Blind video quality assessment using natural video spatio-temporal statistics

In this paper, we propose a natural video statistics (NVS) based VQA model by making use of the measurable deviations in spatio-temporal statistical regularities. Both spatial and temporal statistical features in natural videos (NVS-S and NVS-T) are extracted and synthesized to get the video quality score. NVS-S is similar to what have been used in traditional NSS-based IQA or VQA in spatial domain. For NVS-T, we mainly study the distribution of motion residual DCT coefficients and the corresponding high frequency band energy ratio, which show great correlationship with video temporal distortions. The proposed blind VQA model is trained and tested in the LIVE database and EPFL-PoliMI database. The experimental results indicate that our model outperforms other blind VQAs.