Research On Image Compressed Sensing Algorithm Based On Orthonormal Bases And Multilevel Dictionaries

The conventional optical imaging system is based on Nyquist samplingtheorem,which seriously wastes valuable sensing resources and increases overall systemcost. In recent years, the compressive imaging algorithm based on compressed sensinghas been valued. Its outstanding advantages are the reduction of sampling data, savingstorage room, which makes design the system easy and reduces the requirement ofsampling equipment. Based on the achievements of previous researchers, this paper dosome research on the compressive imaging algorithm based on compressed sensing theory.Firstly, this paper study the basic theory of the compressive imaging algorithm,andthen discuss three hot aspects of the algorithm. In terms of signal sparserepresentation,two classical redundant dictionary learning method isstudied;In terms of thedesign of measurement matrix, owe to the uncertainty of the random observation matrix,The observation matrix suiting to the redundant dictionary is studied;In terms of thereconstruction algorithm, we study the SL0algorithm, Experiments show that when theredundant dictionary and the reconstruct algorithm is the same,the optimization of thesensing matrix is better than the random matrix.Secondly, in the existing block compressed sensing algorithm,to collect a set of linearmeasurements of a2D-image block,columns of the2D-image block are first stacked into alarge column vector,however,ignores the intrinsic row/column-wise structure of the2D-signal.This paper adopts a new method for image compressing, The method is basedon treating small patches from a2D image as matrices as opposed to the conventionalvectorial representation,the patches can be encoded as sparse projections onto a set ofexemplar orthonormal bases.The image reconstrcution from2D random measurements,Taking into account the image patches projected into the exemplars which produced theprojection matrix are sparse and use the2D-SL0algoritm to reconstruct the oringinalimage.The2D-SL0is an algorithm based on a modification of SL0.Experiments show theeffectiveness of the algorithm. Finally, because of the single dictionary can not sparsely represent the image patchesof different types,a multilevel dictionary is learned to enhance the sparsity of therepresentation,the training samples for the next level dictionary learning is choosed fromthe training examples having higher reconstruction error in the previous training round. Inthis paper, we propose to perform image reconstruction using an ensemble of weakdictionary, The image patches reconstructed form each level dictionary is weightedaccording to thel1norm of the sparse coefficient. Experimental results show that theproposed algorithm is better than MLD algorithm.