Structured Sparse Representation For Image Reconstruction

Sparse coding has become the research emphasis of scholars both at home and abroad in the field of image restoration technology recent year.And the structured sparse theory is the latest research result based on the previous sparse theory. It is a common name of some sparse theory methods which is good at using image prior structural informations.It is an ideal image processing algorithm theory. Using the structural information that peculiar to the image, it can effectively improve the accuracy of image restoration and shorten the processing time.This thesis mainly aims at studying several key points of the structured sparse theory. The prior structured knowledge was added to the original sparse algorithm theoretical model, and some achievements based on different circumstances were made. The main innovation points embodied in the following aspects:Due to the shortcomings including slow imaging speed and a large data collection of the original MRI algorithm.We combined the wavelet threshold method of the original sparse theory and the peculiar tree structure of MRI. Wavelet tree is added to the total variation of the original model. The experimental data showed that the new algorithm can use less data to restore a more precise image Due to the universality and robustness of the algorithm, it performed excellently in the recovery experiments of natural images.Total variational model has the superiority of retaining image edge informations, that makes it widely used in the image restoration algorithm. The total variational model has two kinds of algorithms, anisotropy and isotropy. But both still exist many defects. With the help of a large number of natural image statistics,we added adjustable parameterα, and obtained a more excellent method named mix variational method.The experimental data showed that, the new algorithm makes the recovery of image contour more clear, and the gray level distribution more uniform.In terms of the sparse image restoration, block dictionary learning also played an irreplaceable role. However, the original algorithm has high computational complexity and large cost of image restoration. In order to overcome these shortcomings, we used the self-similarity of image for block processing. The experimental data showed that, our algorithm recovered faster and exquisite.