Non-local medians filter for joint Gaussian and impulsive image denoising

Abstract

Image denoising concerns with the development of filters to remove or attenuate random perturbations in the observed data, but at the same time, preserving most of edges and fine details in the scene. One problem with joint additive Gaussian and impulsive noise degradation is that they are spread over all frequencies of the signal. Hence, the most effective filters for this kind of noise are implemented in the spatial domain. In this paper, we proposed a Non-Local Medians filter that combine the medians of every patch of a search window using two distinct similarity measures: the Euclidean distance and the Kullback-Leibler divergence between Gaussian densities estimated from the patches. Computational experiments with 25 images corrupted by joint Gaussian and impulsive noises show that the proposed method is capable of producing, on average, significant higher PSNR and SSIM than the combination of the median filter and the Non-Local Means filter applied independently.