A video restoration system called PriorPool will learn from a database of high-quality footage to clean up distorted videos when no perfect reference exists. Poor lighting, heat haze, or bad weather can ruin video footage, but the original scene is usually lost—so there is no "ground truth" to train a restoration algorithm. PriorPool solves this by building a large library of priors: patterns, textures, and structures extracted from pristine videos that share content with the damaged footage. The system retrieves the most relevant priors and uses them as constraints to guide unsupervised learning, restoring clarity without needing an undamaged version of the same scene. If successful, the framework could sharpen footage from cell microscopy, surveillance cameras, space imaging, and autonomous vehicle sensors—any setting where conditions degrade video quality. The project initially targets natural history filmmaking, a sector where Bristol produces over 40% of the world’s content. Cleaner video means better interpretation by both humans and machines, from tracking wildlife in dim rainforests to spotting obstacles in fog. The approach is fundamental: it tackles blind inverse problems where the type of distortion is unknown, a common hurdle in real-world imaging.
View original technical description
Acquiring high-quality footage in challenging environments, such as low light, heat haze, and adverse weather conditions, presents significant challenges. These conditions not only result in visually unpleasing videos but also make interpretation difficult for both humans and machines. Consequently, post-processing becomes necessary. However, video restoration and enhancement are challenging due to the loss of information. Moreover, ground truth is generally unavailable. Therefore, this research project, PriorPool, aims to address these issues in a novel manner. We propose that prior information extracted from high-quality videos, which share similar content with the distorted videos, can serve as constraints during the learning process of modelling algorithms. This approach allows us to leverage the inherent characteristics and knowledge embedded in high-quality videos, providing valuable guidance for the learning-based restoration and enhancement of distorted videos. PriorPool project aims to develop a comprehensive framework for video restoration and enhancement by addressing blind inverse problems with unsupervised learning. The specific objectives are as follows. To define and acquire a comprehensive database that includes priors relating to high-quality videos serving as references for enhancing distorted videos. To develop a new robust high-level representation of the video content. Distortions will generally alter video characteristics, increasing the difference between the input videos and the corresponding high-quality videos in the database, even if they have the similar content. This will minimise this gap to maximise the accuracy of acquired priors. To develop a prior retrieval system, providing global, local, and context-based priors, along with statistically driven models that provide a reliable basis for video restoration and enhancement process. To address blind inverse problems, where the degradation process during video acquisition is unknown. We will define a network to learn distortion functions from data that simultaneously inform the optimisation in Objective 5. To develop and refine optimisation and learning strategies that are aware of the acquisition context and capable of learning without explicit ground truth information. The aim is to enhance video quality using unsupervised learning approaches. The enhancement of distorted video is important in a number of fields including cell microscopy, space imaging, industrial metrology, surveillance, robotics and autonomous vehicles. While any solutions will have broad applicability, in this work we will initially target natural history filmmaking in challenging environments. The creative industries are a strength in the UK economy and Bristol leads the world, known as the Green Hollywood for natural history content, responsible for well over 40% of the world's productions.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know