Bayesian image separation with natural image prior

Image separation from a set of observed mixtures has important applications in many fields such as intrinsic image extraction. We investigate in this work a natural image prior based image separation algorithm. The natural image prior is modeled via a high-order Markov Random Field (MRF) and is integrated into a Bayesian framework for estimating all the component images. Due to the usage of the natural image prior, which typically leading to non-convex optimization problems, there is no closed form solution for estimating the component images. Therefore, a Markov chain Monte-Carlo based sampling algorithm is developed for solution. Based on this, a Minimum Mean Square Error (MMSE) estimation can be achieved. The proposed method exploits both the mixing observations and the prior distribution of natural images, modeled via an MRF model. Experimental results indicate that the proposed method can generate better results than state-of-the-art image separation algorithms.