Research On Blind Image Deconvolution Algorithm Based On Variational Bayesian Inference

Image is a crucial information carrier.Whether it is daily life,industrail production or scientific research,it depends on sharp images.However,during the image acquisition,multiple sources of degradation lead to image blurring and uselessness.Image restoration technique reverses the degradation process and restores the blurry image to obtain latent sharp image,which helps people recover important lost information.This dissertaion studies the basic theory of image retoration and improves the Variational Bayesian(VB)based blind deconvolution algorithm of Kotera et al.Apply Automatic Relevance Determination prior model to the observed blurry image,latent sharp image and blur kernel,in line with natural image statistics and Bayesian modeling.The ARD prior model,essentially the student's t-distribution,can be expressed as a superposition of a finite number of Gaussians with common mean,whose precision(inverse variance)obeys Gamma distribution.This model plays nicely with the VB inference.Based on VB inference and the ARD prior model,a blind image restoration algorithm is proposed to solve blind image deconvolution problem.The blind image deconvolution algorithm consists of two stages,imcluding blur estimation and non-blind devonvolution.In the improved blind image deconvolution algorithm,the ARD model is defined over blur gradients,considering the intrinsic structure of the blur.To solve the deconvolution problem,the VB inference approximates the posterier probability with distributions in decomposed form,and get to correct solutions of image and blur by minimizing the Kullback-Leibler divergence,which measures the difference between the approximate distribution and the posterior.The VB based blind algorithm involves many variables and parameters,such as the image and precision parameter of its prior,which can be computed alternately by iterative procedure.Besides,the blur should satisfy non-negative and energy conservation in blur estimation phase.The non-blind deconvolution algorithm is based on VB inference as well and adopts the ARD prior model to blurry and sharp image.The blur,estimated in previous stage,is used in the image estimation as a known quantity.On the basis of the proposed algorithm,detailed solutions to problems appeared in the restoration process are given,e.g.ringing artifacts.Finally,experiments are implemented to evaluate the performance of the VB based blind image deonvolution algorithm.Synthectically blurred images and real blurry images are used to perform the evaluation in experiments.The synthetic experiments proves that the quality of blur estimated by the proposed blind deconvolution algorithm is slightly higher than the quality of which estimated by Kotera's algorithm and demonstrates the the robustness of the proposed algorithm to non-Gaussian noise.The real image experiments also shows the quality of blur estimated by the proposed algorithm is higher than Kotera's algorithm,which can effectively suppress the ringing artifacts,and reveals it performs well for motion blurred images.