FUSION OF LOCAL DEGRADATION FEATURES FOR NO-REFERENCE VIDEO QUALITY ASSESSMENT

We propose a blind/No-Reference Video Quality Assessment (NR-VQA) algorithm using models for visibility of local spatio-temporal degradations. The paper focuses on the specific degradations present in H. 264 coded videos and their impact on perceived visual quality. Joint and marginal distributions of local wavelet coefficients are used to train Epsilon Support Vector Regression (epsilon-SVR) models for specific degradation levels in order to predict the overall subjective scores. Separate models for low/medium/high activity regions within the video frames are considered, inspired from the nature of H. 264 coder behavior. Experimental results show that blind assessment of video quality is possible as the proposed algorithm output correlates highly with human perception of quality.