Phase-correlation (PC) is a computationally efficient method for two and three dimensional translation estimation. We presents a projection operator which significantly improves the accuracy and robustness of the PC scheme. The operator projects the estimated correlation function into the space of correlation functions resulting from a certain range of translations, while rejecting components which are unrelated to the estimated motion. Thus, the registration accuracy is improved by an order of magnitude, especially in the registration of noisy images and volumes. In addition, this approach is shown to be complementary with other subpixel phase correlation based techniques.