Light Field Image Quality Assessment Based on Disentangling Bayesian Theory

Quality assessment of light field image (LFI) has attracted increasing attention in recent years. In this paper, we propose a novel light field image quality assessment (LFIQA) method based on disentangling Bayesian theory that attempts to disentangle the coupling information from macro pixel image (MacPI) to accurately evaluate the entire LFI quality. Specifically, since the spatial and angular information of LFI are highly inter-twined with varying disparities, three targeted convolutions are designed to disentangle the spatial, angular and epipolar plane image (EPI) information in MacPI mode. Subsequently, the structural similarity (SSIM) maps are calculated between disentangled feature maps of the original and distorted LFIs. Furthermore, local binary patterns (LBP) operator and natural scene statistics (NSS) in two domains are used to extract the features on the SSIM maps as the quality-aware features of LFI. Finally, support vector regression (SVR) is used to simulate the nonlinear relationship between quality-aware feature and objective quality scores. Experiments show that the proposed model achieves high consistency with human's opinions, and outperforms multiple classical and state-of-the-art LFIQ models.