A Study Of Image Denoising Algorithm Based On Image Patches Prior And Bootstrap

Natural image denoising is a classical problem in image processing fields. The images we learned may contain lots of noise which originate from different noise source. Noise can seriously affect the quality of the images. In order to get reliable results in the latter image processing, it is necessary for ours to remove the noise from images while preserving the texture and edge detail. In recent years, the denoising algorithms using the image patch prior have attracted wide attention. These denoising methods are all based on the statistical characters of image. Zoran et al. proposed a restoration framework based on Expected Patch Log Likelihood. It realized the restoration of a whole image based on the restoration of image patches by using Gaussian mixture model. It has been proved that this algorithm can get effective denoising result. However, the algorithm is sensitive to noise and its parameters estimation method also needs to be modified.Two improved natural image denoising algorithms based on image patches are proposed in this paper. The main jobs of this paper are as follows:(1) In order to improve the accuracy of parameter estimation in the learning process of the EPLL algorithm, a statistical method called Bootstrap is introduced. Bootstrap can estimate the parameters just depending on the observation samples. However, the small sample size may lead the estimation of distribution deviate from the real distribution. To overcome this problem, we improve the Bootstrap method by defining the neighborhood of the sample. It could extent the scope of resampling to the non-observation area, accordingly expand the sampling capacity and improve the accuracy of estimation.(2) To solve the existing problems of the EPLL method, a denoising algorithm based on EM-Bootstrap parameter estimation and image patch adaption EPLL is proposed. Gaussian mixture model is used in the algorithm to learning the image patch prior. The traditional EM is sensitive to the initial parameter value and may lead local optimum problem. Combining the advantages of Bootstrap we propose a new parameter estimation method called EM-Bootstrap to solve the problem mentioned above. Firstly, we use the Bootstrap estimate the observation sample and get the initial parameters of GMM. Then, using the Bootstrap to resample and reestimating the parameter in the M step of EM method. After learning the prior, image patch adaption EPLL framework is use to denoising. Original EPLL ignores the statistical of the noisy image, and it doesn't update the parameter of model in the denoising processing. For this reason, we propose an image patch adaption EPLL to fit the noisy image. Comparing with correlation denoising methods, the method we proposed can get better denoising result while preserving well the texture and edge.(3) We propose an image denoising method which is based on the Spatial distribution factors and the Student-t mixture model. Student-t mixture model is introduced to the image patch adaption EPLL framework. To solve the problem that EPLL is sensitive to the noise, we introduce the information of spatial distribution to the Student-t mixture model. The experiment has proved that the Student-t spatial mixture model is more suitable for denoising by using the image patch adaption EPLL framework. The Improved Bootstrap method is used to estimate parameters. Experiments show this denoising method is better than other methods in preserving details.