Text Image Motion Deblurring Based On The Kernel Sparsity Priors

Image is one of the most important ways to transmit information,and occupies an positions in our daily lives.The specific text images,sometimes contain more important information and thus may be more useful than the ordinary images.But we will get blured text images because of the weather,focus,movement and so on.At this time,restoring those blur text images is a very meaningful research.In this thesis,we study the blind deblurring for the text images by analyzing the sparsities of both the text images and the kernel.The main work will show in the following:First of all,some fundamental knowledge and related work about deblurring are introduced.The degradation model of the process of image degradation as well as the prior knowledge about natural images,text images and the kernels are discussed first.Then energy optimization model built based on the prior knowledge of the images are presented.Next,A new text image deblurring method based on kernel sparsity are proposed,where kernel sparsity is widely ignored by existing text deblurring studies.In the model,the sparsities of both the text images and kernels are taken as prior by L0 norm as constraints.Then a half-quadratic splitting method is adopted to solve this energy minimization of the model.A denoising method is also used to obtain a clean one after the optimization:First,the motion trajectory about the kernel is computed by exracting the skeleto;then,a cross window is recruited to traverse this trajectory and remove the noise;finally,the noise-free kernel is used to restore the clear image.Experiments show that the L0 norm restrained deblurring model can effectivly deblur the image with the denoising step returning abetter results.Finally,block based non-uniform text images deblurring using the textural degree.The non-uniform blur text images is divided into patches with every patch assuming to be blurred by a uniform blur.The L0 norm restrained text images and kernels are used for each block with the energy optimization model.In the last chapter,in order to get an efficient deblurring,the kernel similarity of neighboring patches is analyzed and a method to estimate their kernels are introduced.The kernel of neighboring patch is used as the initial value when estimate the kernel of current patch.When the gradient and the entropy are used to measure the textural degree of each part so that their orders can be obtained subsequently.First the patch which has the richest texture among the blurred pathces is recovered,and then the latent neighbouring patch which has the less texture and the kernel of its neighboring patch is estimated.The kernel denoising method previously presented is also taken.Experiments show the efficiency of the proposed block based method.