A lorentzian stochastic estimation for video super resolution with lorentzian gradient constraint

Abstract

In this paper, a novel super resolution (SR) framework is proposed to protect flat regions and edges of the reconstructed high resolution (HR) image simultaneously. In order to remove outliers and constrain the smoothness of the reconstructed HR image, the Lorentzian stochastic estimation is used for measuring the difference between the estimated HR image and each low resolution (LR) image. Moreover, this paper proposes a new regularization item, termed as Lorentzian gradient constraint, which incorporates with bilateral total variation (BTV) to enhance edges and keep flat regions of the reconstructed HR image. The combination of the two regularization items is superior to existing methods only based on BTV because it considers the balance between eliminating outliers and preserving details. Experimental results are presented to show the image quality and practical applicability of the new SR framework, and additionally demonstrate its superiority to existing SR methods.