No-Reference Virtual Reality Image Quality Evaluator Using Global and Local Natural Scene Statistics

With the rapid proliferation of virtual reality (VR) technologies, the usage of VR in multimedia, education, and social media platforms has increased due to realistic and immersive 3-D viewing experiences. In particular, VR refers to a computer-generated synthetic environment where the users can experience 180(degrees) x 360(degrees) spherical VR content through head-mounted displays (HMDs). Due to the 180(degrees) x 360(degrees) viewing range, the quality assessment (QA) of VR images becomes quite difficult compared to conventional 2-D image QA (IQA) models. To alleviate this problem, in this article, we propose a supervised no-reference (NR) VR IQA model based on global and local natural scene statistics (NSS) of a VR image. The global features are computed based on joint dependencies between adjacent pixels of equirectangular projection (ERP) VR image using generalized Gaussian distributions (GGDs). Specifically, we compute the model parameters of GGDs at multiple scales and show that the model parameters are distortion discriminable. Furthermore, we compute local features based on statistical properties of spatial and spectral entropy maps of cube map projection (CMP) faces of a VR image. Since the local feature extraction is carried out at the CMP face level, we compute the average of the face-level features to obtain the overall local feature set of a VR image. Finally, global and local feature sets are combined and given to support vector regressor (SVR) to map the quality-aware feature set to VR image quality with labels as human assessment scores. The performance of the proposed VR IQA model is verified on three omnidirectional VR IQA datasets, such as CVIQD, OIQA, ODIQA, and one stereoscopic VR IQA dataset, namely, LIVE S3D-VR. Experimental results show that the predicted scores of the proposed VR IQA model correlate very well with the subjective scores of the aforementioned VR IQA datasets and achieved state-of-the-art performance numbers compared to the existing 2-D and 3-D IQA models.