Image Denoising Research Based On Statistic Learning

In recent years, denoising methods via dictionary learning has become a hot topic in the image processing field. Most of the developed denoising methods are based on the statistics learning theory, which learn adaptive dictionaries from a large amount of image patches and produce effective denoised images.Based on the research on sparse representation and over-complete dictionary learning, several dictionary learning methods are discussed, including KSVD algorithm, online dictionary learning, non-parametric Bayesian dictionary learning and so on. Based on the analysis of self-similarity of images, an image denoising method based on dictionary learning and self-similarity is proposed. Based on the analysis of multiscale transform, an image denoising method via dictionary learning, self-similarity and multiscale transform is proposed. Moreover, combining with the recently proposed denoising algorithm of BM3D and multiscale transform methods, a denoising method based on dictionary learning and local constraint is presented. The contributions of our work is summarized as follow:(1) A natural image denoising method based on self-similarity regularization and dictionary learning is proposed. By combining the global dictionary learning with nonlocal self-similarity strategy, we design a self-similarity regularizer of image. The method can overcome the disadvantage of KSVD dictionary learning method that does not fully use the self-similarity structure of the image. The efficiency of the proposed method is demonstrated through experiments on natural images. Some experiments are taken on natural images and the experimental results show that our methods removes the noise in homogeneous regions effectively and simultaneously graciously retains the the texture information, the contour as well as the detail information of edge, and the higher evaluation guidelines are also obtained.(2) A natural image denoising method based on self-similarity regularizer and dictionary learning is proposed, based on the multiscale transform. Several multiscale multiscale coefficients images can be obtained from the nature image by a multiscale transform. At each scale, different residual is used to denoised, the image using the dictionaries and self-similarity based method, which can lead to an improved denoising result. The proposed method can overcome the disadvantage of the single dictionary learning method. Some experiments are taken on natural images and the result shows the feasibility of our method and its improvement over natural images on restoring the tiny texture information of original images as well as the smoothness of uniform area, and the evaluation guidelines also proves its efficiency. Compared with the sparse representation and dictionary learning denoising method on single scale in the third chapter, the proposed method is more efficient.(3) Based on the proposed method in the third chapter, by using the advantage that BM3D uses block matching to mine self-similarity of an image, we propose a method based on multiscale dictionary learning and BM3D constrained regularization for image denoising. We establish a denoising model that includes constraints derived from the denoising result of BM3D, and analyze the calculation of the regular factor. Using the multiscale transform introduced in the fourth chapter, the multiscale coefficients are denoised scale by scale. The method can overcome the disadvantages that the single scale dictionary learning cannot make full use of the self-similarity information and the structure information. Some experiments are taken on natural images and the experimental results show that the proposed method improves the smoothness of homogeneous regions simultaneously retains the textures, contours, edges and so on. Compared with the dictionary learning denoising method, the BM3D algorithm and the methods proposed in the third chapter and the fourth chapter, the visual results and the evaluation guidelines are improved significantly.