Image Despeckling Using Non-local Means with Diffusion Tensor

Abstract

This paper presents a novel modification of one of the varieties of the non-local means (NLM) algorithm for speckle reduction in images. This modification comes in the form of replacement of the structure tensor used in the NLM algorithm by the diffusion tensor. The diffusion tensor originally was used in the nonlinear coherent diffusion algorithm making possible intensification of the diffusion in the direction parallel to edges and inhibition in the direction perpendicular to edges. It is shown in this paper that using the diffusion tensor in calculating the weights for the NLM leads to an improvement of the quality of despeckled images. The NLM algorithm has a tendency to smooth the image in such a way that the despeckled image is covered by relatively flat areas typical for mosaic images. This tendency is undesirable since flat areas form visible contours that are not related to the object visualized. The superiority of the new despeckling filter is confirmed by examples of filtering the ultrasound (US) images as well as by image quality measures.