Research On Blind Motion Deblurring Based On Single Image

To solve blind motion deblurring problems, the paper has studied some image restoration methods,such as image deconvolution algorithms and kernel estimation methods, and proposed a multi-scaledmotion deblurring framework for single image. Experimental result shows that the proposed methodcan remove motion blurring in noisy conditions. The main work is divided into the following areas:1) Classic image restoration algorithms are discussed in the paper. Firstly it analyses imagedegradation model and summarizes the main difficulties we are facing. Then it reviews and classifiesthe classic image deconvolution methods.2) The paper mainly analyses the image deconvolution based on natural gradient distributionpriors. A math model, such as Gaussian distribution or Laplace distribution, is used as a regularizationitem to fit the distribution of gradients in natural scenes in order to get better result. Thedeconvolution using hyper-laplace priors can generate high-quality result in noise-free conditions, butwith coarse and noisy surfaces under noisy environment.3) Recent studies have shown that natural images can use sparse representation technique toexclude noise outside in the reconstruction process. To overcome the shortcomings of above method,sparse representation prior is used as another regularization item cooperating with natural gradientdistribution prior and helping to smooth the restored image. The method is solved by breaking downthe problem into two sub-problems: image updating problem and sparse representation problem.4) The paper proposes a multi-scaled motion deblurring framework for single image. Firstly thepaper studies and analyses the motion blurring degradation. Then it applies shock filter to predictsharp edges which help to estimate kernel. Finally the complex deconvolution method mentionedabove is used to restore the whole image with the estimated kernel. Some kernel optimizationmethods are used in the iterative process of kernel estimation to avoid amplifying errors. Multi-scaledscheme is applied to solve big kernel problem. In order to improve computational efficiency, thesimply deconvolution based on gaussian priors is used in iterative process.5) The paper designs and builds a barcode recognition system. It applies the image restorationmethod mentioned above in the stage of image preprocessing. The experiments show that theproposed algorithm works well on barcode image.