Edge-preserving wavelet thresholding for image denoising (Q2475370)

The problem of wavelet-based image denoising by means of thresholding wavelet image estimators is considered. The denoised image is obtained by minimizing the quality index being the sum of the thresholded image estimate error and a penalty term associated with a roughness of the obtained solution. The used penalty functions are based on the gradient magnitude of the image and correspond to potential functions applied in the non-linear diffusion filtering theory. It is shown that the proposed penalties and particular families of dyadic wavelets make it possible to preserve and restore by the penalized estimates both edges and smooth regions of the image. The simple denoising algorithm, consisting in a two-step iterative shrinkage procedure for each couple of dyadic wavelet coefficients, is provided and its convergence is justified. Simulation experiments for classical test images corrupted by white Gaussian noise with various noise levels show that the proposed shrinkage method outperforms the traditional techniques, both in the PSNR value sense and visual quality.