Research Of Image Restoration Based On Compressed Sensing And Sparse Prior

Due to weather,light,equipment and other factors,images are prone to blur,noise and other problems during shooting,so image restoration is a hot issue in image processing.This paper mainly recovers degraded images in datasets and real scenes,including image denoising and image deblurring.It recovers images using compressed sensing and image sparse prior to achieve faster and more accurate clear images.The specific contents are as follows:（1）An image denoising model based on compressed sensing and image segmentation is proposed.In this model,the compressed sensing theory is mainly used to restore the image.Based on the compressed sensing and image segmentation,combined with filtering,a method of image restoration based on F,IS and CS,called F-IS-CS,is proposed.The experimental results of three different image restoration methods show that when the compression ratio of the proposed F-IS-CS method is 0.5,the proposed F-IS-CS method performs mean filtering on the image containing Poisson noise,and then performs segmentation,compression,reconstruction and merging to obtain the best image restoration effect.Compared with the F-CS method,F-IS-CS has a higher peak signal to noise ratio.It reached 32.2198.（2）An image deblurring model based onl₁/l₂-norm is proposed.In this model,the sparse difference between the edge information of clear image and blurred image is used to constrain the blurred image withl₁/l₂-norm,and the blur kernel is constrained with₂l-norm.The image deblurring problem is transformed into an unconstrained optimization problem.The half-quadratic splitting method is used to solve the problem.After the blur kernel is finally estimated,the Split Bregman method is used to effectively obtain the clear image.The final structural similarity between the estimated clear image and the original image reaches to 0.7990.（3）An image deblurring model based on multiple priori is proposed.The dark channel prior,intensity prior and gradient prior are combined to establish a deblurring model.In the estimation stage of blur kernel,different weights are given to the three priori,and three kinds of prior are combined to estimate the blur kernel.The₀l-norm regularization method and total variation method are used to deblur the image in the final non-blind deblurring process.The structure similarity between the estimated clear image and the original image is 0.8256.