基于深度学习的SAR与光学影像配准方法综述

Review of SAR and Optical Image Registration Based on Deep Learning

遥感影像配准在遥感数据后续应用中发挥着基础性的作用,光学影像作为应用最广泛的数据源与全天时全天候获取数据的合成孔径雷达(SAR)影像综合利用能够获取更丰富的信息,SAR与光学影像配准已成为前沿热点问题。鉴于深度学习方法在光学影像配准中取得的成功,其在SAR与光学影像配准上也得以发展,对基于深度学习的SAR与光学影像配准方法进行归类与总结,根据是否直接使用深度学习网络提取描述异源影像特征将其分为特征描述符学习和风格迁移2类方法,将现有研究从使用的网络模型、损失函数和数据集等方面进行总结,并简要介绍了适用于SAR与光学配准的公开数据集以及图像匹配评价指标。

Remote sensing image registration plays a fundamental role in the subsequent application of remote sensing data.As the most widely used data source,optical images utilized in combination with SAR images which obtain all-day and all-weather data can provide more abundant information.SAR and optical image registration has become a hot topic of research.Given the successful application of deep learning in optical image registration and its development in SAR and optical image registration,the methods of SAR and optical image registration based on deep learning are classified and summarized.According to whether the deep learning network is directly used to extract and describe the features of heterogeneous images,the methods are divided into two categories,i.e.feature descriptor learning and style transfer.Current researches are summarized in terms of network model,loss function and data set,etc.Besides,the public data set and evaluation index for SAR and optical registration are briefly introduced.