Bayesian non-parametric gradient histogram estimation for texture-enhanced image deblurring

Abstract

Image deblurring aims to restore the latent clean image with textures and details from the blurry observation, and is a classical yet active inverse problem in image processing and low level vision. Even though various methods based on image priors have been proposed, the deblurring results by the existing methods usually tend to be over-smoothed and cannot recover fine scale textures. On the other hand, gradient histogram prior has been introduced for texture-enhanced image denoising but the gradient histogram estimation model cannot be used to estimate reference histogram from blurry image. In this paper, we first suggest a gradient histogram preserving (GHP) based image deblurring method, where the reference histogram is parameterized by Hyper-Laplacian distribution. Considering the complexity of blurring process, a Bayesian non-parametric method, Gaussian Processes regression, is utilized for estimating histogram parameters. The experiments demonstrate that, the histogram parameter estimation method is effective, and the proposed GHP based image deblurring method can well restore image textures and improve image quality.