NSS Models for Detecting Picture Artifacts
Netflix and other video content providers are tasked with delivering top-notch video quality to hundreds of millions of subscribers. As these providers continue to increase the sizes of their collection, a substantial percentage of the acquired video content will contain visual artifacts produced at the time of the video's production. These artifacts can include de-interlacing errors, up-sampling distortions, and other annoying visual defects that could greatly reduce the perceptual quality and ultimately the quality of experience of the subscriber/viewer. Finding this low-quality content, which is called “source inspection,” is valuable for improving the quality of the Netflix catalog.
Ph.D. student Todd Goodall and Prof. Alan C. Bovik are developing methods based on Natural Scene Statistics (NSS) models for detecting picture artifacts without requiring a reference signal. NSS-based image quality prediction models are regarded as the state-of-the-art for conducting reference-free picture quality prediction on 2D and 3D pictures and videos. The research team’s recent efforts have focused on developing specialized NSS-based methods for detecting/predicting upscaling and telecine artifacts. The goal of this research is to create a framework by which Internet video providers such as Netflix can make better decisions as they add to their sizeable collections.