Research On A Single Image Motion Deblurring

A single image deblurring is an important research topic in image restoration. Image blur is a very common phenomenon which caused by motion. There are many reasons that can let image blur. One reason is that the camera shaking. There are many methods for image deblurring, but most methods are heavy computation, slow speed, small kernel assumption, ringing artifact, multi-image, additional hardware. This text introduce to some methods about image deblurring. Everyone has some advantage and disadvantage.We introduce to Qi Shan present iterative image deblurring that appears in the paper 2008 and Lu Yuan present multi-scale progressive non-blind deconvolution that appears in the paper 2008.We introduce to Qi Shan present iterative image deblurring that appear in the paper 2008. In the case that kernel is unknown. Establish an iterative deblurring algorithm based on a maximum posteriori probability (MAP). The algorithm consists of two parts. Estimates kernel and recover a clear image. Use Bayesian theory. Seeking the establishment of a maximum a posteriori probability (MAP) problem。And translate them into the minimum energy problem. Set the number of iterations. Iterate kernel and clear image. The final image can be restored satisfactorily. At the same time we made some improvements. The constraints of natural images have been improved. Gaussian distribution of natural images is satisfied with the heavy-tailed distribution. Algorithm is relatively simple to implement. Reduce the complexity of the algorithm. Calculation of the kernel was made improvements. Landweber iteration method used to solve. Landweber iteration method is easy to understand. And the iterative method is simple.We introduce to Lu Yuan present multi-scale progressive non-blind deconvolution that appears in the paper 2008. In the algorithm, a gradual multi-scale framework for solving non-blind deconvolution. from coarse to fine-scale evolutionary scale the image non-blind deconvolution. At the same time we made some improvements in the calculation of kernel. For calculate the kernel. In 2008 Lu Yuan in the paper only use the 2006 Rob Fergus paper that mentioned the method of statistical learning. However, sometimes this method of calculation kernel can not get the optimal solution, especially product ringing. In this paper, we first deal with image noise using bilateral filter. recover the strong edge of the image using shock filter. then using the method of statistical learning that appear in 2006 Rob Fergus paper. Because remove the image noise and restore the strong edge of the image, it is easier to calculate the kernel. Experiments show that this algorithm can significantly reduce the ringing and easy handle large kernel.