Authors: Shaoguo Liu Haibo Wang Jue Wang Sunghyun Cho Chunhong Pan
Publish Date: 2014/06/29
Volume: 31, Issue: 5, Pages: 733-746
Abstract
Existing image deblurring approaches often take the blurkernelsize as an important manual parameter When set improperly this parameter can lead to significant errors in the estimated blur kernels However manually specifying a proper kernel size for an input image is usually a tedious trialanderror process In this paper we propose a new approach for automatically estimating the underlying blurkernelsize value that can lead to good kernel estimation Our approach takes advantage of the autocorrelation map automap of image gradients that is known to reflect the motion blur information We show that the standard automap suffers from structural edges in the image and cannot be directly used for kernel size estimation To alleviate this problem we develop a modified automap method that contains a directional attenuation component which can effectively reduce the influence of structural edges leading to more accurate and reliable kernel size estimation Experimental results suggest that the proposed approach can help stateoftheart deblurring algorithms achieve accurate kernel estimation without relying on manual parameter tweaking
Keywords: