lin1,2,ZHANG Jing-lei1,2.

Person re-identification based

on joint loss and siamese network

[J]. Computer Engineering & Science.

[1]	马玉民, 钱育蓉, 周伟航, 公维军, 帕力旦·吐尔逊. 基于孪生网络的目标跟踪算法综述[J]. 计算机工程与科学, 2023, 45(09): 1578-1592.
[2]	王可超, 郭军军, 张亚飞, 高盛祥, 余正涛, . 基于回译和比例抽取孪生网络筛选的汉越平行语料扩充方法[J]. 计算机工程与科学, 2022, 44(10): 1861-1868.
[3]	张俊飞, 张贵英. 改进残差神经网络实现心音分类[J]. 计算机工程与科学, 2022, 44(08): 1514-1520.
[4]	冯琪堯, 张惊雷, . 基于混合注意力机制的目标跟踪算法[J]. 计算机工程与科学, 2022, 44(02): 276-282.
[5]	方梦华, 姜添, . 基于无监督学习的无人机目标跟踪[J]. 计算机工程与科学, 2021, 43(06): 1024-1031.
[6]	杨德志, 柯显信, 余其超, 杨帮华. 基于RCNN的问题相似度计算方法[J]. 计算机工程与科学, 2021, 43(06): 1076-1080.
[7]	曾上游, 贾小硕, 李文惠. 基于双线程LSTM在线更新的视频追踪算法[J]. 计算机工程与科学, 2021, 43(03): 480-485.
[8]	熊炜, 杨荻椿, 艾美慧, 李敏, 李利荣. 基于多支路特征融合的行人重识别研究[J]. 计算机工程与科学, 2021, 43(02): 329-339.
[9]	熊炜, 熊子婕, 杨荻椿, 童磊, 刘敏, 曾春艳. 基于深层特征融合的行人重识别方法[J]. 计算机工程与科学, 2020, 42(02): 358-364.
[10]	熊炜1,2，冯川1，熊子婕1，王娟1,2，刘敏1,2，曾春艳1,2. 基于CNN的改进行人重识别技术[J]. 计算机工程与科学, 2019, 41(04): 665-672.