图像特征检测与匹配方法研究综述  

A survey of image feature detection and matching methods

作　　者：唐灿[1] 唐亮贵[1] 刘波[1] TANG Can;TANG Lianggui;LIU Bo(School of Computer Science and Information Engineering,Chongqing Technology and Business University,Chongqing 400067)

机构地区：[1]重庆工商大学计算机科学与信息工程学院,重庆400067

出　　处：《南京信息工程大学学报（自然科学版）》2020年第3期261-273,共13页Journal of Nanjing University of Information Science & Technology（Natural Science Edition）

基　　金：重庆工商大学开放课题(KFJJ2019067);重庆市教委课题(1792079)。

摘　　要：几十年来,图像特征检测与匹配一直是图像处理的最核心领域之一,是计算机视觉的基石.没有特征检测与匹配就没有SLAM、Sfm、AR、通用图像检索、图像配准、全景图像等视觉任务.本文在回顾几十年来的经典检测算法的基础上,阐述了引用最新的以深度学习为首的机器学习算法后,在本领域取得的最新进展,包括特征点、局部特征子、全局特征子、匹配及优化、端到端框架等所有关键点,展示了算法各自的优缺点.总而言之,面对工业界的宽基线、实时、低算力检测的要求,图像特征检测和匹配仍然是一项未能完整攻克的任务,融合特征点、局部特征子、全局特征子、匹配及优化的多任务全局框架成为未来发展的趋势.For decades,image feature detection and matching has been the foundation of computer vision.Without feature detection and matching,there would be no visual tasks such as SLAM,Sfm,AR,image retrieval,image regis⁃tration,or panoramic images.Based on the review of classic detection algorithms in the past decades,this paper describes the latest progress in image feature detection and matching after the introduction of machine learning algo⁃rithm led by deep learning.The survey includes all the key points such as feature points,local descriptor,global de⁃scriptor,matching and optimization,and end⁃to⁃end framework,and compares the merits and demerits of each algo⁃rithm.In summary,facing the requirements of wide baseline,real⁃time,and low computing load detection from the in⁃dustrial sector,image feature detection and matching is still a hard task.The multitasking global framework which fu⁃ses feature points,local descriptor,global descriptor,matching and optimization has become the trend of future re⁃search.

关 键 词：图像特征检测 描述子 匹配算法 深度学习 

分 类 号：TP13[自动化与计算机技术—控制理论与控制工程]