R&D engineer working on developing optimised ML solutions for image and video quality enhancements
Computer Vision research at the BBC: Restoring image and video content from our rich archives
BBC Archives manages one of the world’s largest multimedia archives, documenting BBC’s broadcasting history. Over 1.5 million video tapes and 7 million photographs are stored in numerous collections, which the BBC aims to preserve, catalogue, and digitalise. However, in some cases, the original content is of inadequate quality, containing noise and artefacts due to the technology it was recorded with or the limitations of older digitalisation techniques. Therefore, a restoration process is needed to convert the footage to modern broadcast standards. This process traditionally requires highly skilled manual intervention, which in some cases is too time-consuming and therefore expensive for productions to consider.
At BBC Research & Development, we are exploring how to automatise this restoration process to assist the skilled work of human restorers and allow productions to use a greater range of archive clips. We aim to do this using recent breakthroughs in Machine Learning (ML) and other image and video processing techniques. Specifically, we train neural networks to mitigate blur, noise, scratch and dust artefacts, and enhance old black & white content with ML colourisation techniques.