Inspection and evaluation of artifacts in digital video sources

Abstract

Streaming digital video content providers such as YouTube, Amazon, Hulu, and Netflix collaborate with production teams to obtain new and old video content. These collaborations lead to an accumulation of video sources, some of which might contain unacceptable visual artifacts. Artifacts may inadvertently enter the video master at any point in the production pipeline, due to any of a number of equipment and user failures. Unfortunately, these artifacts are difficult to detect since no pristine reference exists for comparison. As of now, few automated tools exist that can effectively capture the most common forms of these artifacts. This work studies no-reference video source inspection for generalized artifact detection and subjective quality prediction, which will ultimate inform decisions related to acquisition of new content.

Automatically identifying the locations and severities of video artifacts is a difficult problem. We have developed a general method for detecting local artifacts by learning differences in the statistics between distorted and pristine video frames. Our model, which we call the Video Impairment Mapper (VID-MAP), produces a full resolution map of artifact detection probabilities based on comparisons of excitatory and inhibatory convolutional responses. Validation on a large database shows that our method outperforms the previous state-of-the-art of even distortion-specific detectors.

A variety of powerful picture quality predictors are available that rely on neuro-statistical models of distortion perception. We extend these principles to video source inspection, by coupling spatial divisive normalization with a series of filterbanks tuned for artifact detection, implemented using a common convolutional framework. We developed the Video Impairment Detection by SParse Error CapTure (VIDSPECT) model, which leverages discriminative sparse dictionaries that are tuned to detect specific artifacts. VIDSPECT is simple, highly generalizable, and yields better accuracy than competing methods.

To evaluate the perceived quality of video sources containing artifacts, we built a new digital video database, called the LIVE Video Masters Database, which contains 384 videos affected by the types of artifacts encountered in otherwise pristine digital video sources. We find that VIDSPECT delivers top performance on this database for most artifacts tested, and competitive performance otherwise, using the same basic architecture in all cases.