Locally adaptive multiscale Bayesian method for image denoising based on bivariate normal inverse Gaussian distributions

Recently, the use of wavelet transform has led to significant advances in image denoising applications. Among wavelet-based denoising approaches, the Bayesian techniques give more accurate estimates. Considering interscale dependencies, these estimates become closer to the original image. In this context, the choice of an appropriate model for wavelet coefficients is an important issue. The performance can also be improved by estimating model parameters in a local neighborhood. In this paper, we propose the bivariate normal inverse Gaussian (NIG) distribution, which can model a wide range of heavy-tailed to less heavy-tailed processes, to model the local wavelet coefficients at adjacent scales. We will show that this new statistical model is superior to the conventional generalized Gaussian (GG) model. Then, a minimum mean square error-based (MMSE-based) Bayesian estimator is designed to effectively remove noise from wavelet coefficients. Exploiting this new statistical model in the dual-tree complex wavelet domain, we achieved state-of-the-art performance among related recent denoising approaches, both visually and in terms of peak signal-to-noise ratio (PSNR).