Blur kernel re-initialization for blind image deblurring

Hyukzae Lee, Changick Kim

Abstract: We propose a simple yet effective blur kernel re-initialization method in a coarse-to-fine framework for blind image deblurring. The proposed method is motivated by observing that most deblurring algorithms use only an estimated blur kernel at the coarser level to initialize a blur kernel for the next finer level. Based on this observation, we design an objective function to exploit both a blur kernel and an latent image estimated at the coarser level to produce an initial blur kernel for the finer level. Experimental results demonstrate that the proposed algorithm improves performance of the existing deblurring algorithms in terms of accuracy and success rate.