Non-blind Image Deblurring from a Single Image

Bo Zhao, Wensheng Zhang, Huan Ding, Hu Wang

Abstract: Conventional non-blind image deblurring algorithms often involve in maximum a posteriori (MAP) estimation and natural image priors. However, MAP estimation has several disadvantages which limit its application. To address these issues, we propose to use Bayesian minimum mean squared error (MMSE) estimation instead of MAP to perform deblurring. The new method is based on high-order non-local range–Markov random field (NLR-MRF) prior, which is an effective statistical framework to model prior knowledge of natural images. The high-order NLR-MRF prior can be integrated into MMSE framework naturally. Then, an efficient Gibbs sampling algorithm is employed to compute MMSE estimation. For convenience of computation, we convert to solve a least-squares problem for sampling latent sharp images. The proposed method frees the users from determining regularization parameter beforehand, which relies on unknown noise level. Both quantitative and qualitative evaluations show superior or comparable results to the state-of-the-art deblurring methods.