Dual-Resolution Local Attention Unfolding Network for Optical Remote Sensing Image Super-Resolution

Abstract

Single image super-resolution technology based on deep learning is widely applied in remote sensing. In recent years, the deep unfolding super-resolution strategy has been proposed, which combines the neural networks with traditional optimization-based algorithms, making the neural networks interpretable and achieving high performance. However, the typical deep unfolding algorithms usually treat different kinds of blurring kernels in the same way, so the algorithms cannot take advantage of the properties of blurring kernels, limiting the algorithm’s performance. To design a super-resolution network that can fully use the properties of Gaussian blurring kernels, a dual-resolution local attention unfolding network (DLANet) is proposed. Based on the Gaussian blurring functions, a low-resolution (LR) space branch is designed to supplement the high-resolution (HR) space branch. Specifically, for Gaussian blurring kernels, the closer the pixel is to the center, the greater the weight is. It means that the pixel points retained after downsampling will contain more information about the original corresponding pixel points, and it could be easier to estimate their original pixel values. So we design two branches. The HR branch completes the estimation of the whole image, and the LR branch only estimates the points retained after downsampling. To better complete the feature fusion of the two branches, we propose a row-column decoupling local attention module. This module can retain more information when fuse features and the row-column decoupling strategy can reduce computational complexity. Comprehensive experiments demonstrate the superiority of our method over the current state-of-the-art on remote sensing datasets.