Comparison of nonlocal means despeckling based on stochastic measures

Abstract

This work presents the use of stochastic measures of similarities as features with statistical significance for the design of despeckling nonlocal means filters. Assuming that the observations follow a Gamma model with two parameters (mean and number of looks), patches are compared by means of the Kullback-Leibler and Hellinger distances, and by their Shannon entropies. A convolution mask is formed using the p-values of tests that verify if the patches come from the same distribution. The filter performances are assessed using well-known phantoms, three measures of quality, and a Monte Carlo experiment with several factors. The proposed filters are contrasted with the Refined Lee and NL-SAR filters.