Fast gradient distance based image registration

Abstract

Phase Correlation is many times faster than feature correspondence based image registration methods, but is still relatively computationally complex. To register images using phase correlation related by similarity transforms a translation invariant log-polar remapping of the images is required. This operation requires the computation of an additional two 2D FFTs plus two log-polar transforms. This paper proposes an alternative approach that eliminates the need to compute the log-polar transforms that reduces the number of FFTs required to register images and reduces overall computation time.