No-Reference Video Quality Assessment (VQA) Using Novel Inter Sub-band 3-D DWT Features

Abstract

Major goal of blind video quality assessment (VQA) is to predict visual quality of videos for enhancing the quality of service (QoS) without any reference. However, the conventional VQA model uses video as a two-dimensional image sequence and extracts the features on a frame to frame basis; which completely neglects temporal nature of the video data and the corresponding distortion content as well. In this work, we come up with a novel no-reference VQA model that describes and exploits the inter sub-band statistics of the three-dimensional discrete wavelet transform (3-D DWT) coefficients of video blocks. First, the 3-D DWT transform decomposes the video block of selected size into eight 3-D DWT sub-bands. Then we propose various novel statistical features using the sub-band coefficients. Inter sub-band statistics depicts the spread of the various frequency components and correlation between them. 3-D DWT features automatically take care of the temporal distortions along with spatial distortions and subsequently the support vector regression (SVR) model estimates them to predict the visual quality score of distorted videos. Experimental results on LIVE database demonstrate the superiority of the proposed VQA model over the other state-of-the-art methods.