Font Size: a A A

Image Denoising Based On Sparse Representations

Posted on:2015-06-07Degree:MasterType:Thesis
Country:ChinaCandidate:C C DongFull Text:PDF
GTID:2298330452458997Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
Image denoising is the most basic problem in image processing. With the rise andpopularization of compressed sensing, more and more scholars begin to pay attention to the sparserepresentation theory and its applications, image denoising based on image sparse representation hasbecome a frontier research topics in this field in recent years. The paper studies the sparserepresentation theory, discusses two key problems, which is the atomic database structure and sparsedecomposition. On this basis, the image denoising method based the over complete dictionary sparserepresentation is studied.In the sparse representation theory, there are two methods for constructing the dictionary: one isthe analytical dictionary composed with fixed base groups; the other is adaptive dictionary based onthe training sample learning. Although the first cannot adapt the structure characteristics of therepresentation signal, it has a very fast implementation, so it is still widely used in practice. Atoms isstructured by the all phase biorthogonal transform previous proposed (APBT), and the basisfunction are combined into a hybrid atomic library in the paper, the author puts forward imagedenoising method based on the class dictionary representation, and achieve good results.Redundant dictionary based on learning method can more accurately extract structurecharacteristics of signals; it is also a research hotspot in recent years. Based on the study imagedenoising algorithm on the basis of KSVD dictionary learning, the author combines the correlationcoefficient matching criterion and dictionary cutting method, proposes an improved dictionarylearning algorithm. Furtherly, in order to use non local self-similarity information of the image,theauthor puts forward combining the self-similarity as a constrained regularization into the imagenoise model and the image denoising algorithm based on the improvement of dictionary learningand non-local self-similarity.In a large number of experiments, compared with the traditionalKSVD denoising method, the method can improve the smoothness of homogeneous regions whilestill preserving more texture and edge details.
Keywords/Search Tags:Sparse representation, all phase biorthogonal transform, dictionary learning, self-similarity regularization
PDF Full Text Request
Related items