Performance evaluation of non-local means (NLM) Algorithms for grayscale image denoising

Non-local Means (NLM) algorithms are state of the art nonlinear techniques for image denoising. Although computationally intensive, NLM methods have gained growing attention due to their interesting combination of noise removal and detail preservation. This paper aims at providing a quantitative evaluation of these key filtering features without resorting to existing metrics. For the first time, the true values of filtering distortion and residual noise are formally obtained from NLM theory. Many computer simulations dealing with test images corrupted by Gaussian noise are reported in the paper in order to study how the filtering errors depend upon different parameter settings. Error maps are also provided in order to show the exact location and nature of filtering effects at the pixel level.