Abstract:
The goal of zero-shot image super-resolution (sr) is to generate high-resolution (hr) images from never-before-seen image distributions. This is challenging, especially, because it is difficult to model the statistics of an image that the network has never seen before. Despite deep convolutional neural networks (cnn) being superior to traditional super-resolution (sr) methods, little attention has been given to generating remote sensing scene-based hr images which do not have any prior ground truths available for training. In this paper, we propose a framework that harnesses the inherent tessellated nature of remotely images using continuity to generate hr images that tackle atmospheric and radiometric condition variations. Our proposed solution utilizes self tessellations to fully harness the image heuristics to generate an sr image from a low resolution (lr) input. The salience of our approach lies in a two-fold data generation in a self-preservation case and a cascaded attention sharing mechanism on the latent space for content preservation while generating sr images. By learning a mapping from lr space to sr space while keeping the content statistics preserved helps in better quality image generation. The attention sharing between content and tessellations aids in learning the overall big picture for super-resolution without losing an eye on the main image to be super-resolved. We showcase our results with the generated images given the low resolution (lr) input images in zero-shot cases comparable to state-of-the-art results on eurosat and patternnet datasets with metrics of ssim and psnr. We further show how this architecture can be leveraged for non-remote sensing (rs) applications. © 2021, springer nature switzerland ag.