PDF version of this document
The objective we are after is the minimisation of some cost - a cost is associated with the model and the images which we transform. Convergence of the algorithm is declared once that cost can no longer be reduced. We have found that very much computational effort is spent on refining existing transformations even when the solution is yet far away. As it turns out, at early stages of the optimisation, coarse changes to the images should suffice.
Amongst the more important developments of this project, we managed to find a way of substantially decreasing the amount of time which is required to register sets of 1-D data. This was done primarily by tweaking the optimisation which is involved in the process.
We decided to aim for a different optimiser tolerance depending on our advancement towards the correct solution. By doing do, little time should be spent on obtaining lower costs at the early stages of registration. This rational observation motivated us to implement a similar algorithm in a separate domain2.