Image denoising using Gabor filter banks

Abstract

We introduce a method for denoising a digital image corrupted with additive noise. A dyadic Gabor filter bank is used to obtain localized frequency information. It decomposes the noisy image into Gabor coefficients of different scales and orientations. Denoising is performed in the transform domain by thresholding the Gabor coefficients with phase preserving threshold and non-phase preserving threshold where both approaches have been formulated as adaptive and data-driven. For the non-phase preserving approach the BayesShrink thresholding methods have been used. Finally using the thresholded Gabor coefficients of each channel the denoised image has been formed. It has been found that for smoothly varying images the modified BayesShrink method outperforms both the BayesShrink and the phase preserving approaches whereas for images with high variations the phase preserving approach performs better.