Learning Varying Dimension Radial Basis Functions for Deformable Image Alignment

This paper presents a method for learning Radial Basis Functions (RBF) model with variable dimensions for aligning/registrating images of deformable surface. Traditional RBF-based approach, which is mainly based on a fixed dimension parametric model, often suffers from severe parameter over-fitting and complicated model selection (i.e. select the number and locations of centers determination) problems which lead to inaccurate estimation and unreliable convergence. Our strategy for solving both the parameter over-fitting and model selection problems is through the use of a probabilistic Bayesian inference model to obtain a posterior estimation of the alignment as well as the model parameters simultaneously. To learn the parameters of the Bayesian model, a reversible jump Markov Chain Monte Carlo (RJMCMC) algorithm is employed, allowing us to handle large deformation image registration. Our approach is demonstrated successfully on real image sequences of different deformation types, with results compared favorable against other existing approaches