The Research On Image Restoration Algorithm Based On Sparse Representation

In the image acquisition process, there are many factors such as video framemotion blur, noise, degradation of resolution, which reduce the quality of images.These result in loss of some useful information in the image and decrease of thevalue of the image. However, in some fields such as public administration,security, medicine and military affairs, image provides essential information andit is an indispensable tool for the social development. Therefore, it is of greatsignificance to do some necessary restoration by the poor quality parts to obtainclean and high-quality images.Based on the latest theory of Sparse Representation, this paper mainlyanalyzes the image denoising and super-resolution reconstruction.Sparse Representation theory is often used to decrease the dimension ofsignals. After sparse decomposition, a two-dimensional image can be linearlyexpressed by a small number of atoms in a sparse dictionary. Considering theimportance of sparse dictionary, this paper studies the learning algorithms ofsparse dictionary. The dictionaries can be divided into three categories: the basictransform base, redundant global dictionary and redundant local dictionary, andthen apply these dictionary learning algorithms into the image denoisingprocessing.The sparse image denoising algorithm considers that image can bedecomposed sparsely but noise can not. Thus noise in a image can be removedthrough the sparse representation. This paper simulates the representativeDCT-based dictionary, KSVD global dictionary and KLLD local dictionary toachieve the purpose of denoising respectively.In view of the superiority of the redundant dictionary algorithms, we applythem to the single-frame image super-resolution processing. Taking thecharacteristics of image super-resolution reconstruction into account, weestablishe a clear image sample set, and then through different learningalgorithms obtain both the high and low resolution dictionaries. This paper makesanalysis of three kinds of global dictionary learning methods: the initial samplepieces dictionary method, the Lagrangian training method and KSVD training method.Taking the global dictionary algorithm limitations in detail images intoconsideration, an improved algorithm, named Dual_KLLD algorithm is proposedvia the effective combination of the image classification theory and Dualdictionary learning method. For detailed images, this algorithm performs better inimage reconstruction than global dictionary. However, the disadvantage ofDual_KLLD algorithm is that there is obvious blockiness in the reconstructionresult. On this point, this paper focuses on the PCA_KLLD algorithm. Afterimage classification, we use the PCA matrix decomposition to train thedictionaries, which makes up the lack of Dual_KLLD algorithm and achieves abetter treatment effect.In addition to dictionary, the solution of coefficient matrix is also a part ofthe sparse algorithm. We simulate and analyze the classical OMP algorithm,ROMP algorithm and the symbolic search algorithm in this paper, and improvethe processing efficiency while ensuring the sparse effectiveness.