Research On Image Restoration Based On Domain Decomposition

With the rapid development of artificial intelligence,research on machine vision,such as driverless technology,is heating up.Image recovery is an important part of image preprocessing in machine vision research.Clear image and fast real-time processing provide a favorable guarantee for subsequent image intelligence research.Due to the imperfection of the imaging system,recording equipment and transmission medium,the image is prone to degradation: the image contains noise and the image is subject to blurry interference.In addition,due to the improvement of device storage capacity and the pursuit of image resolution,large-scale image processing problem increases the difficulty of real-time image processing.In this thesis,the large-scale image recovery problem is studied,and the large-scale image recovery problem is transformed into a small problem with the help of the region decomposition method,so as to achieve parallel solution,aiming at reducing the image recovery time and ensuring the image recovery effect.The research contents of this thesis are as follows:(1)Firstly,this thesis studies the domain decomposition algorithm for image denoising problem,and proposes a domain decomposition algorithm for TV-L1(total variational model guided by L-1 norm)image denoising model.By using successive subspace correction method,the original image domain was decomposed into overlapping subdomains,and the primal-dual method was used to solve the subproblems in each subdomains.Numerical experiments show that the algorithm has good acceleration ratio and efficiency.In addition,an improved TV-L1 image denoising algorithm is proposed in this thesis,which solves the shortcoming of TV-L1 image denoising model that is prone to generate staircase effect in the image smoothing area.(2)In addition,this thesis applies the region decomposition method to the TV-L1 image deblurring problem.Since the image deblurring model involves the convolution operation of the blur kernel,the domain decomposition will destroy the structure of the blur kernel due to the image segmentation,thus increasing the difficulty of the algorithm.In this thesis,the larger size of the overlapping area is selected in the successive subspace correction method to overcome the above problems.Numerical experiments show that the algorithm can restore clear image deblurring effect,and has good acceleration ratio and efficiency.(3)Furthermore,the problem of blind image deblurring with unknown blur kernel is a more practical problem.Therefore,this thesis also studies the application of domain decomposition method in image blind deblurring.Numerical experiments show that the algorithm can restore the blur kernel,greatly reduce the computing time of the blind image deblurring model,and restore the image to a PSNR value similar to that of the blind image deblurring algorithm without domain decomposition.(4)This thesis further proposes a blind deblurring algorithm for single color motion blur image.The idea of decomposition is applied to the color image space,and R,G and B color three-channel are used to restore the real color image and blur kernel.Numerical experiments show that this algorithm can estimate the blur kernel through R,G and B color three channels and restore the clear image with only one color blurry image,which has theoretical and practical significance.To sum up,this thesis studies the application of domain decomposition method in image restoration.All the experiments of image domain decomposition and recovery shows that the domain decomposition algorithm proposed in this thesis can not only restore the clear image effectively,but also improve the computational efficiency greatly,thus accelerating the realtime processing of large-scale images.