Study On Key Techniques Of Image Processing Based On Sparse Representation Theory

In the area of signal processing,we hope to represent the signal with the most"economic" form.As a burgeoning method,sparse representation realizes the goal above.So called "sparse",which means that the coding of signal has only few nonzero entries.But it is surprising that these nonzero entries can express the essence of signal completely.So sparse representation attracts attention in the worldwide and become a more and more important issue in theory of signal representation.In addition,with the similarity to the vision mechanism,sparse representation has been applied to many areas,such as computer vision,pattern recognition,image processing,automatic control,and so on.This paper focuses on sparse representation theory and its application to image processing.It mainly includes multi-scale geometric analysis(MGA),dictionary learning,sparse coding,spare regularization and their typical applications to image processing.The concrete research contents are shown as follows:(1)Discuss some basis concepts about sparse representation theory,which mainly includes multi-scale geometric analysis and spare representation with overcompleted dictionary.Research their mathematic model,computational method and some typical applications to image processing.(2)Analysis different multi-scale geometric analysis method,such as Wavelet,Curvelet,Contourlet and Non-subsampled Contourlet(NSCT).Compare their performance on basis function construction,multi-dimension filter banks,capturing the line singularity and directional expansion.Proposed a fusion algorithm with entropy measurement based non-sbsampled contourlet.With the redundancy theory,the proposed algorithm produced a mask by the low-frequency coefficients of NSCT,which made the measurement more efficient.In addition,with the non-subsampled character,we can fusion images by point to point conveniently.(3)With the advantage in capturing signal characters by different scale,the paper proposed an image multi-scale enhancement method based on illumination partition.The proposed algorithm divided the image into different illumination space under the weber vision model.And then,we selected the different scale parameter combination for different space,which could gain the interesting information in different space adaptively and promoted the visual contrast.(4)The paper proposed a multi-exposure images enhancement method with MGA and Retinex model.We extended the general Retinex model for single image to the 3D image sequences.Meanwhile,we protect the edge information by NSCT and extract the exposure information from images with different exposure time.Next,combine the information above and gain the enhancement image according to the Retinex model.(5)With some research on current dictionary learning algorithm and sparse coding,the paper proposed a novel dictionary learning method based on compressive sensing(CS).The novel algorithm can learn high-dimension dictionary from the low-dimension compressive samples by coding the dictionary.We also take it into the issue about CS-based image reconstruction and get good performance.(6)The paper proposed a novel sparse scheme for image restoration with dual-prior constraint models.The new scheme has two main contributions.In one side,it establishes a restoration framework with total variation and nonlocal sparsity,which can gain the good performance both of the two constraint models.In other side,we developed a modified iterative Split Bregman to solve the framework above.With two auxiliary variables,we divide the main objective function into three optimizations,which can solve the framework conveniently.