Fast spatially varying object motion blur estimation
Yi Zhang, Keigo Hirakawa · 2015
Object motion results in spatially varying image blur. We propose an efficient method to recover a dense estimation of blur kernel. Proposed method takes advantage of the sparse representation of double discrete wavelet transform (DDWT) to simplify the wavelet analysis of blurry image. Our optimal solution includes separating the estimation of blur direction and length by investigating the cross-correlation; and exploiting mean absolute summation (MAS) function for noise-robust estimation. We demonstrate by experiments the considerable improvement in speed and handling noise.