Research On The Technique Of Image Deblurring

Image is already one of the most important means of people’s access to and storage of information in today’s society.With the rapid development of science and technology,the devices of image acquisition are becoming more and more common,and they have not only limited to medical,military,production and other professional science,and more become the tool of people in the daily life.However,due to environmental factors,human factors and other reasons make the image produce quality degradation in the process of image obtaining.For these images which can’t get again,image blur becomes a very serious problem.Motion blur is a kind of degradation of image quality,and the main reason of motion blur is due to the relative motion between a camera and the target.According to whether the blur kernel is known,the image motion deblurring algorithm is divided into blind motion deblurring and nonblinded motion deblurring.Blind image motion deblurring algorithm is closer to reality,that is,only a blur image is used to estimate the kernel and the latent image.But this algorithm have more problems,especially the estimation of kernel,the accuracy of kernel estimation is directly related to the quality of latent image.Recently,the blind motion deblurring algorithm which based on maximum posterior probability has gotten more and more attention by researchers.In this thesis,a fast and robust kernel estimation algorithm is proposed to solve the problem that the kernel estimation is not accurate,the robustness is poor and the calculation speed is slow in the previous algorithm.Firstly,this thesis uses the image smoothing method via L0 gradient minimization to detect salient structures,and improves the algorithm so that it can be better applied to the image motion deblurring.Blur image structure extraction is a prediction of the edge of the latent image.According to the previous research and analysis,the accuracy and robustness of the kernel estimation can be improved by using the predicted edge.After extracting the structure information of the image,this thesis applies a novel algorithm based on L0 norm and L2 norm to estimate the kernel.The algorithm can not only guarantee the sparsity of kernel,but also ensure the connectivity of kernel by using the L2 nor.In the past kernel estimation algorithm,the connectivity problem of kernel is often neglected,so the latent image has ringing artifacts.Since the L0 norm and the L2 norm can be solved by the conjugate gradient method,the speed of kernel estimation algorithm is fast in this thesis.This thesis utilizes a connectivity-based algorithm to suppress high-frequency noise in kernel.In addition,since the saturation region in the blur image has an adverse effect on the image recovery,this thesis proposes a saturation region detection method based on HSV space,and guarantees the recovery of the image by suppressing the saturation region.The experimental results show that the algorithm proposed in this thesis can not only estimate the clear image with good quality,but also obviously improve the recovery speed by shortening the computing time by more than 50%.