航空学报 > 2024, Vol. 45 Issue (3): 28796-028796   doi: 10.7527/S1000-6893.2023.28796

基于深度学习的图像匹配方法综述

刘海桥1, 刘萌1, 龚子超1, 董晶2()   

  1. 1.湖南工程学院 电气与信息工程学院,湘潭 411104
    2.中南大学 航空航天技术研究院,长沙 410083
  • 收稿日期:2023-04-03 修回日期:2023-04-24 接受日期:2023-07-06 出版日期:2024-01-16 发布日期:2023-07-21
  • 通讯作者: 董晶 E-mail:dongjing@csu.edu.cn
  • 基金资助:
    国家自然科学基金(62173134);湖南省教育厅科研项目(21B0661)

A review of image matching methods based on deep learning

Haiqiao LIU1, Meng LIU1, Zichao GONG1, Jing DONG2()   

  1. 1.School of Electrical and Information Engineering,Hunan Institute of Engineering,Xiangtan 411104,China
    2.Research Institute of Aerospace Technology,Central South University,Changsha 410083,China
  • Received:2023-04-03 Revised:2023-04-24 Accepted:2023-07-06 Online:2024-01-16 Published:2023-07-21
  • Contact: Jing DONG E-mail:dongjing@csu.edu.cn
  • Supported by:
    National Natural Science Foundation of China(62173134);Scientific Research Foundation of Hunan Provincial Department of Education(21B0661)

摘要:

图像匹配是飞行器视觉导航中的一项关键技术。基于深度学习的图像匹配方法在近几年快速发展,其特征提取网络比传统方法具有明显优势与广阔的应用前景。基于深度学习的图像匹配方法可以按照网络结构的不同分为单环节网络模型匹配方法和端到端网络模型匹配方法。首先对单环节网络模型中的特征检测模型、描述符学习模型、相似度度量模型和误差剔除模型逐一进行了深度调研及分析,然后对端到端匹配网络模型中的单网络结构方法和多网络结构组合方法进行了针对性的综述,并对经典的端到端匹配网络模型算法进行了介绍和分析。最后,结合目前基于深度学习的图像匹配方法存在的问题,指出未来可能的发展趋势和方向,为后续研究者在深度学习图像匹配的研究提供一定参考。

关键词: 深度学习, 图像匹配, 视觉导航, 单环节, 端到端, 特征检测

Abstract:

Image matching is a key technology in aircraft visual navigation. The image matching methods based on deep learning have developed rapidly in recent years. The feature extraction network of the methods has obvious advantages over traditional methods, and has broad prospects for application. The image matching method based on deep learning can be divided into the single-link matching network model method and the end-to-end matching network model matching method according to different network structures. In this paper, the feature detection, descriptor learning, similarity measurement and error elimination network model of the network model of the single-link matching method network model are first investigated and analyzed. Then, the single-network structure method and multi-network structure combination method in the end-to-end matching network model are reviewed, and the classic end-to-end matching network model algorithm is introduced and analyzed. Finally, the problems of current image matching methods based on deep learning are pointed out, and the possible development trend and direction in the future are discussed to provide a certain reference for subsequent research on deep-learning based image matching.

Key words: deep learning, image matching, visual navigation, single link, end-to-end, feature detection

中图分类号: