Anisotropic Smoothing Using Double Orientations

Gabriele Steidl, Tanja Teuber

Abstract: To improve the quality of image restoration methods directional information has recently been involved in the restoration process. In this paper, we propose a two step procedure for denoising images that is particularly suited to recover sharp vertices and X junctions in the presence of heavy noise. In the first step, we estimate the (smoothed) orientations of the image structures, where we find the double orientations at vertices and X junctions using a model of Aach et al. Based on shape preservation considerations this directional information is then applied to establish an energy functional which is minimized in the second step. We discuss the behavior of our new method in comparison with single direction approaches appearing, e.g., when using the classical structure tensor of Förstner and Gülch and demonstrate the very good performance of our method by numerical examples.