[1]张誉馨,张索非,王文龙,等.面向行人重识别的多域批归一化问题研究[J].计算机技术与发展,2022,32(01):91-97.[doi:10. 3969 / j. issn. 1673-629X. 2022. 01. 016]
　ZHANG Yu-xin,ZHANG Suo-fei,WANG Wen-long,et al.Research on Batch Normalization for Multi-domain PersonRe-identification[J].,2022,32(01):91-97.[doi:10. 3969 / j. issn. 1673-629X. 2022. 01. 016]
点击复制

面向行人重识别的多域批归一化问题研究()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
32
期数:
2022年01期
页码:
91-97
栏目:
图形与图像
出版日期:
2022-01-10

文章信息/Info

Title:
Research on Batch Normalization for Multi-domain PersonRe-identification
文章编号:
1673-629X(2022)01-0091-07
作者:
张誉馨1 ; 张索非2 ; 王文龙1 ; 吴晓富1
1. 南京邮电大学 通信与信息工程学院,江苏 南京 210003;
2. 南京邮电大学 物联网学院,江苏 南京 210003
Author(s):
ZHANG Yu-xin1 ; ZHANG Suo-fei2 ; WANG Wen-long1 ; WU Xiao-fu1
1. School of Telecommunications & Information Engineering,Nanjing University ofPosts and Telecommunications,Nanjing 210003,China;
2. School of Internet of Things,Nanjing University of Posts and Telecommunications,Nanjing 210003,China
关键词:
计算机视觉; 深度学习; 行人重识别; 多域训练; 批归一化
Keywords:
computer vision; deep learning; pedestrian Re-ID; multi-domain training; batch normalization
分类号:
TP391. 41
DOI:
10. 3969 / j. issn. 1673-629X. 2022. 01. 016
摘要:
近年来基于深度神经网络的行人重识别算法取得了长足的进步, 被广泛应用于网络中的批归一化 ( batchnormalization) 模块发挥着重要作用。 批归一化模块在多数情况下可有效提高网络收敛速度和训练稳定性,然而当多个独立标注的数据库混合在一块进行跨域或者多域训练时,数据之间的分布差异使得目前的批归一化算法工作逻辑存疑。 由于不同批次下训练数据的分布差异较大,归一化过程中的统计参数不稳定导致批归一化效果恶化。 该文聚焦于多数据集合并下的行人重识别模型训练问题,通过对多数据集分布差异导致的多域模型批归一化存在的问题进行分析。 然后针对模型批量归一化算法面对的多域差异,提出了一种解决策略,在多个数据集并行训练下提高了模型的泛化能力。 实验结果表明:所提出的多域归一化方法在多域训练下能有效提高模型最终的泛化能力,获得更高的识别准确度,并且可应用于其他行人重识别网络以进一步提升模型性能。
Abstract:
In recent years,the pedestrian re-identification ( Re-ID) algorithms based on deep neural networks ( DNNs) have made considerable progress. The use of batch normalization modules in constructing DNNs is key to the success of deep Re - ID, which caneffectively improve the convergence speed and training stability in most cases. However,when multiple independently labeled datasets aremixed together for cross- domain or multi - domain training,the domain gap makes the use of the current batch normalization modulequestionable. The distribution difference among training batches may result in unstable estimates for various statistical parameters,whichmay lead to possible performance deterioration for the employed batch normalization. We focus on the training of pedestrian re -recognition models under the combination of multiple data sets and analyze the problems in the batch normalization of multi - domainmodels caused by the distribution differences of multiple data sets. Aiming at the multi - domain difference of the batch modelnormalization algorithm,a strategy is proposed to improve the generalization ability of the model under the parallel training of multipledata sets. Experiment shows that the proposed domain - specific batch normalization algorithm can effectively improve the finalgeneralization ability under multi-domain training and obtain higher recognition accuracy. Furthermore,it can be applied to existing Re-ID networks for further improving their performance.

相似文献/References:

[1]黄艳 赵越.3D靶标的摄像机三步标定算法与实现[J].计算机技术与发展,2010,(01):135.
　HUANG Yan,ZHAO Yue.Algorithm and Realization of Three-step Camera Calibration Based on 3D-Target[J].,2010,(01):135.
[2]付海洋 牛连强 刘守琳.一种基于平面模板的单应矩阵求解方法[J].计算机技术与发展,2010,(04):69.
　FU Hai-yang,NIU Lian-qiang,LIU Shou-lin.A Solving Homography Matrix Method Based on Planar Pattern[J].,2010,(01):69.
[3]张铖伟 王彪 徐贵力.摄像机标定方法研究[J].计算机技术与发展,2010,(11):174.
　ZHANG Cheng-wei,WANG Biao,XU Gui-li.A Study on Classification of Camera Calibration Methods[J].,2010,(01):174.
[4]毛雁明 杨慧玲.一种新的立体匹配算法[J].计算机技术与发展,2011,(03):105.
　MAO Yan-ming,YANG Hui-ling.A New Stereo Matching Algorithm[J].,2011,(01):105.
[5]杨晟,李学军,王珏,等.连续尺度复合分析核线重排列影像准稠密匹配[J].计算机技术与发展,2013,(04):111.
　YANG Sheng,LI Xue-jun,WANG Jue,et al.Continuous Scale Multi-change Detecting Quasi-dense Matching for Epipolar Resample Images[J].,2013,(01):111.
[6]卢振宇,郭星,魏赛,等.基于计算机视觉的虚拟安全空间预警技术[J].计算机技术与发展,2014,24(02):237.
　LU Zhen-yu,GUO Xing,WEI Sai,et al.A Surveillance Technology for Virtual Security Space Based on Computer Vision[J].,2014,24(01):237.
[7]李孟,周波,孟正大,等. 三目立体相机的标定研究[J].计算机技术与发展,2015,25(02):69.
　LI Meng,ZHOU Bo,MENG Zheng-da,et al. Study on Trinocular Stereo Camera Calibration[J].,2015,25(01):69.
[8]陈浩翔,蔡建明,刘铿然,等. 手写数字深度特征学习与识别[J].计算机技术与发展,2016,26(07):19.
　CHEN Hao-xiang,CAI Jian-ming,LIU Keng-ran,et al. Deep Learning and Recognition of Handwritten Numeral Features[J].,2016,26(01):19.
[9]程龙乐[][],许金林[],李皙茹[][],等. 基于图像处理的跑步机速度自适应技术研究[J].计算机技术与发展,2016,26(10):92.
　CHENG Long-le[][],XU Jin-lin[],LI Xi-ru[][],et al. Research on Speed-adaptive Technology of Treadmill Based on Image Processing[J].,2016,26(01):92.
[10]贺飞翔,赵启军. 基于深度学习的头部姿态估计[J].计算机技术与发展,2016,26(11):1.
　HE Fei-xiang,ZHAO Qi-jun. Head Pose Estimation Based on Deep Learning[J].,2016,26(01):1.
[11]施泽浩,赵启军.基于全卷积网络的目标检测算法[J].计算机技术与发展,2018,28(05):55.[doi:10.3969/j.issn.1673－629X.2018.05.013]
　SHI Ze-hao,ZHAO Qi-jun.Object Detection Algorithm Based on Fully Convolutional Neural Network[J].,2018,28(01):55.[doi:10.3969/j.issn.1673－629X.2018.05.013]
[12]许必宵,宫 婧,孙知信.基于卷积神经网络的目标检测模型综述[J].计算机技术与发展,2019,29(12):87.[doi:10. 3969 / j. issn. 1673-629X. 2019. 12. 016]
　XU Bi-xiao,GONG Jing,SUN Zhi-xin.A Survey of Object Detection Models Based on Convolutional Neural Networks[J].,2019,29(01):87.[doi:10. 3969 / j. issn. 1673-629X. 2019. 12. 016]
[13]陈晓艺,陆一鸣,沈加炜,等.基于深度学习的灾后建筑物损坏程度检测综述[J].计算机技术与发展,2023,33(09):1.[doi:10. 3969 / j. issn. 1673-629X. 2023. 09. 001]
　CHEN Xiao-yi,LU Yi-ming,SHEN Jia-wei,et al.Review of Post-disaster Building Damage Detection Based on Deep Learning[J].,2023,33(01):1.[doi:10. 3969 / j. issn. 1673-629X. 2023. 09. 001]
[14]卜子渝,杨 哲,刘纯平.基于 EfficientNet 的无锚框目标检测模型[J].计算机技术与发展,2024,34(01):37.[doi:10. 3969 / j. issn. 1673-629X. 2024. 01. 006]
　BU Zi-yu,YANG Zhe,LIU Chun-ping.An Anchor-free Object Detection Model Based on EfficientNet[J].,2024,34(01):37.[doi:10. 3969 / j. issn. 1673-629X. 2024. 01. 006]

更新日期/Last Update: 2022-01-10