Blur Measurement for Partially Blurred Images with Saliency Constrained Global Refinement

Blur measurement of partially blurred image is still far from being resolved. This calls for more distinctive blur features and, even more importantly, a global refinement strategy that has not been considered by existing studies. In this paper we propose a new spatial and frequencial coupled blur descriptor by composing the number of extreme points, the vector of all singular values and the entropy-weighted pooling of the high frequency DCT coefficients. We also introduce a global refinement scheme to explore the merits of saliency for further refining the initial measurements. Consequently, we propose a novel saliency constrained blur measurement method by integrating a neural network based blur metric and a superpixel-scale blur refinement together. Experimental results show the efficiency of our method qualitatively and quantitatively, especially for the images with flat textures.