Image denoising based on nonsubsampled shearlet transform domain Laplacian mixture model and bilateral filter and its method noise thresholding

Abstract

In this paper, we present a spatially-adaptive non subsampled shearlet transform (NSST) based Bayesian technique for denoising the natural images. The NSST can provide better directional selectivity and give nearly optimal approximation for a piecewise smooth function and is approximately shift invariant. The NSST coefficients of the images are modeled as 2-state Laplacian mixture(LM) distribution that uses local parameters. Through Kolomogrov-Smirnov (KS) goodness of fit test, it is shown that a 2-state LM distribution is highly appropriate for modeling the NSST image coefficients. This prior distribution is then employed to obtain a maximum a posteriori (MAP) estimator. An adaptive bilateral filtering (ABF) is applied on the MAP output to smooth out the visual artifacts in the homogeneous regions while preserving the textures. The NSST domain MAP estimator with LMM prior followed by ABF is finally implemented in its method noise thresholding framework for image denoising. The performance of proposed denoising framework is validated on a variety of benchmark images at different noise levels. The proposed framework achieves highly encouraging results in terms of fine detail preservation and exhibit less distortion compared with some recent relevant methods and BM3D method.