Multi-Reference Frame Image Registration for Rotation, Translation, and Scale; Master's thesis

摘要

This thesis investigates applications of multi-reference frame image registration for image sets with various translation, rotation, and scale combinations. It focuses on registration accuracy improvement over traditional pairwise registration, and also compares the quality of scene estimation from frame averaging. Three experiments are developed which use cross-correlation to estimate translation, the Radon transform to estimate translation and rotation, and the Fourier-Mellin transform to estimate translation, rotation, and scale. Results from applying multi-reference frame registration in these experiments show distinct improvements in both registration accuracy and quality of frame averaging compared to single-reference frame registration. Furthermore, it is shown that the new registration technique is equivalent to the optimal Gauss- Markov estimator of the relative shifts given all pairwise shifts.