Visual Image Restoration Methods Based On Image Statistical Features

Images always suffer from degradation in the imaging process,which damages the visual quality a lot,and image blur is one of the major source of them.E.g.the vibration from the imaging system or the relative movement of the target,out of focus,and the density fluctuation of aerosol caused by environmental factors will all lead to different degrees of blur degradation of the observed images,and seriously affect the follow-up process in engineering.Aiming at the problem of image blur and degradation,this dissertation discusses and researches the prior models of natural images,the blur feature of degraded images,and the estimation of point spread functions,etc.And different solutions are proposed according to the degradation in different situations.The main contents and novelties are summarized as follows:1.An image prior model based on the characteristics of mathematical statistics of natural images is proposed,and the Bayesian hierarchy theory is adopted to design a deconvolution algorithm so as to recover the latent image.In this dissertation,the mathematical statistical analysis on a large number of natural images is carried out and an image,resulting in an image model that combines the gamma curve and Gaussian distribution that regards as the prior model of natural image.Then with the help of Bayesian hierarchy theory,and the likelihood function derived from the additive random noise model,a posteriori distribution about the original image and the observed one is obtained from the prior model mentioned before,and the cost function is formulated based on the a posteriori,as well as an optimization method with a tail recursive structure is designed.The optimal solution is found by alternating searches in the negative gradient and its conjugate direction of the cost function,thus obtaining the optimum estimation of the original image under the perspective of maximum a posterior(MAP),which is regarded as the sharp image after restoration.2.A measurement of the degree of image blur based on the theory of sampling distribution and a blur feature based on the auto-correlation in the image gradient domain are proposed.Combined with deep neural network technology,networks which are capable of predicting the parameters of the blur model are built and well trained.In this dissertation the process of image blurring is analyzed from the point of view of central limit theorem at first,the difference between the original image and the blurred image in the gaussianity is demonstrated.Then,from the perspective of signal separation theory,it discusses the feasibility of using high-order statistics to measure the degree of image degradation and solving the image restoration problem from the parameter space of the degradation model.In addition,the amplitude of auto-correlation of a blurred image in the gradient domain is found to be very strongly correlated to the blur model,besides,the surface of the amplitude changes as the degree of blur in a certain degradation.On this account,a large number of experimental images are blurred artificially by specific blur models with different parameters,then blur features are extracted to a generalized recursive neural network for training so as to predict the blur parameters of any input blurry images.Finally,the original sharp image can be solved by the deconvolution algorithm of total variation regularization.3.A point spread function(PSF)estimation method based on the sparse characteristics of the image proposed.In a sharp and clear image,the sudden changes in the pixel value of the targets and background at the boundary are act as the distinct curved lines in the gradient domain.However,the pixel values change slowly and gently at these boundaries in the case of blurred image,and the curved lines in gradient domain are blurred too.The pixel histogram counted from the gradient domain indicates that the proportion of non-zero values in the blurred image is much higher than that of the sharp one,but its sparsity is reduced.Based on this phenomenon,the Kronecker-Delta function is used to count the non-zero values in the gradient domain,and a sparse optimization based PSF estimation method is proposed with the help of the MAP framework.Thereafter an image deconvolution algorithm derived from the heavy-tailed distribution-based prior model is introduced and the clear latent image is thus obtained.