Abstract
Among the various means to evaluate the quality of video streams, light-weight No-Reference (NR) methods have low computation and may be executed on thin clients. Thus, these methods would be perfect candidates in cases of real-time quality assessment, automated quality control and in adaptive mobile streaming. Yet, existing real-time, NR approaches are not typically designed to tackle network distorted streams, thus performing poorly when compared to Full-Reference (FR) algorithms. In this work, we present a generic NR method whereby machine learning (ML) may be used to construct a quality metric trained on simplistic NR metrics. Testing our method on nine, representative ML algorithms allows us to show the generality of our approach, whilst finding the best-performing algorithms. We use an extensive video dataset (960 video samples), generated under a variety of lossy network conditions, thus verifying that our NR metric remains accurate under realistic streaming scenarios. In this way, we achieve a quality index that is comparably as computationally efficient as typical NR metrics and as accurate as the FR algorithm Video Quality Metric (97% correlation).