Noise model discrimination for digital images based on variance-stabilizing transforms and on local statistics: Preliminary results

Most of the image restoration algorithms assumed the noise model and its parameters as an a priori information. Nevertheless this is not necessarily the case for real scenarios. Moreover, lack of knowledge about the noise parameters leads to heuristically approaches to choose the restoration algorithm's parameters. Given a non-texture observed image, which can be noise-free or corrupted with some kind of noise (we consider Gaussian, Poisson, Gamma and Rayleigh) we propose a simple yet effective method to discriminate the noise model (or lack of) that corrupts the observed image by first applying a set of variance-stabilizing transforms and then proceed to estimate the variance using a local statistics estimator; the estimated variance will be unitary only for the particular variance-stabilizing transform that matches the correct noise model.