A coarse to fine procedure for very precise and fast subpixel image registration in frequency domain

摘要

Automatic Image Registration (AIR) is a challenging task in many image processing applications. In this paper, an AIR method is proposed which is mainly useful for registration of satellite images that are captured in the same frequency band. In this work, images are registered precisely, by means of a coarse to fine approach which is fully implemented in the frequency domain. To find a coarse estimation of motion parameters, enhanced windowing and modified Vandewalle's method are introduced. Then, accuracy of rotation estimation is increased to a fraction of a degree, by use of Taylor expansion on the log-spectrum of the images. Afterwards, accuracy of shift estimation is enhanced to sub-pixel, by using the phase difference procedure. The performance of our method is evaluated in three experiments. Results show that the suggested framework is able to estimate large and small motion parameters with high speed and great accuracy. Suggested method is also robust to noise and is superior in performance to other five methods which are being used in this paper.