Abstract: The traditional FFT-based algorithm is found to be less stable numerically to register images with translations, scalings and rotations. This paper proposes a new FFT-based approach to estimate these affine transforms. We have proved that the cross-power spectrum method is poor in registering basically monotone-smooth signals with translations. These signals often exist in Fourier-magnitude representation data of an image in a log-polar coordinate system for each polar angle. By truncating the data at low coefficients in the logarithm coordinate system, the local maximum effect of undersirable signals can be reduced. Experimental results show that the proposed method outperforms the traditional FFT-Based image registration algorithm.

Key words: aliasing, cross-power spectrum, interpolation, registration