Variance Based External Dictionary for Improved Single Image Super-Resolution

Abstract

In this paper, we have proposed a novel method for image super-resolution using single image. Based on structural similarity index, an image similar to input low-resolution (LR) image is selected from the database and two separate external dictionaries i.e. smooth and textured, are formed from the selected image based on their variances. Different features are used for representation of different type of patches. For smooth patches norm luminance is used as feature vector and for textured patches it consist of first and second order gradients. In neighbor embedding process, a new parameter in combination with Euclidean distance has been introduced to eliminate outliers. Extensive simulations are performed to show superiority of the method.