This article presents an unsupervised virtual reality (VR) image quality assessment (IQA) model based on the feature-fusion technique. A distilled feature selection approach is employed to obtain the optimal features of computationally efficient 2-D IQA models. Further, the obtained optimal features are used to compute quality-aware features from the viewports of the VR images. Principal components and projection matrices of each pristine viewport are obtained by principal component analysis (PCA) which are further used to obtain the projection vector of a test viewport feature set. Multivariate Gaussian modeling is performed to compute the mean vector and covariance matrices from the pristine and distorted projection vectors. The modified Mahalanobis distance is obtained to estimate the viewport-level quality score of a VR image from the computed mean vector, covariance matrices, and projection vector of the test viewport. Finally, the viewport level quality scores are spatially pooled with the location and content weights to estimate the perceptual deviation score in an omnidirectional VR image. Experimental results show that the proposed unsupervised 3-D VR IQA model outperforms existing completely blind 2-D IQA models and shows competitive performance to 2-D and 3-D supervised VR IQA models.