A Deep Unfolding Method for Satellite Super Resolution

Abstract

Despite that existing deep-learning-based super resolution methods for satellite images have achieved great performance, these methods are generally designed to stack unaccountable and dense modules (i.e., residual blocks and dense blocks) to reach an optimal mapping function between low-resolution and high-resolution patches/images at the expense of computing resources. To address this challenge, we propose a deep unfolding method (LDUM) that includes two major components: 1) the pretreatment network and 2) the unfolding blocks. The main responsibility of the pretreatment network is to generate initial high-resolution images. Further, we model high-resolution images with the prior information, which can be seen as a combination of low-resolution and high-frequency residual images, and solve the optimization problem via the iterative proximal strategy. Specifically, we unfold the iterative process into a deep neural network to refine the reconstructed results, as each layer serves as an iterative step of the proposed optimization model. Thus, the proposed method is able to iteratively generate residual maps and high-resolution images by combining the powerful feature extraction capability of data-driven deep-learning-based methods and the interpretability of traditional model-driven algorithms. Experiments show that the proposed method, featured by its interpretable and lightweight merits, outperforms other state-of-the-art methods from quantitative and qualitative perspectives. The source code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/jiaming-wang/LDUM</uri> .