Generative Super-Resolution

Introduction

In this work, we present a Natural image prior based Bayesian image SR (NBSR) approach with a flexible high-order MRF model as the prior for natural images. The Minimum Mean Square Error (MMSE) criteria is used for  estimating the HR image. A Markov Chain Monte Carlo (MCMC) based sampling algorithm is presented for obtaining the MMSE solution. The proposed method can not only enjoy the benefits offered by the flexible prior, but also has the advantage of making use of the probabilistic modeling to perform a posterior mean estimation, thus is less sensitive to the local minima problem as the MAP solution. Experimental results indicate that the proposed method can generate competitive or better results than state-of-the-art SR algorithms.

Advantages

1) the proposed method is a fully Bayesian approach, which can incorporate the a priori knowledge on the latent HR image as well as other uncertainties such as the noise level into the framework  in a natural way;

2) a more advanced prior is used  for capturing the HR image statistics rather than other very simple priors. Specifically, a high-order MRF model is used for modeling the natural image statistics, which, at the same time, enables generative sampling from it;

3) the Bayesian MMSE criteria is applied for estimating the HR image rather than the MAP criteria. The MMSE approach has several advantages over MAP. It does not require ad-hoc modifications as MAP to achieve desirable restoration performance. Also, it is less sensitive to the local minima in the solution space than MAP, especially when heavy-tailed priors are applied.

Results

1. Image SR Quality Comparison  ( PSNR, zooming factor = 3 )

2. Color Image SR ( zooming factor = 3 )