Quality Assessment of Blur Images Via Saliency and Multiscale Features

Abstract

Blur is the most common distortion in images, so it is significant to evaluate the quality of blur images. Considering the view saliency and multiscale characteristics of human vision, this paper proposes a fast and efficient blur quality assessment method. Firstly, spatial sharpness and saliency maps are extracted, and the saliency map is used as the weight of the sharpness map. At the same time, images are downsampled to multiscales and the texture naturalness of the image on each scale is measured. Then all features are used to train the support vector regression model. Finally, the model is employed to evaluate the quality of blur images. Experimental results conducted on four image databases demonstrate that comparing with existing blur quality assessment methods and general quality metrics, the presented method has obvious superiority on performance and computation speed.