测绘学报 ›› 2018, Vol. 47 ›› Issue (6): 844-853.doi: 10.11947/j.AGCS.2018.20170627

• 数字摄影测量与深度学习方法 • 上一篇    下一篇

卫星影像匹配的深度卷积神经网络方法

范大昭, 董杨, 张永生   

  1. 信息工程大学地理空间信息学院, 河南 郑州 450001
  • 收稿日期:2017-12-13 修回日期:2018-03-23 出版日期:2018-06-20 发布日期:2018-06-21
  • 作者简介:范大昭(1973-),男,博士,教授,研究方向为数字摄影测量的理论与应用。E-mail:fdzcehui@163.com
  • 基金资助:
    国家自然科学基金(41401534);地理信息工程国家重点实验室开放基金(SKLGIE2013-M-3-1)

Satellite Image Matching Method Based on Deep Convolution Neural Network

FAN Dazhao, DONG Yang, ZHANG Yongsheng   

  1. Institute of Geospatial Information, Information Engineering University, Zhengzhou 450001, China
  • Received:2017-12-13 Revised:2018-03-23 Online:2018-06-20 Published:2018-06-21
  • Supported by:
    The National Natural Science Foundation of China (No.41401534);The Open Fund of State Key Laboratory of Geographic Information Engineering (No.SKLGIE2013-M-3-1)

摘要: 本文侧重于智能化摄影测量深度学习的第一个方面:深度卷积方法。传统的影像同名点对提取算法通常利用人工设计的特征描述符及其最短距离作为匹配准则进行匹配,其匹配结果易陷入局部极值,造成部分正确匹配点对的遗漏。针对这一问题,本文引入深度学习方法,设计了一种基于空间尺度卷积层的两通道深度卷积神经网络,采用其进行影像间的匹配模式学习,实现了基于深度卷积神经网络的卫星影像匹配。试验表明,在处理异源、多时相、多分辨率的卫星影像情况下,本文方法比传统匹配方法能提取到更为丰富的影像同名点对,且最终匹配提纯结果正确率优于90%。

关键词: 影像匹配, 深度学习, 面向对象, 卷积神经网络, 卫星影像

Abstract: This article focuses on the first aspect of the album of deep learning: the deep convolution method.The traditional matching point extraction algorithm usually uses the manually-designed feature descriptor and the shortest distance between them to match as the matching criterion.The matching result is easy to fall into the local extreme value,which causes the missing of the partial matching point.Aiming at this problem,we introduce a two-channel deep convolution neural network based on spatial scale convolution,and performs matching pattern learning between images to realize the satellite image matching based on deep convolution neural network.The experimental results show that the method can extract the richer matching point in the case of heterogeneous,multi-temporal and multi-resolution satellite images,compared with the traditional matching method.And the accuracy of the final matching results can be maintained at above 90%.

Key words: image matching, deep learning, object-oriented, convolution neural network, satellite image

中图分类号: