Multilateral filtering: A novel framework for generic similarity-based image denoising

We present a novel iterative nonlinear filtering framework, termed multilateral filtering, based on the idea of generic local similarity. A set of local features is computed for each pixel using its local neighborhood. Two pixels are considered to be similar if the Euclidean distance between their corresponding feature vectors is small and vice versa. Multilateral filtering results in image smoothing while preserving edge and textural features. Our experimental results show that the proposed method produces comparable and often better results than the state-of-the-art denoising methods.