Abstract
Image deblurring is an important and challenging task in image processing. The main goal of blind deconvolution is to recover latent image from blurry image without knowing blur kernel. To estimate the blur kernel, a typical method is MAP (Maximum a Posteriori) approach which uses the predicted latent image and the blur kernel it estimated to minimize the error iteratively. However, the method has complex computation and takes long time if a whole image is used to recover the latent image. In our research, we find out that it is not necessary to use whole image to predict the blur kernel. By using a patch of the image with enough edge information can efficiently reduces the computational time and improves the results for image deblurring since it eliminates the smooth regions and repetition edges’ influence. In our method, we use an anisotropic Partial Differential Equation which can keep strong edges’ information to find out informative points, patch them by clustering points to avoid fix patch size and wrong patch’s center location, and treat every patch center as a player to use a coalitional game to find out the best patch. In our experiment, we use our algorithm to modify two blind deconvolution methods and the results show improvement with our method both on real world and synthetic images.