A Stochastic Approach for Non-Rigid Image Registration

This note describes a non-rigid image registration approach that parametrizes the deformation field by an additive composition of a similarity transformation and a set of Gaussian radial basis functions. The bases' centers, variances, and weights are determined with a global optimization approach that is introduced in this work. This approach consists of simulated annealing with a particle filter based generator function to perform the optimization. Additionally, a local refinement is performed to capture the remaining misalignment. The deformation is constrained to be physically meaningful (i.e., invertible). Results on 2D and 3D data sets demonstrate the algorithm's robustness to large deformations.