Very High-Resolution Satellite Image Registration Based on Self-supervised Deep Learning

Image registration that unifies coordinates between images is a fundamental preprocessing process for utilizing very high-resolution satellite imagery. In this study, we propose an image registration methodology that aligns coordinates between "VHR (Very High-Resolution) satellite" images using extracted conjugate points by a self-supervised deep learning network. MagicPoint detector is built using a synthetic shape dataset to learn overall characteristics of feature points. The MagicPoint detector is advanced using feature points extracted from various VHR remote sensing open-source datasets. Pseudo-label is generated in the VHR satellite imagery by the advanced MagicPoint detector. At this time, homographic adaptation is applied to consider the various geometric environments. SuperPoint is built using the VHR satellite imagery and the pseudo-label. The image registration is conducted using an estimated non-linear transformation model based on extracted conjugate points through the proposed method. Based on experiments conducted from KOMPSAT-3 imagery acquired over the Daejeon city, the proposed method evenly extracted numerous conjugate points over the overlapping areas, and registration accuracy was measured as RMSE (Root Mean Square Error) 1.563 pixels and CE90 (Circular Error 90%) 1.971 pixels. Therefore, it was demonstrated that the proposed method can contribute to the improvement of registration accuracy by extracting the conjugate points reflecting the characteristics of VHR satellite imagery.