A no-reference Image sharpness metric based on structural information using sparse representation

As a ubiquitous image distortion, blur casts non-trivial influence on image visual quality. Many image sharpness assessment methods have been proposed in the views of edge information, gradient map, frequency spectrum, or other natural image statistics features. In this paper, we propose a no-reference image sharpness metric based on structural information using sparse representation (SR). We observe that the dictionary atoms learned by SR algorithm convey clear structural information. Considering the distinct sensibility of human visual system (HVS) to different structures, we use the learned dictionary to encode the patches of the blurry image. To embed the locality of the representation, a multi-scale spatial max pooling scheme is incorporated. The final sharpness score is given by an efficient linear support vector regression (SVR) model. We evaluate our approach on three public databases, i.e., LIVE II, TID2008, and CSIQ. The experiments demonstrate that the proposed method achieves competitive performance compared with the state-of-the-art blind image sharpness assessment algorithms.