Blind image quality assessment on real distorted images using deep belief nets

Abstract

We present a novel natural-scene-statistics-based blind image quality assessment model that is created by training a deep belief net to discover good feature representations that are used to learn a regressor for quality prediction. The proposed deep model has an unsupervised pre-training stage followed by a supervised fine-tuning stage, enabling it to generalize over different distortion types, mixtures, and severities. We evaluated our new model on a recently created database of images afflicted by real distortions, and show that it outperforms current state-of-the-art blind image quality prediction models.