基于SuperPoint和SuperGlue的图像特征匹配  

Image Feature Matching Based on SuperPoint and SuperGlue

作　　者：杨金玲 马俊海 曹先革 YANG Jinling;MA Junhai;CAO Xiange(School of Surveying and Geoinformation Engineering,East China University of Technology,Nanchang 330013,China;Shenzhen Xinwei Surveying and Mapping Technology Co.,Ltd.,Shenzhen 518100,China)

机构地区：[1]东华理工大学测绘与空间信息工程学院,江西南昌330013 [2]深圳市新维测绘科技有限公司,广东深圳518100

出　　处：《测绘与空间地理信息》2024年第9期8-11,共4页Geomatics & Spatial Information Technology

基　　金：东华理工大学博士科研启动基金(DHBK2019193,DHBK2019194)资助。

摘　　要：基于SuperPoint和SuperGlue算法原理,将两者级联在一起,实现了完整的深度神经网络图像特征匹配流程。基于设计的深度神经网络,选取4组多源图像数据进行了特征匹配,并与5种传统特征匹配算法进行了对比。实验结果表明:本文设计的深度神经网络在多源图像特征提取、描述和匹配中具有稳定性强、鲁棒性高、识别特征点多且分布合理、正确匹配率高等特点。This article first introduces the principles of SuperPoint and SuperGlue algorithms,and achieves a complete deep neural network image feature matching process by cascading the two together.Based on the designed deep neural network,four sets of multisource images were selected for image feature matching,and compared with five classic feature matching algorithms.The experimental results show that the deep neural network designed in this paper has strong stability,high robustness,multiple and reasonably distributed recognition feature points,and high accuracy in feature extraction,description,and matching of multi-source images.

关 键 词：SuperPoint SuperGlue 深度学习 深度网络 特征匹配 

分 类 号：P208[天文地球—地图制图学与地理信息工程]