High-resolution Solar Image Registration Using Unsupervised Deep Learning

High-resolution solar images captured at different wavelengths are essential for understanding solar activity. However, such images often exhibit geometric discrepancies due to varying instrument resolutions and observation conditions, making image registration a critical preprocessing step. In this study, we propose an unsupervised deep learning-based framework named hourGlass and haRdnEt (GRE) for accurate solar image registration. The method detects keypoints in both the reference and moving images, extracts local feature descriptors, performs bidirectional matching to establish reliable correspondences, and estimates affine transformation parameters to align the images. The proposed framework was evaluated on quiet Sun images from the New Vacuum Solar Telescope and active region (AR) images from the Goode Solar Telescope, covering both photospheric and chromospheric features. A synthetic data set with known transformations was also used to assess registration accuracy under controlled conditions. Registration performance was quantitatively measured using mutual information (MI) and structural similarity index (SSIM) methods, and results were compared with those obtained using the scale-invariant feature transform and intensity-based two-step methods. The experimental results demonstrate that the proposed method achieves accurate registration across different solar features and imaging scenarios. The method maintains structural consistency in both AR and quiet Sun observations, including time-series data, with MI and SSIM improvements over baseline methods. The approach provides a validated tool for solar image alignment, suitable for further solar physics studies.