Attention-based Residual Network for Single Image Remote Sensing Super-resolution

Single image super-resolution (SISR) reconstruction is crucial in meeting the growing demand for remote sensing imaging applications that need high spatial resolution. With the advancement of deep convolutional neural networks (CNNs), remote sensing SR has received considerable attention and has shown promising performance in the recent years. An effective feature extraction approach of CNN-based SR methods determines the quality of reconstructed images as well as increase the representation ability of CNN by more accurately extracting feature abstraction. In order to enhance the representational ability of CNN, a deep residual spatial and channel squeeze-and-excitation (RSCSE) SR network is proposed for remote sensing images (RSIs) to extract the deeper features in terms of channel-wise and spatially. Furthermore, a local feature fusion approach is incorporated in RSCSE module to preserve local features adaptively. Experiments conducted on two remote sensing datasets demonstrate that the proposed RSCSE network performs better than state-of-the-art methods both visually and quantitatively.