Methods for Blind Super-Resolution of satellite images

Recently, deep neural networks have demonstrated remarkable efficiency at improving the resolution of low-resolution (LR) images. This is called the image super-resolution (SR) task. Numerous researchers have come up with network architectures to deal with the problem of multi-image super-resolution (MISR) by using supervised learning. This entails the availability of ground-truth high-resolution (HR) pictures for the purpose of training. However, collecting LR and HR images from the same device to avoid introducing any additional mismatch between images could be problematic, especially for satellite images, which involve cameras hundreds of kilometres from the subject. The goal of this thesis is to provide a method to deal with MISR in a blind, or unsupervised, setting by leveraging other well-known networks to make blur kernels from LR images and do super-resolution. Specifically, blur kernels were used to create a set of images called coarse-resolution (CR) images that have lower resolution than the LR images. The LR and CR images are then used to train a supervised MISR network. The backbone of this method is the MISR network PIUnet, proposed by Valsesia and Magli. At the same time, for the kernel estimation task, this thesis explores two different solutions: MANet, introduced by Liang et al., and the architecture DIP-FKP, also by Liang et al. Moreover, this thesis proposes a method to fine-tune the SR network PIUnet using a novel architecture inspired by DIP-FKP, the so-called PIUNET-FKP. Experiments on the Proba-V challenge dataset showed good results for the aforementioned technique in terms of corrected peak signal-to-noise ratio (cPSNR), moving a step towards the performance of the supervised case. This work encompasses several components, including a description of the devised methodologies and the outcomes achieved. Additionally, it provides an overview of the context of image and satellite image super-resolution, along with a literature review of the existing methodologies in the field. Furthermore, it presents a theoretical background on the instruments employed in the study.