Survey of remote sensing image registration based on deep learning

Abstract

Remote sensing image registration is the process of spatial alignment of two or more images through geometric transformation. It is an important preprocessing operation for image fusion, change detection, agricultural monitoring and other remote sensing applications. Considering that remote sensing images have the characteristics of large-scale changes, complex ground covers and imaging modalities, although a large number of registration methods have been developed, there is still a lack of methods that can be widely used in different scenarios. Therefore, research on registration algorithms with high efficiency, high robustness, high precision and wide applicability is of great significance. In recent years, deep learning, which has achieved great success in the field of natural image and medical image registration, has provided a new method for remote sensing image registration. First, we introduced two kinds of traditional registration methods and analyzed the advantages and disadvantages of area-based and feature-based registration methods in detail from the aspects of registration accuracy, efficiency and algorithm robustness. Generally, there are two main problems in traditional methods: poor applicability and insufficient utilization of the deep semantic information of the image. Second, we focused on the important progress of deep learning in area-based registration methods and feature-based registration methods. According to the specific application purpose of deep learning, we made a more detailed division of the above two methods and summarized the advantages and disadvantages of the existing research. In addition, considering the importance of datasets for deep learning, we sorted and shared some public datasets for remote sensing image registration. Due to the great progress of earth observation technology, an increasing number of remote sensing images are being applied. Image registration is the key step of remote sensing image preprocessing and the basic research content of quantitative remote sensing analysis. In recent years, research on remote sensing image registration algorithms based on deep learning has shown an increasing trend, but it is still in the early stage, and the framework is not mature. It mainly includes but is not limited to the following shortcomings: (1) lack of open source standard datasets; (2) difficult to apply to large-scale remote sensing images; (3) insufficient utilization of geospatial information and spectral information of remote sensing images; and (4) long training time and the large computing overhead. From the perspective of data and methods, we looked forward to the application of deep learning in the field of remote sensing image registration and put forward four main research directions: (1) remote sensing image registration datasets; (2) registration methods based on hybrid models; (3) registration methods based on different neural networks; and (4) training strategies based on small samples.