Machine Learning Based UHD Up-conversion Using Generative Adversarial Neural Networks

摘要

Ultra High Definition (UHD) is now part of mainstream TV production. Various consumerelectronic companies now offer a wide range of high quality UHD TV sets, which now account for asignificant proportion of all new TVs sold globally. The popularity of UHD has raised the viewer’sexpectations for a higher quality experience. To date, online providers have moved quickly to satisfythis UHD demand via Adaptive Bit Rate (ABR) delivery.However, there remains a limited amount of content available that is natively UHD and traditionalcontent providers have a wealth of great content that could be offering their consumers even morevalue in a UHD format. As consumer expectations grow, broadcasters will need to provide higherquality experiences, moving from a few UHD events to full-time or pop up channels, even if this isdelivered only over ABR infrastructure. The question about where to get valuable UHD content remainsopen-ended, even though traditional up-conversion techniques can result in an end user experiencethat is more ‘HD-like’ than UHD.This paper explores a different approach to up-conversion, using Generative Adversarial NeuralNetworks (GANs) to synthesize detail in the upconverted image, leading to an experience that is muchcloser to native UHD leading to more compelling, higher quality experiences for consumers.