基于深度学习SuperGlue的枪弹痕迹自动识别方法  

Automatic Identification Method of Bullet Marks Based on Deep Learning SuperGlue

作　　者：耿乐 沐春华 张浩 虞浒 GENG Le;MU Chun-hua;ZHANG Hao;YU Hu(School of Mechanical and Power Engineering,Nanjing Tech University,Nanjing,Jiangsu 211800,China)

机构地区：[1]南京工业大学机械与动力工程学院,江苏南京211800

出　　处：《计算技术与自动化》2023年第1期174-178,共5页Computing Technology and Automation

基　　金：国家自然科学基金资助项目(51205211、51305207)。

摘　　要：为实现枪弹痕迹自动匹配的准确性,提出了将基于深度学习的SuperPoint特征提取和SuperGlue匹配算法引入枪弹痕迹自动识别研究。通过SuperPoint网络提取弹痕图像特征点位置与描述子向量;研究了SuperGlue的匹配机制,包括注意力机制的图神经网络(GNN)及优化匹配层,将提取的弹底窝痕迹的特征点和描述子使用SuperGlue算法进行匹配。实验表明SuperPoint特征通过SuperGlue匹配,相较于机器学习算法实现了更高的匹配准确度,正确匹配数量提高,为枪支鉴定增加科学性。In order to achieve the accuracy of automatic matching of bullet traces,it is proposed to introduce the deep learning-based SuperPoint feature extraction and SuperGlue matching algorithm into the automatic identification of bullet traces.The position and descriptor vector of bullet marks image feature points are extracted through SuperPoint network;the matching mechanism of SuperGlue is studied,including attention graph neural network(GNN)and optimized matching layer,and SuperGlue algorithm is used to extract the feature points and descriptors of bullet bottom pit marks to match.Experiments show that SuperPoint features are matched by SuperGlue,which achieves higher matching accuracy than machine learning algorithms,and increases the number of correct matches,adding scientificity to gun identification.

关 键 词：枪弹痕迹 超级点算法 超级胶水算法 痕迹匹配 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]