Image correlation, non-uniformly sampled rotation displacement measurement estimation

Abstract

A rotation invariant image correlation algorithm is described which measures and locates the rotational error of individual spatial frequency components. This information may then be interpreted to track signal dependent signatures, analyse spatial frequencies with much higher bandwidths and form optimal matched and generalised correlation filters. The technique is based on image re-sampling and a non-uniform sampling interval that is adjusted depending on its distance from the origin of the polar map. A nearest-neighbour polar interpolated grid scheme, comparable to linear interpolation error, achieves accuracies of 0.1% of a degree. Preliminary measurements based on images containing natural and rigid structure are presented. The algorithm also has potential applications for data-driven image registration and deformation analysis with small variations.