Exact Computation of Residual Noise and Collateral Distortion in Nonlinear Image Filtering: A Case Study

Residual noise and collateral distortion are two key features for any image denoising filter. The former is the amount of noise still affecting the data after filtering, the latter represents the price to be paid in terms of detail blur. Measuring these effect is of paramount importance for the validation of a denoising algorithm. This work focuses of Non-Local Means (NLM) filtering that represents one of the most effective approaches to grayscale image denoising. The exact values of residual noise and collateral distortion are derived from NLM theory and an in-depth analysis of these features is provided for different input data and different parameter settings.