Automatic and Precise Image Registration Based on Phase-Correlation Combined with Surface Fitting

The original method for image registration based on phase-correlation is only valid, if the shift vector is of integer values. So the error could be up to half of a pixel in theory. Since the estimation of the parameters on rotation and scaling is based on the estimation of these shift values, their accuracy is also restricted. In response to this limitation, we improved the method of search the maximum point of the impulse function. Using surface fitting methods like Parabolic Interpolation or Gaussian Function Fitting Method we could get the location of extreme point in sub-pixel level. So the error would be significantly reduced. The result of some registration experiments showed that the algorithm was valid for dealing with shifted, rotated and scaled images. Through the analysis we can see the improved method greatly increased the accuracy of registration without increasing the computational complexity.