An Improved Non-Local Filter for Image Denoising

Abstract

An improved image denoising algorithm based on nonlocal means (NLM) framework is proposed in this paper. The NLM estimates pixel values by a weighted mean of pixels in the corresponding searching area. The weight of each pixel is computed by a similarity function. The main idea of NLM is a searching area and a similarity function. The similarity function is a Euclidean distance function which is weighted by the standard Gaussian kernel. We modified the similarity function to become a flexible Euclidean distance function depending on image noise to balance smoothness in homogeneous areas and importance of the center pixel. We extend the searching area to the searching space, which not only covers a block of the image but also includes a dimension of pixel value space. This extension renders our proposed approach to exclude most of irrelevant pixels in the searching area which often disturb thin structures.