Discrete multiscale Bayesian image reconstruction

Statistical and discrete-valued methods can substantially improve reconstruction quality by incorporating prior information about both the imaging system and the object being imaged. A statistical method shown to perform well in the tomographic setting is Bayesian MAP estimation. However, computing the MAP estimate in the tomographic domain is a computationally involved optimization problem. Furthermore, discrete-valued MAP reconstruction requires accurate knowledge of the density or emission levels in the cross-section. In this paper we present an efficient multiscale algorithm for discrete-valued MAP reconstruction including estimation of the discrete levels. Experimental results indicate that the multiscale algorithm has improved convergence behavior over fixed scale reconstruction and is more robust with respect to local minima.