Kernel refinement based on best light streak for motion deblurring

Abstract

This paper introduces a blur kernel refinement method that produces a more accurate kernel estimation based on the best light streak that is selected from a motion blurred image. The best image patch that contains a clear light streak is firstly selected and the blur kernel is estimated from the patch by solving an optimization problem. Then, a kernel refinement method based on region growing is proposed to extract the motion trajectory to be the refined kernel and avoid the disturbance from the background. At last, a non-blind deconvolution method is used to obtain the restored sharp image using the refined kernel. Experimental results of both synthetic images and real world images demonstrate that the kernel refinement can improve the quality of deconvolution and yield a better sharp image with less ringing artifacts. Also, the normalized cross-correlation is utilized to evaluate the similarity between refined and ground truth kernel and verifies the improvement of refined kernels.