Research On Blind Deblurring Algorithms For Global Motion Blurred Images

Image is an important carrier for transmitting information,and its quality directly affects the accuracy of information transmission.With the popularization of mobile phones,cameras and other devices,the global motion blur caused by shaking is more and more frequent,which brings great inconvenience to people's life,study and work.How to recover highquality clear images from global motion blurred images is a hot research topic in the field of digital image processing,which has extremely important significance and wide application value.In the case where the image blur kernel is unknown,the use of blind motion deblur technology to obtain high-quality clear images is a serious ill-posed problem.In order to solve this problem,this thesis deeply studies the existing algorithms and analyzes them.It mainly optimizes and improves the algorithms for accurately estimating blur kernels,suppressing the ringing effect in image restoration process and speeding up image recovery time.The main research contents are as follows:First,because the traditional algorithm is susceptible to noise interference,which affects the acquisition of priors and the estimation of blur kernels.In order to obtain a richer prior of texture details,the global motion blurred image needs to be preprocessed.At first,a double filtering algorithm is proposed,which combines the median filtering and the bilateral filtering to eliminate the high and low frequency noise in the image to a large extent.Then,according to the differences of gradient distribution between the blurred image and the clear image,the Canny operator with stronger directivity is used to obtain the gradient information of the blurred image,and enhance the edge features to get a more favorable strong edge prior of image.Second,in order to obtain a more accurate motion blur kernel and recover high-quality and clear image,a blind image motion deblurring algorithm based on maximum a posterior probability(MAP)is proposed.At first,a blind estimation model of the multi-scale motion blur kernel based on MAP is designed according to the fidelity and sparsity characteristics of blurred image priors.In accordance with the gradient prior of the L1 and L2 regular joint constraints,the algorithm obtains the optimal solution of motion blur kernel step by step through multi-scale iterative calculation.Then,an improved RL deblurring algorithm is designed to suppress the residual noise.Combining the optimal blur kernel and the gradient prior with sparsity,the deconvolution operation is used to obtain a high-quality restored image with more complete texture details.The experimental comparisons shows that the algorithm not only effectively weakens the ringing effect and shortens the image recovery time,but also has some universality.Third,due to the strong dependence between the blur kernel estimation and the image deconvolution recovery process in the traditional algorithm,it is easy to accumulate errors or even magnify errors,which may lead to some problems,such as low image deblurring accuracy and slow processing speed.Aiming at the above problems,this thesis proposes an improved generative adversarial network model to implement blind image motion deblurring in an end-to-end manner.In order to generate high-quality deblurred images,the residual network is used to build the generator of the model.Meanwhile,considering the sparsity of the gradient information and the consistency of the content before and after image motion deblur,L1 and L2 losses are used to constrain the gradient and content of images respectively in the model.In order to enhance the robustness and applicability of the network,the model uses a global discriminator and a local discriminator to determine the global content and local details of the restored image respectively.Finally,the model is trained by the publicly dataset.The comparative analysis results of a large number of experiments prove the superiority and stability of the improved algorithm.