Study On Several Issues Related To Image Denoising

Recently, digital images become a key media for information acquisition and transmission.However, images are easily corrupted by noise, such as white guassian noise. Thisdegradation leads to a significant reduction of image quality and then makes more difficult toperform high-level vision tasks such as segmentation, target classification and recognition, etc.Therefore, denoising is an essential step in image processing and analysis. During the pastthree decades, a variety of denoising methods have been developed in the image processingand computer vision communities.Nowadays, machine learning and statical model become more and more popular in theimage denoising.This paper’s theme is image denoising. TWSVM, HMT-MRF model and anew type of HMT combined with interscale, intrascale, across-scale model were studied andsummarized in this paper, completed the following work:1. Based on twin support vector machine (TWSVM) theory, we propose anon-subsample Shearlet-based image denoising using TWSVM. Firstly, the noisy image isdecomposed into different subbands of frequency and orientation responses using thenon-subsample Shearlet transform. Secondly, the feature vector for a pixel in a noisy image isformed by the spatial regularity in non-subsample Shearlet domain, and the TWSVM model isobtained by training. Then the non-subsample Shearlet coefficients are divided into twoclasses (noise coefficients and noise-free ones) by TWSVM training model.2. Based on extended Shearlet transform and Hidden Markov Tree (HMT) model, a newimage denoising using extended Shearlet domain HMT models is proposed. Firstly, theextended Shearlet transform is performed on the noisy image. Then, the extended Shearletcoefficients are modeled using HMT model, and the HMT model parameters are estimatedutilizing maximum posterior probability. Finally, the trained extended Shearlet coefficientsare transformed back into the original domain to get the denoised image.3. Based on Bessel K form(BKF) and Hidden Markov Tree (HMT) model, we propose aExtended Shearlet-based image denoising using a new type of HMT. Firstly,we model thesignificance of the noise-free extended Shearlet coefficients in a local window using a newsignificance measure. Secondly, we combine the interscale, intrascale, across-scale with anew hidden Markov tree model to capture the dependencies between the extended Shearletcoefficients.