Brian Moser

• For decades, low-resolution images have captured the attention of viewers worldwide, highlighting a persistent need for advancements in digital imaging. For this reason, image Super-Resolution (SR) is a critical machine learning task that aims to enhance resolution, often by a factor of four or more. Such enhancements are vital across various application fields, from medical imaging to satellite imagery, from refining image quality to augmenting visual details. The core challenge is accurately reconstructing high-resolution images that maintain the authenticity and detail of the original, a challenge amplified by outdated imaging hardware or lossy compression methods. Nowadays, deploying advanced deep learning models that leverage extensive datasets to infer missing details accurately is at the heart of this research field. To further advance the field, this Thesis addresses three fundamental aspects: data representation, dataset quality, and the deep learning model itself. Regarding the first aspect, this Thesis explores the frequency domain as an alternative for representing data for both regression-based and generative SR models. By analyzing image data through its frequency components, SR models can more effectively emphasize and restore high-frequency details, contributing to sharpness and clarity. The second focal point of this Thesis is improving the quality of training datasets. It presents strategies for dataset optimization, including pruning to retain only the most impactful samples and introducing benchmarks to evaluate SR model performance in federated learning settings. Federated learning is particularly beneficial when data cannot be centrally collected. It allows collaborative model training across multiple decentralized clients without compromising data privacy. Lastly, the Thesis introduces innovative methods for generating high-resolution images using diffusion models. These include a technique for training diffusion-based SR models with a selective focus on detail-rich areas, thereby enhancing the ability to reconstruct more accurate and detailed high-resolution images. Moreover, it explores how dataset distillation can benefit from the insights of the image SR domain by including the latest state-of-the-art image generation methods and known SR concepts like latent alignment. By addressing these three dimensions, the Thesis contributes significantly to the field of image SR, advancing the capabilities of SR technologies and broadening their applicability.