Blind motion deblurring is a very challenging yet important problem in image recovery that receives enduring attention in last decade. Regularization has been one main tool for blind deblurring, which imposes certain prior on target image for estimation. Different from non-blind image deblurring, a good/accurate prior for natual images is indeed NOT necessarily a good image prior for blur kernel estimation, the critical part in blind deblurring. In this talk, in the context of image recovery, we will discuss how regularization and statistical learning techniques can interact each other to enable the introduction of better image prior. Such a discussion leads to a powerful variational expectation maximization (VEM) algorithm with adaptive image prior for blind deblurring, which achieves unprecedented performance on standard benchmark datasets.