A novel hierarchy features and matching based no-reference image blur assessment framework

The proposed framework is to treat the process of no-reference blur assessment as a matching process between the psychophysical features of the test image and memory existing common representations of clear images. The proposed framework has three key stages: knowledge of reality accumulation stage, hierarchy features extraction stage, and feature matching and blur metric computation stage. The knowledge of reality accumulation stage is to extract common features of clear images such as natural image statistics by learning. The hierarchy features extraction stage is to simulate the relevant components of human visual system to extract observable or unobservable visual features to represent an image; three levels of features such as low-level features, middle-level features and latent levels of features are proposed. The feature matching and blur metric computation stage is to compute the blur metric by the matching degree between the representation of the test image and the accumulated knowledge. Experimental results on synthetic images and public available images show that the proposed framework has better performance on monotonicity and anti-noise ability, and are also consistent with human visual system.