基于深度度量学习的足迹图像检索算法  

A Footprint Image Retrieval Algorithm Based on Deep Metric Learning

作　　者：朱明[1] 江畅 于小勇[2] 殷克华 唐俊[4] 王年[4] ZHU Ming;JIANG Chang;YU Xiaoyong;YIN Kehua;TANG Jun;WANG Nian(School of Integrated Circuits,Anhui University,Hefei 230601,China;Institute of Forensic Science,Nanjing Municipal Public Security Bureau,Nanjing 210012,China;Criminal Investigation Corps of Public Security Department of Zhejiang Province,Hangzhou 310009,China;School of Electronic and Information Engineering,Anhui University,Hefei 230601,China)

机构地区：[1]安徽大学集成电路学院,合肥230601 [2]南京市公安局刑事科学技术研究所,南京210012 [3]浙江省公安厅刑侦总队,杭州310009 [4]安徽大学电子信息工程学院,合肥230601

出　　处：《刑事技术》2023年第3期283-291,共9页Forensic Science and Technology

基　　金：安徽高校自然科学研究重大项目(KJ2021ZD0004);公安部刑事技术“双十计划”重点攻关任务(2020SSGG0203);安徽省科技合作专项-科技强警(2022k07020006)。

摘　　要：针对赤足迹光学图像相似度高,传统算法无法快速提取样本有效信息的问题,提出了一种基于深度度量学习的赤足迹图像检索算法,该算法使用多尺度特征融合结构构建主干网络并辅以批去除结构监督学习,同时在嵌入空间内构建动态难样本三元组,解决传统难样本三元组挖掘方式不合理的问题。构建了一个包含250人、2 500枚的赤足迹光学图像数据集,该算法在所构建的数据集上平均检索精度达到88.86%,高于传统难样本三元组算法,可有效地应用于赤足迹光学图像检索。该算法对于足迹领域的自动化识别具有积极意义。In order to solve the problem of high similarity of footprint optical image and traditional algorithms cannot extract the effective information of samples quickly;this paper proposes a footprint image retrieval algorithm based on deep metric learning.The algorithm uses multi-scale feature fusion structure to construct backbone network,supplemented by batch removal structure supervised learning,and constructs dynamic hard sample triplet in the embedded space,and solves the unreasonable problem of traditional hard sample triplet mining method.A footprint optical image data set containing 2500 images of 250 people is constructed.The average retrieval accuracy of this algorithm on the constructed data set reaches 88.86%,which is higher than the traditional difficult-sample triple algorithm.It can be effectively applied to the footprints optical image retrieval.This algorithm has positive significance for automatic recognition of footprints.

关 键 词：深度度量学习 赤足迹图像检索 多尺度特征融合 嵌入空间 动态难样本三元组 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] D918.91[自动化与计算机技术—计算机科学与技术]