Super-resolution reconstruction of remote sensing images based on Swin Transformer fusion attention network

Abstract

Image super-resolution reconstruction technology in remote sensing can improve the spatial resolution of remote sensing images with the breakthrough of physical hardware limitations. With the development of deep learning technology, more and more algorithms proposed in the field of natural images are applied to the field of remote sensing super-resolution. Due to the large difference in the size of the objects in remote sensing images and the high complexity of the image, the reconstructed image will be blurred when the algorithm in the field of natural images is directly used. To address this problem, this paper proposes a shallow feature extraction feature fusion with multiple convolutions, followed by the extraction of high-frequency information using the Swin Transformer module with a fusion attention mechanism. The edge details of the image are extracted using the gradient of the image in the final reconstruction process, and complementary fusion is performed at the end of the network, which can effectively supplement the lack of shallow features caused by the deep network. Finally, experiments show that the proposed model obtains satisfactory reconstruction results of remote sensing images.