Robust rigid registration by scanning multiple phase correlation peaks

Rigid image registration is an important image processing problem with applications in medical imaging, biometrics, image mosaicking, video stabilization, and others. A popular technique for rigid image registration consists in estimating the location of the maximum of the phase correlation function. However, multiple motions, non-rigid transformations, non-textured regions and noise introduce artefacts and spurious peaks that may confound the phase correlation method. A solution, which is proposed here, is to explore multiple peaks and estimate an error measure for each plausible transformation, in order to find the correct one. Our tests, performed with both artificially generated and real cases, show that the proposed approach can increase the probability of successful registration by 25–40% and reduce the registration error by up to 35%.