Motion measurement and quality variation driven video quality assessment

Many video quality assessment (VQA) methods generally use frame difference to represent motion information which may cause foreground apertures. Besides, quality variations of video contents significantly affect visual quality predictions. To remedy these, we develop a novel video quality assessment model that considers the impact of motion estimation and quality variations on distortion perception. The model firstly decomposes a video into a reflectance component and an illumination component via the intrinsic decomposition, and extracts scene features from these two components. To alleviate the foreground apertures, a simple structure difference is calculated as the motion measurement between adjacent frames. Motion difference features are extracted from the structure difference. We additionally consider analyzing quality variations to boost video quality predictions. Finally, a support vector regressor (SVR) is used to map generated features to predicted quality scores. We evaluated our proposed model on the LIVE, CSIQ, CVD2014 and LIVE-VQC video quality databases. The results show that the proposed model achieved competitive performance in comparison with state-of-the-art methods. Source code is freely available at: https://github.com/Aca4peop/QVDVQA.

Automated summary

The paper proposes a novel video quality assessment model that considers the impact of motion estimation and quality variations on distortion perception. By decomposing videos into reflectance and illumination components, extracting scene features, calculating structure differences to mitigate foreground apertures, and analyzing quality variations, the model aims to improve video quality predictions.