Image registration using an adaptive LMS algorithm

摘要

The adaptive LMS registration algorithm has shown the potential to be as accurate as the Fourier shift theorem (FST) registration method for small translational displacements, however with larger displacements the accuracy of the LMS method can degrade relative to the FST approach. The LMS algorithm also appears to be less robust to noise than the FST. The LMS method was over twice as fast as the FST method for the experiments described with large unknown rotations. Further speed increases would be expected using a variable step-size algorithm and automatic detection of convergence. A variable step-size algorithm may also reduce the variations in performance between images of different structure. A cursory investigation of the effect of grey-scale stretching on registration error found no benefits when applied uniformly to the whole image.