Auroral Image Super-Resolution via Dynamic Nonlocal Intensity Fusion and Low-rank Regularization

Analyzing auroral images is particularly significant for studying the relationship between auroral phenomena and magnetospheric dynamics. Using artificial intelligence methods, we have implemented auroral classification and event detection based on auroral images. However, in previous studies, only images with a single resolution were considered. In this paper, facing the natural image dataset with time continuity such as auroral image, we propose a novel super-resolution method of dynamic nonlocal intensity fusion and low-rank regularization. First, based on the fact that similar patterns are more likely to appear in adjacent frames than in the current image, we fuse the adjacent frames information of the target image, and dynamically divide the image into overlapping image patches along the time dimension. Then, an asymptotic nonlocal algorithm is adopted to assign different weights to image patches, which is used to distinguish the representation contribution to local image patches. Finally, the alternating direction method of multipliers (ADMM) and the augmented lagrangian method (ALM) are used to solve the convex optimization problem of the low-rank super-resolution algorithm. Through comparative experiments on the all-sky auroral image dataset, it is shown from both subjective and objective aspects that this method can retain details and boundary information, and is superior to the current mainstream super-resolution models.