Current blind image quality prediction models rely on benchmark databases comprised of singly and synthetically distorted images, thereby learning image features that are only adequate to predict human perceived visual quality on such inauthentic distortions. However, real world images often contain complex mixtures of multiple distortions. Rather than a) discounting the effect of these mixtures of distortions on an image's perceptual quality and considering only the dominant distortion or b) using features that are only proven to be efficient for singly distorted images, we deeply study the natural scene statistics of authentically distorted images, in different color spaces and transform domains. We propose a feature-maps-driven statistical approach which avoids any latent assumptions about the type of distortion(s) contained in an image, and focuses instead on modeling the remarkable consistencies in the scene statistics of real world images in the absence of distortions. We design a deep belief network that takes model-based statistical image features derived from a very large database of authentically distorted images as input and discovers good feature representations by generalizing over different distortion types, mixtures, and severities, which are later used to learn a regressor for quality prediction. We demonstrate the remarkable competence of our features for improving automatic perceptual quality prediction on a benchmark database and on the newly designed LIVE Authentic Image Quality Challenge Database and show that our approach of combining robust statistical features and the deep belief network dramatically outperforms the state-of-the-art.
