Multiscale LMMSE based statistical estimation for image denoising

Abstract

Image Denoising is the process of removal of the noise from the image contaminated by additive Gaussian noise without loss of features of image. It is a fundamental process in pattern recognition and image processing.In this thesis,a wavelet based new denoising scheme for estimation of parameters such as variance of the multiscale Linear minimum mean square error(LMMSE) estimator to derive optimal threshold using maximum a posterior (MAP) estimator of the noisy coefficients in wavelet domain has been proposed. Our proposed scheme modify the parameter of LMMSE. Input image is decomposed in four wavelet subband then for each subband the LMMSE estimator is then applied.Denoised image is reconstructed after applying inverse wavelet transform. Each schemes is studied separately and experiments are conducted on test images to evaluate the performance.This denoising scheme shows the best performance for highly corrupted image in terms of the structure similarity index measure(SSIM)the peak signal-to-noise ratio (PSNR).