The Unnoticed Mapping Problem in Phase Correlation Image Registration

Phase correlation image registration is a suitable approach to global motions including translation, rotation, and zoom. This motion model is of special interest in real-time applications for remote sensing, image super-resolution, and manipulation by industrial robots. However, it is common that phase correlation leads to incorrect zero degrees rotation estimations, when under real application conditions. These errors are usually attributed to rotation-dependent aliasing or spectral leakage. In this work, it is shown that this problem has a deterministic cause, being the result of the mapping from Cartesian coordinates to Polar coordinates foreseen in the method, apparently unnoticed by the scientific community in the last 25 years. Preliminary results illustrate what can be done for solving this problem and illustrate the performance gain.

Automated summary

Phase correlation image registration is a suitable approach to global motions. However, under real application conditions, it often leads to incorrect rotation estimations. This study reveals that the problem is caused by the mapping from Cartesian to Polar coordinates, which has been overlooked for the past 25 years. Preliminary results demonstrate the solution to this problem and the improvement in performance.