面向图像匹配的局部特征提取研究进展  

Research Progress of Local Feature Extraction for Image Matching

作　　者：刘向增 徐雪灵 刘如意[1] 宋建锋[1] 苗启广[1] LIU Xiang-zeng;XU Xue-ling;LIU Ru-yi;SONG Jian-feng;MIAO Qi-guang(School of Computer Science and Technology,Xidian University,Xi’an 710071,China)

机构地区：[1]西安电子科技大学计算机科学与技术学院,陕西西安710071

出　　处：《计算机技术与发展》2022年第2期1-13,共13页Computer Technology and Development

基　　金：国家重点研发计划(B019030051);国家自然科学基金(61771369);陕西省自然科学基金(2020JM-195,2020JQ-330);西安市科技计划项目(GXYD6.2);中央高校基本科研业务费(20101216855)。

摘　　要：图像匹配作为计算机视觉领域的重要研究方向,广泛应用于图像配准、图像融合、变化检测、视觉导航、3D重建、视觉同时定位与地图构建(SLAM)等领域,精确稳健的局部特征提取是实现其高效处理的前提与关键。以图像匹配研究为导向,从传统特征设计到现代特征学习对局部特征提取方法进行了分类总结,首先,为增强对现代局部特征提取方法的理解,重点介绍了基于传统特征设计的相关方法,接着回顾了基于经典机器学习的方法,搭建起传统方法到深度学习方法的桥梁,最后详细讨论了基于深度学习的现代特征提取方法。针对跨传感器、多视角、不同时段环境下的图像匹配需求,全面分析了各阶段主流方法的优缺点,提出了目前存在的问题与挑战,并给出了相应的研究建议,为相关研究人员全面深入理解图像局部特征提取方法并利用深度学习方法对其进行改进提供基础性参考。As an important research direction in the field of computer vision, image matching is widely used in image registration, image fusion, change detection, visual navigation, 3 D reconstruction, simultaneous visual positioning and map construction(SLAM) and other fields. Among them, accurate and robust local feature extraction is the prerequisite and a key step to achieve efficient feature matching. Guided by image matching research, the local feature extraction methods are classified and summarized from traditional feature design to modern feature learning. Firstly, in order to enhance the understanding of modern local feature extraction methods, we focus on the related methods based on traditional feature design, then review the methods based on classic machine learning, build a bridge from traditional methods to deep learning methods, and finally discuss in detail based on modern feature extraction methods for deep learning. In view of the requirements of image matching under cross-sensor, multi-view, and different time periods, the advantages and disadvantages of mainstream methods at each stage are comprehensively analyzed, and the current problems and challenges of local feature extraction in image matching are put forward, and corresponding research recommendations are proposed. It provides a basic reference for related researchers to fully understand the topic of local feature extraction and improve the results with deep learning.

关 键 词：图像匹配 局部特征 特征描述 不变特征 深度学习 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]