| Image Registration is the process by which the most accurate match is determined between two images. Images considered are misaligned by a four parameter rigid transformation, consisting of scale, rotation and/or x- and y-translations. The search for the matching transformation can be automated with the use of a suitable metric. This can be time consuming and somewhat tedious, and a simple search strategy based on a stochastic gradient is used to speed up the automated search. In this work we use mutual information as a similarity measure, together with a wavelet-based multi-resolution pyramid to speed up the overall registration process. We then modify the stochastic gradient optimization algorithm to include second-order Hessian effects with the aim of accelerating its convergence rate to get even closer to the optimum. Two sets of results are presented here. Results of the first order optimization scheme are presented comparing two similarity metrics, Mutual Information and Correlation. Data is provided to show that the second order optimization scheme can be configured to act like the first order scheme by suitable choice of parameter value. |
