Revealing More Details: Image Super-Resolution for Real-World Applications

Digital images have become an integral part of our daily lives, but their quality may be compromised due to technical limitations. This PhD thesis, conducted from 2020 to 2023, focuses on investigating novel methods to enhance the visibility and quality of such images through Super-Resolution (SR). The goal of SR is to restore high-resolution details from low-resolution images. State-of-the-art deep-learning-based SR models utilize learned priors, which can lead to unwanted artifacts when the input image resides outside the training distribution. Consequently, there is a growing interest in developing methods that generalize to real-world images. In this PhD thesis, we continued this line of research and explored even more challenging scenarios. Speciőcally, we investigated SR of face images from surveillance footage and proposed a method to handle the artifacts present in such images, which obtained improved performance. Considering the scarcity of paired real-world low-resolution and high-resolution training image pairs, we explored the use of semantic segmentation guidance, which yielded highly improved results. Additionally, we developed a method to estimate per-pixel degradations and adapt the SR process accordingly, addressing the challenge posed by images with spatially variant degradations. To facilitate objective evaluation, we assembled the őrst dataset comprising such images and ground truths. Our method showcased superior performance when evaluated on this demanding dataset. Furthermore, we investigated SR of images degraded by both lowlight and low-resolution, for which we curated another comprehensive …

Recommended citation: Aakerberg, Andreas (2023). "Revealing More Details: Image Super-Resolution for Real-World Applications." .
Download Paper