Reconstruction Of Compressed Light Field And Denoising Of Images Via Improved K-SVD Dictionary

Large-scale data processing is one of the challenges of modern signal processing and it has become a current hot research.This thesis concludes the results of the researches based on compressed sensing and light field imaging in detail,and we study on reconstruction of compressed light field and denoising of images via improved K-SVD dictionary.The completed works are as follows:1.We study the basic mathematical framework of compressed sensing theory,sparse representation,measurement matrix and the algorithm of reconstruction,and analyze sparse representation and reconstruction algorithm in detail.Then we derive and conclude the characteristics of some kinds of model via numerical simulation and mathematical analysis.2.An improved gradient reweighted non-local averaging algorithm over K-SVD dictionary is proposed here.The improved algorithm can denoise the strong noise images.Detailedly,the improved algorithm finds the inherent structure of original signals using non-local averaging algorithm with gradient reweighting,which is obtained by total variation.Then combined the sparse prior and similarity of the structure of images,in order to solve the optimization of inverse problem.We contrast the improved algorithm and the traditional K-SVD dictionary denoising by numerical simulation under different noise intensity.As the result of experiments,when the noise standard deviation is as high as 50 and the input noise SNR is 7.954 dB,we verify the improved algorithm has the beset effects,showing in numerical PSNR has improved 0.2dB and SSIM has improved 0.02.When the noise standard deviation is as high as 100 and the input noise SNR is 1.961 dB,the effects of denoising images perform better with the input noise SNR between 7.954 dB and 1.961 d B than the traditional one.Also,we contrast different kinds of images to prove the improved algorithm can keep the outline of the images’ detail better,and the improved one is more suitable for processing covering a small amount of texture information and a large number of contour information images.3.The optimal threshold lasso reconstruction light field regression algorithm is proposed.Firstly,we study on the mathematical framework of light field imaging: light field information function,encoding light field acquisition,and compressed light field reconstruction.Secondly,we mainly focus on compressed light field reconstruction,by K-SVD dictionary in order to obtain the light field atoms,and we propose the optimal threshold lasso reconstruction in order to reconstruct the light field.Lastly,we compare the improved algorithm with the traditional solver of l₁ norm problem,by reconstructing 5′5 light field,to prove the improved one is more suitable for higher dimensional signals and has higher reconstruction precise and faster convergence speed.However,the improved algorithm has narrowed the depth of field of the reconstructed images.