A No-Reference Image Sharpness Metric Based on Large-Scale Structure

Abstract

Most of no-reference image sharpness metrics suffer interference of tiny textures and the contrast around edges. Based on large-scale structure, this paper proposes a novel metric to avoid this problem. Firstly prominent edges are extracted by a weighted least-squares filter. Then, we calculate the statistics proportion of edges with different width and the mean value of the maximum gradients from the 51th to 250th as the features of the image. Finally, the support vector regression is used to get the relationship model between the features and the subjective assessment result. The metric can reduce the influence of textures, contrast and the content of images effectively, accordingly enhancing the assessment accuracy.