Estimation of Noise Using Non-local Regularization Frameworks for Image Denoising and Analysis

Abstract

In this paper, we propose a novel model which adaptively estimates the noise probability distribution and noise parameters from the input image and restores the data accordingly choosing appropriate regularization model designed for it. In most imaging applications the noise characteristics are assumed prior to the restoration process. This assumption is generally based on the previous experimental study of the images from a specific modality. The adaptive detection of the noise distribution from the data makes it robust and highly suitable for automated signal and image restoration systems. The non-local framework implemented using fast numerical solvers catalyzes the convergence rate of the model. Here we analyze three different noise distributions such as Gamma, Poisson, and Gaussian. Among this Gaussian is additive and source independent, Gamma is multiplicative and source dependent, and finally Poisson is data dependent (neither multiplicative nor additive). The model can be extended to the other source-dependent distributions such as Rayleigh and Rician by appropriately tuning it. The experimental results conform to the assumption regarding the noise distribution and noise parameters estimation capability of the model.