Study On ICA Algorithms And Its Applications To Image Denoising

Image noise seriously affects the acquisition of image information.As a pre-process,image denoising is very important.Recently,due to good properties in theory,denoising methods based on sparse coding have been drew much attention.Highly related with sparse coding,independent component analysis(ICA)is a widely used blind source separation method,which can obtain original source signals while the transmission channel characteristics are unknown.This method has been used in audio separation,biomedical science,wireless communication system,etc.Especially,ICA methods can extract image features or information connected with high order moments.When combined with shrinkage estimation,ICA can also be used to image denoising.This thesis focus on the performance of ICA based denoising methods,and the main works are summarized below:Firstly,the thesis reviews the development history of ICA.Basic theory and method,such as the basic model,model restrictions and ambiguities,pre-process and optimization methods are introduced briefly.Some classical ICA algorithms are shown,too.Secondly,noisy ICA model and its connection with sparse coding are analyzed.Feature extraction performances of ICA algorithms based on different nonlinear functions or different optimal algorithms are compared.Experiment results show that,the chosen of nonlinear functions affects the feature extraction performance more.Methods of which nonlinear functions are conducted by super-Gaussian density can extract features localized in space,frequency,and orientation.While the extracted features based on ICA algorithms with sub-Gaussian density do not have these properties.Besides,in the applications of image denoising based on shrinkage estimation,the feature coefficients' density modeled by sup-Gaussian density could get good denoising performance.Finally,the thesis considered the basic idea of two methods which are based on the self-similarity and redundancy property of images,nonlocal mean methods and simultaneous sparse coding model with Gaussian scale mixture methods.Combined self-similarity with ICA features and Laplacian scale density,a simultaneous sparse coding model with Laplacian scale mixture is proposed.Compared with shrinkage estimation,nonlocal mean methods or the Gaussian scale mixture with a few iterations,the new model can denoise better.