Research On Reconstruction Algorithm Of Image Compression Perception

Compressed sensing theory is a new sampling theory.Compared to the traditional sampling theory,it has two advantages.One is that it makes the signal compression during the sampling.Another is that a small number of random projections of a compressible signal contain enough information for exact reconstruction.As one of the crucial issues,the reconstruction algorithm plays a key role in the application of compressed sensing and affects its practical usage.Recently the signal reconstruction problem has been used in many practical field including image inpainting,language processing,geological exploration and so on.How to design a reconstruction algorithm with high reconstruction accuracy to reconstruct signal,has been a hot study.Under this background,dissertation has deeply studied the compressed sensing reconstruction algorithms for image in order to find effective reconstruction algorithms.This thesis studies the main contents are as follows:(1)An improved Compressive Sampling Matching Pursuit(CoSaMP)algorithm is proposed in this thesis.CoSaMP algorithm is one of the matching pursuit algorithms,its signal reconstruction performance is better than other matching pursuit algorithms.Taking that the product operation does not maximize the degree of correlation between the atoms in residual vector and sensing maxtrix with CoSaMP algorithm into account,while the correlation coefficient can represent the degree of correlation of two vectors better,we propose the optimized CoSaMP algorithm based on correlation coefficient.We perform experiments on one-dimensional analog signal and image signal,the experiments show that the optimized algorithm improve the quality of the reconstructed signal compared to CoSaMP algorithm..(2)We research image reconstruction algorithm of compressive sensing which is based on the GPU parallel computing technology.Nonlocal Low-Rank regularization(NLR-CS)algorithm can significantly outperform existing state-of-the-art CS techniques for image recovery,but the algorithm has lower efficiency.We introduce the GPU parallel computing technology and propose the parallel NLR-CS algorithm based on GPU.We find out hot spots of the algorithm with profiling and performance analysis under the precondition of understanding the principle of the algorithm,then analyze the hot spots to check whether they can be paralleled or not and utilize CUDA programming model to design and implement the parallel algorithm.We perform experiments on image signal,the experiments show that the running time of optimized algorithm can be reduced obviously and the reconstructed images are almost unaffected.(3)An improved NLR-CS algorithm is proposed in this thesis.NLR-CS algorithm exploits low-rank regularization to reconstruct the image,not considering the relationship between adjacent pixels,the reconstructed image could not keep texture retails.In order to solve this problem,we propose an algorithm which combines the low-rank with total variation regularization.The new model combines the local with nonlocal information of an image,which is solved by Alternating Direction Multiplier Method to obtain the reconstructed Image.We perform experiments on image signal,the experiments show that the optimized algorithm improve the quality of the reconstructed images.