Image Rotation from the Quadratic Sum of Phase Correlation

This paper proposes a new method to align two images that are rotated and translated with respect to each other. Existing approaches usually take the strategy of constructing translation invariant representations before recovering rotation as a result of phase correlation between them. In comparison, our method directly constructs an evaluation function of rotations, which comes from the quadratic sum of the one dimensional phase correlation between the Radon transforms of the two images to be aligned. The coordinate of this function’s maximum value gives the estimate for rotation. Experimental results show that our method compares favorably to phase-only bispectrum based method and can estimate large rotations and medium translations both robustly and accurately.