Research On Image Denoising Based On ICA Sparse Coding And Contourlet Transform

Image denoising is a classical and hot research problem in the field of signal processing. Noise is a major factor which influences on the quality of picture. It not only influences the visual effect but also gets huge interference when Image processing. In order to keep subsequent reliability, image need to deal with denoising in advance. When removing noise, the image need keep edge texture details of the original image as far as possible, as well as be denoised at the greatest extent. Therefore, image denosing is always a hot issue which researchers focus on.In this thesis, it mainly gets research and analysis on denoising arithmetic of image preprocessing, concentrates on studying the ICA sparse coding arithmetic and analyzing the contourlet transform, and develops correspond way to improve the arithmetic based on its shortage. The content of the thesis is mainly as follows:(1) This thesis studies the ICA sparse coding algorithm in image denoising and puts forward a modified ICA sparse coding algorithm according to the problems, such as being difficult to estimate the probability density function of the original signal, possibly resulting in blurred image details during soft threshold shrinkage function denoising process and so on. Through the establishment of Bayesian threshold in ICA transform domain and a new shrinkage function, this algorithm not only avoids the estimation of the original signal, but also it effectively solves the problem of blurred image details. Experimental results show that this algorithm not only has a low degree of computation complexity and is much easier to implement in practice, but also its denoising effect is better than that of traditional algorithm.(2) Based on the contourlet Transform, this thesis puts forward an improved threshold denoising arithmetic of the NSCT (Nonsubsampled Contourlet Transform). This algorithm builds a new Bayesian threshold, which can be adaptive in the concentration direction of sub-band energy and can better estimate the signal variance and noise variance. Referring to the result of the simulation experiment and compared to traditional NSCT threshold denoising arithmetic, the denoising effect and peak signal to noise ratio of the algorithm in this thesis are both significantly improved.(3) In high noise, the ability of ICA sparse coding is much better than the contourlet transform, while the result is the opposite in low noise. Therefore, the thesis designs a new image denoising method which combines ICA sparse coding and contourlet transform. For the problem that it is difficult to obtain the noise variance in practical engineering, the proposed method can achieve better denoising effect through calculating the standard deviation of the "flat" areas of image to estimate the magnitude of noise and selecting algorithm according to the standard deviation. Experimental results show that this method can precisely and adaptively choose denoising algorithm and achieve the goal of optimization. Therefore, it has a certain application value.