Dual-branche attention network for super-resolution of remote sensing images

Remote sensing (RS) images are considered to be reflections of the real world. However, RS images often suffered from low resolution, making further research difficult to follow. Although super resolution (SR) techniques based on deep learning have achieved considerable breakthroughs, they show limited performance when dealing with low-quality RS images with complicated backgrounds; for instance, the SR results tend to loss details and have undesired structural distortion. Thus, this paper proposes an innovative dual-branch attention network (DBAN) to produce sufficient details and preserve clear structural information for SR results of RS images. It consists of two components: a feature extraction branch and a high-frequency information learning branch. The features extraction branch, formed as a densely residual structure, combines a series of dual attention blocks that are designed to exploit valid features from different dimensions, and then all these multi-scale features are reused through a global concatenation. The high-frequency information extraction branch, incorporating noise removing units (NRU) and high-frequency attention units (HFU), is responsible for producing the high-frequency features without noise, which enables DBAN to handle the problem of structural distortion. Meanwhile, a composite loss function based on a Laplacian pyramid is proposed to maximize the structural similarity between reconstruction results and real high-resolution RS images. The proposed network is efficient and lightweight because of its strong and effective attention to feature learning. Experimental results on three open-source RS image datasets and the JiLin-01 dataset demonstrate the effectiveness of our DBAN where higher accuracy over state-of-the-art methods in super-resolving complicated images was achieved.