PDF version of this document

Adaptive Precision for Images

This work is associated with the change in optimiser tolerance and the context is now different. This approach was never tried previously, but it seems to get good results and it incorporates a novel and elegant algorithm which seeks a multi-scale approach to choice of warps. Rather than scaling the data, it scales the level of warp quality and, in that sense, it is multi-scale in an unorthodox sense.

Figure: Automatic precision and the differing rates of convergence for image registration.

Figure shows how the choice of tolerance (or precision homologously) affects the rate of convergence. What is less obvious to the mind, is the way this optimisation improves in terms of time. While this is a possible experiment to perform at present, registration versus time will be shown only later, when a different approach is discussed.

Figure shows quite clearly the differing rates of objective function improvements in its evaluation. Different curvatures correspond to different choices of tolerance.

Figure: Adaptive precision requirement resulting in different rates of convergence. This curve is drawn in the context of shapes and selection of landmark point.