Subpixel registration of images

Phase correlation has been studied as a tool for registering pairs of images to subpixel resolution. Shekarforoush et al. (1994) investigated an algorithm that works well for when observed data are samples of ideally bandlimited images. In fact, images of real scenes captured by modern optics are not ideally bandlimited images. This paper examines how aliasing in sampled images effects the ideal phase relationships between continuous images undergoing translational shifts. We show that the performance of standard phase correlation methods for subpixel image registration can be severely degraded with moderate amounts of aliasing, and we investigate two approaches for modeling aliasing effects in order to improve subpixel image registration accuracy. The first approach, a very simple algorithm, both conceptually and computationally is based on detecting those frequency components that have become unreliable estimators of shift due to aliasing, and removing the detected components from the shift estimate. The second approach, more ambitious and more complex, attempts to undo the effects of aliasing and to use all de-aliased frequency components in the shift estimator.