Researches On Image Registration

Image registration is a vital preprocessing step in image processing. The basic methods for this are feature-based methods and area-based methods.A fast algorithm based on covering theory is developed for pattern matching. Using the sub-image from the search image, we can get a high speed for searching the template image in the search one by switching the roles of the template image and the search image.For the problems of image registration in the Fourier domain, we have studied the phase of the classical smooth signals, and give the reason why the phase correlation method may give wrong results in translation detection for decreased smooth signals. We also present an Pseudo-Log-Polar Fourier based registration algorithm called PLPFFT algorithm for estimating translations, rotations, and scalings (TRS) among single sensor images, multi-sensor images and multi-spectral images. Interpolated by pseudo-log-polar discrete Fourier transform (DFT), the log-polar DFT values of images can be approximated more accurately than conventional techniques, and the detection ability is enhanced for single sensor images. An extra step is added for multi-sensor/spectral images. This is done by edge detection, and salient and distinctive features can be extracted as the images for registration. Experimental results with various kinds of remote sensing images have verified robustness and high accuracy of this algorithm.