基于多尺度特征的赤足足迹图像人身识别算法  

Algorithm of Personal Recognition Based on Multi-scale Features from Barefoot Footprint Image

作　　者：金益锋 于霄雪 王丽 李岱熹 蒋雪梅 程坚 谢敏 欧阳巍嘉 JIN Yifeng;YU Xiaoxue;WANG Li;LI Daixi;JIANG Xuemei;CHENG Jian;XIE Min;OUYANG Weijia(People’s Public Security University of China,Beijing 100038,China;Institute of Forensic Science,Ministry of Public Security,Beijing 100038,China;Dalian Everspry Sci-Tech Co.,Ltd.,Dalian 116023,Liaoning,China;Institute of Criminal Science and Technology of Jiangxi Provincial Public Security Department,Nanchang 330006,China;Unit 31056 of PLA,Beijing 100036,China)

机构地区：[1]中国人民公安大学,北京100038 [2]公安部鉴定中心,北京100038 [3]大连恒锐科技股份有限公司,辽宁大连116023 [4]江西省公安厅刑事科学技术研究所,南昌330006 [5]31056部队,北京100036

出　　处：《刑事技术》2022年第6期587-592,共6页Forensic Science and Technology

基　　金：公安部刑事技术“双十计划”重点攻关任务项目(2020SSGG0201、2020SSGG0205);公安部科技强警基础专项(2019GABJC17)。

摘　　要：本文提出一种基于深度学习技术的赤足足迹图像人身识别算法。以ResNet50为基础网络,结合水平金字塔匹配(horizontal pyramid matching,HPM)技术提取赤足足迹图像的多尺度特征,并利用三元组损失函数Separate Triplet Loss对赤足足迹进行人身度量学习。结果表明,本文基于6433人的赤足足迹进行训练,在11028人的开集赤足数据集上进行测试,所提出的算法的首位度识别准确率达到了96.2%,并且在CMC和mAP各项指标上均远优于常规的ResNet50网络结合交叉熵损失(cross-entropy loss)以及ArcFace损失的深度学习方法。实验证明,本文提出的基于赤足足迹的人身识别算法达到了很好的识别效果,在万人级别的采集数据上达到了较高的识别水准。A personal recognition algorithm was here to propose on the basis of deep learning with barefoot footprint images.ResNet50,a network-based approach to carry out deep learning into image recognition,was used as the basic configuration to collect the barefoot footprint images through extraction of their features against which HPM(Horizontal Pyramid Matching)manipulation was adopted to successively separate and recombine on multiple scales.Separate Triplet Loss was selected to perform the personal metric learning about the fully-featured barefoot footprints.All the related formulae were indicated on requirement.With the training of barefoot footprints of 6433 persons plus testing into an open barefoot dataset of 11028 people,the algorithm adopted here and supported with the relevant formulae did achieve its first-rank accuracy up to 96.2%,rendering far more superior CMC and mAP indications to the algorithms of both the ResNet50 coupling with Cross-Entropy Loss and ArcFace Loss.The here-adopted personal recognition algorithm is of good recognition effect for barefoot footprint,reaching a high recognition level at the collected data of ten thousand people.

关 键 词：人身识别 赤足足迹 深度学习 特征提取 度量学习 

分 类 号：DF794.1[政治法律—诉讼法学]