Research Of Image Restoration Algorithm Based On Regularization

Quality of an image is degraded when there are some undesirable elements in imaging process and image transmission, which affects use and following processing of the image. It is paid more attention that how to restore the degraded image into a perfect image with clear and of rich content, and important demand in many fields. That is subject of image restoration, which includes denoising, deblurring, inpainting and super-resolution and so on. This thesis focuses on image denoising, and image deblurring research.Based on analyzing image degraded model, this thesis studies prcessing procedure of image restoration, reviews.four image restoration categories, which includes statistical signal processing based, partial differential equation based, and using potential functions of a Markov random fields model and using regularization method. Our work mainly is based on regularization method, which involves motion deblurring, image denoising and deblurring based on Field of Experts model (FoEs), and image restoration based on both image sparse representation and non-local information, respectively.Motion deblurring algorithms are researched in this thesis. On the basis of regularization method, both finding the solution of image restoration model equations and selecting regularization parameters are lucubrated. A conjugate gradient algorithm that is blocked symmetric successive over relaxation is presented in. finding the solution of image restoration model equations. And an adaptive selecting parameters method based on splitting regions is presented in selecting regularization parameters.Field of Experts model applied to image restoration is studied in this thesis, which is widely paid attention. A technical way to apply FoEs to image denoising is discussed, that is, the probabilistic models are built by using of FoEs model and solved following with MAP. In the thesis, a new FoEs model for image denoising is presented, which is called adaptive FoEs model, and has superior performance. In image deblurring, a new potential function is inserted into object function in regularization, which is created with FoEs model for describing prior constraint condition embodied. The better deblurring result that can keep boundary information is obtained by optimization processing of the object function.Image sparse representation applied to image restoration is studied in this thesis. An image deblurring based on both sparse representation and regularization is presented. In the method, firstly a new sparse regularization item as prior constraint condition is fused in the object function and normal regularization method is used, then a dictionary of image sparse representation can be obtained by learning based on K-SVD and constantly updated in the procedure of restoration, and better restoration performance can be got.Lastly, image restoration algorithm based on non-local information is studied in this thesis. A new denoising algorithm is presented. In the method, firstly a new regularization function is defined, which is based on earth mover's distance for computing the similarity between image; secondly image patches are roughly classified with their dominical directions in order to computing time of regularization function; finally a new split Bregman iteration algorithm is designed for solving new image restoration model. The restoration image with the algorithm can better keep image structure information.