The traditional polar transform usually suffers from the non-uniform sampling problem, which means that the low-frequency components are often over-sampled, while the high-frequency components are relatively under-sampled. Consequently, the inappropriate sampling rates will affect the registration accuracy, or else increase the computation cost vainly. To conquer the drawbacks mentioned above, we develops a novel frequency-domain registration algorithm using the improved polar transform. The reference image and the image to be registered are both carried out Fourier transform individually, and the corresponding spectrum images are sequentially mapped into the improved polar coordinate. Then projection operations are done along the angular and radius direction, respectively. As a result, the rotation and scale parameters between the two spatial images can be easily induced from the corresponding projection curves. Eventually, the shift parameters are retrieved with the weighted phase difference, after the inverse rotation and scale operations are implemented for the image to be registered. The experimental results show that the registration precision of our algorithm is much higher than the algorithm using the traditional polar transform or the pseudo-polar transform, while the required computation costs are almost equivalent.