Research On Blind Image Restoration Algorithm Of Fuzzy Image Under Tensor Frame

In the process of acquisition,storage,transmission and processing,the quality of an image is generally degraded by many adverse factors to appear as a blurred image,a noisy image,an image with low resolution,an image with loss of local information,and so on.Image degradation brings great difficulties for the subsequent image processing and analysis.Image restoration refers to restoring the latent image from a degraded version according to the degradating factor,and typical image restoration processing includes image deblurring,image denoising,image super-resolution and image inpainting.Blurring degradation is one of the most common image degradation types.There are many factors leading to blurring degradation,such as inaccurate focus in an optical imaging system,relative motion between an imaging system and imaging objects,atmospheric turbulence,image compression,image preprocessing and so on.Image deblurring refers to restoring the latent image using a blurred version,and it has become one of the most important research topics in image processing with application in many fields,such as astronomy,military detection,public safety,intelligent transportation,medical imaging,daily life and so forth.Therefore,a research on theories,algorithms and key technologies for image deblurring has an important academic significance and many practical application values.Image deblurring can be generally categorized into non-blind and blind types,according to whether the blurring kernel is known or not.Blind restoration for a blurred image is a typical ill-posed problem,since both the blurring kernel and the latent image are unknown.Usually the process for blind image deblurring is divided into two steps: One is to estimate the blurring kernel,and another one is to deblur the blurred image using the estimated blurring kernel and a certain restoring algorithm.This thesis makes an in-depth research on blind image restoration algorithms and the main work includes analysis of blurring degradation models,estimation of the parameters of blurring kernel models,and development of deblurring algorithms.The related work is established both in the vector framework and two tensor frameworks,and the main contents and contributions are as follows:First,in the vector framework,wherein a gray image is concatenated to a vector,the characteristics and limitations of the traditional convolution degradation model is studied and the inherent blurring factors caused by the optical diffraction and anti-aliasing filtering in optical imaging systems are analyzed.At the same time,in a view of function approximation,a weighted Gaussian blurring kernel model is proposed,in which an unknown blurring kernel is approximated by weighted sum of Gaussian kernels.And based on the weighted Gaussian kernel model,a parameter estimation algorithm with closed-form solution is proposed.And at the same time,several existing deblurring algorithms based on total variation(TV)regularization are reviewed and analyzed,and a deblurring algorithm based on image local structure information is proposed,in which multiple images deblurred with different TV regularizations are fused into a latent image.Experimental results show the effectiveness of the proposed algorithm.Then,the weighted Gaussian kernel model in the vector framework is first extended to the traditional tensor framework wherein a Z-direction difference is included to form a three-order operator,and a three-order tensor TV model is proposed by extending traditional TV model to the tensorial version.On the basis of three-order tensor TV model,a blind restoration algorithm based on the weighted Gaussian kernel model is proposed,and its feasibility is illustrated by experimental results,and notice that the algorithm may be used to restore videos,high-order medical images,and so forth.In addition,in considering the problems existed in general blind restoration algorithms wherein sufficient effective edge information is needed to estimate the blurring kernel,and generally an alteration strategy is adopted for blurring kernel estimation and latent image restoration and thus it leads to high time complexity,a novel blurring kernel estimation method based on tensorial two-dimensional principal component analysis(T2DPCA)is proposed,based on studies of linear two-order tensor space and the T2 DPCA subspace projection method in the Kilmer's tensor framework,wherein the conclusion in image processing study that the power spectra of a natural image has power-law feature is taken for whitening of the magnitude spectrum of a blurred image,by projecting the magnitude spectrum into a suitable subspace formed by T2 DPCA method.Finally,the blurring kernel is reconstructed by the phase-retrieval algorithm.Experimental results show satisfactory performance of the proposed method.Furthermore,the restoration problem for color images is studied in the Kilmer's tensor framework,in which a color image is regarded as a third-order tensor and the problem existed in general blind restoration algorithms that only the gray information of a color image is utilized to estimate the blurring kernel,and thus a restored image may be unsatisfactory if the size of a blurred image is too small or the salient edge in a blurred image is too little,is taken into consideration,and then a tensor-total variation(t-TV)model and the associated algorithms for model parameter estimation and image restoration from a blurred color image are proposed.Experimental results show that,in comparison with other mainstream algorithms,in particular when dealing with blurred images with low-resolution,the estimated blurring kernel is more precise,and thus the restored image has more sharp edges and more details,and at the same time the ringing effects in the restored image is also greatly suppressed.