Abstract
Deblurring from a single image has been extensively discussed. Most of single-image blind image deblurring methods using whole image to estimate the blur kernel based on a coarse-to-fine MAP approach and they are usually computational expensively due lots of iterations. We observed that using whole image to estimate blur kernel is not always a good option and may ruin the kernel estimation process but also need more computation time. In this paper, we focus on accelerating the blind deconvolution algorithm and increasing the accuracy of kernel estimation. We propose a coalition game based patch selection method to choose an informative patch for kernel estimation. We first find the informative pixels which is useful for kernel estimation using blur image gradient magnitude and a strong structure map. However, we consider that single pixel is not informative enough. For each pixel we found, we form a small patch centered at it. Our goal is to find a group of informative patch and united them into a large patch. We apply coalitional game to solve this problem. In our coalitional game, each patch represents a player, and they seek to join a coalition to improve their payoff. We design the utility for each coalition and compute the Shapley value to fairly distribute the utility to each player in the coalition. After the game, we will have a coalition such that no other player can obtain an outcome better than the current assignment and then, we use it to form our final patch. We show the speed-up and the quality improvement of our method both on real-world and synthetic images.