基于深度学习的图像匹配:方法、应用与挑战  

Image Matching in Deep Learning Era:Methods,Applications and Challenges

作　　者：孔庆群[1,2] 吴福朝 樊彬[3] KONG Qing-Qun;WU Fu-Chao;FAN Bin(Institute of Automation,Chinese Academy of Sciences,Beijing 100190;University of Chinese Academy of Sciences,Beijing 100049;School of Intelligence Science and Technology,University of Science and Technology Beijing,Beijing 100083)

机构地区：[1]中国科学院自动化研究所,北京100190 [2]中国科学院大学,北京100049 [3]北京科技大学智能科学与技术学院,北京100083

出　　处：《计算机学报》2024年第7期1485-1520,共36页Chinese Journal of Computers

基　　金：北京市自然科学基金(4202073);国家自然科学基金(62222302,61876180)。

摘　　要：图像匹配旨在建立图像之间的点对应关系,是许多计算机视觉任务的关键环节.近年来,随着深度学习技术的发展,图像匹配方法已从以手工设计特征为主转变为基于深度网络的方法,基于深度学习的图像匹配方法在多个标准数据集上展现出卓越的性能,推动着多个相关应用的发展围绕图像匹配涉及的若干关键问题,如:特征点检测、特征点描述、稠密点匹配、误匹配去除,本文对深度学习图像匹配方法进行了系统性总结.首先分析了领域内基于深度学习的典型方法和关键技术,随后介绍了与图像匹配密切相关的几个典型应用并给出其现状分析,最后,根据对图像匹配领域技术发展的分析总结,结合作者在该领域的长期研究积累,本文给出了目前图像匹配所面临的主要挑战以及未来发展趋势。Image matching is a crucial technique within the field of computer vision,primarily focused on identifying and establishing point correspondences between two different images depicting the same scene.It seeks to find points in one image that correspond to points in another,thus enabling a wide range of computer vision tasks that rely on the analysis of multiple images of the same object or scene from different viewpoints or at different times,including but not limited to 3D reconstruction,motion tracking,image stitching for panoramic views,and visual localization.Traditionally,this process has leaned heavily on the use of hand-crafted keypoint detectors and local descriptors,i.e.,algorithms and methodologies designed to pinpoint and describe discriminative features within a local image region,aiming to achieve invariance to scale,rotation,and changes in lighting and perspective.In recent years,with the revolutionary development of deep learning in many areas of computer vision,image matching methods have switched from handcrafted design style to relying on deep learning.The advent of deep learning technologies has catalyzed significant advancements in the area of image matching,and numerous deep learning based image matching techniques have emerged,showcasing promising results across a wide range of benchmarks.This has also significantly accelerated the development of many downstream applications of image matching,notably including structure from motion,visual localization,and simultaneous localization and mapping(SLAM),among others.This paper aims to provide a comprehensive overview of deep learning-based image matching methods that have emerged in recent years.By delving into the core challenges of image matching,including keypoint detection,local feature description,dense matching,and mismatch removal,it offers a detailed summary of the innovative deep learning approaches devised to tackle these issues.This systematic review not only highlights the advancements in the field but also sheds light on how these cu

关 键 词：图像匹配 特征点匹配 稠密匹配 三维重建 视觉定位 同时定位与建图 深度学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]