Quality over Quantity: Enhancing and Assessing Image and Video Quality

In modern society, images and videos have a strong influence on how people perceive and engage with the world, as they play a pivotal role in shaping communication, education, and entertainment. In this context, the visual quality of the content is critical, since low-quality media can negatively impact the user experience, thus decreasing user satisfaction and engagement. This emphasizes the importance of reliable approaches to improve and assess multimedia quality. Various factors, including low resolution, noise, motion blur, poor lighting, and compression artifacts, contribute to visual degradation, hindering media interpretability and aesthetic appeal. Image and video restoration methods aim to enhance the perceptual quality of multimedia by removing visual distortions to recover clean content from its degraded counterpart. Evaluating the performance of such approaches requires methods for assessing the quality of images and videos. Subjective strategies involve collecting multiple human evaluations, making them expensive and time-consuming. To address this issue, No-Reference Image Quality Assessment (NR-IQA) methods aim to automatically predict the image quality in accordance with human judgments without the need for a reference image. Traditional NR-IQA methods rely on natural scene statistics, struggling with real-world images subjected to authentic distortions. Modern NR-IQA approaches employ deep learning on labeled datasets in a supervised way, thus requiring a large amount of costly annotated data. This thesis aims to design methods for enhancing and assessing the perceptual quality of images and videos. First, we address the video restoration task, with a specific focus on two applications, namely videoconferencing and analog videotapes. In particular, we develop reference-based approaches that take advantage of the highest-quality frames of a video to restore lost details in the degraded frames. Our experiments reveal that popular NR-IQA metrics struggle to predict quality ratings that correlate with human judgments. Therefore, we propose two NR-IQA metrics that leverage recent advancements in self-supervised learning and vision-language models to remove the need for expensive labeled data and to predict quality scores that better correlate with human perception.