A Non-local Means Filter Image Denoising with Directional Enhancement Neighborhood Windows

Non-Local Means (NLM) filter is an effective method for image denoising. However, it only focuses on the geometry structure of image, ignoring the appearance model and directional information. In this paper, a new Non-Subsampled Shearlet Descriptor (NSSD) is proposed and employed to model the appearance of image patches and the measurement of similarity between two image patches is more robust. According to NSSD, a more effective Shearlet Non-Local Means (SNLM) algorithm is proposed by combining the NSSD with non-local computation model. For another, for texture images with directional information, a direction enhance window is proposed, which increases the weights on the main direction in the neighborhood window in the measurement of similarity. Experiment results show that the proposed NLM algorithm gets better performance on natural image denoising than the traditional NLM algorithm. Moreover, for texture image, the algorithm based on direction enhance neighborhood window can not only remove the noise but also preserve the detail information such as edges and textures and show great advantages on denoising.