INVESTIGATIONS INTO SOME AUTOMATIC IMAGE REGISTRATION TECHNIQUES

Abstract

Image registration is a key to many image processing tasks such as image fusion, image change detection, GIS overlay operations, 3D visualization etc. The task of image registration needs to become efficient and automatic to process enormous amount of remote sensing data. A number of feature and intensity based image registration techniques are in vogue. The aim of this study is to evaluate the applicability and performance of the two intensity based similarity metrics for temporal and view point image registration. The two metrics namely Mutual Information and Cluster Reward Algorithm have been explored in this study. Image registration task has been mapped as an optimization problem. A global optimizer namely Genetic Algorithm and a local optimizer namely Nelder Mead Simplex algorithm have been used to search registration parameters from the coarsest to the finest level of the image pyramid formed using wavelet transformation. For sound investigations, registration of remote sensing images acquired with varied spatial, spectral, radiometric characteristics have been considered. The data sets include images from IRS 1C LISS III, ERS 2 SAR, Landsat ETM+ and Terra ASTER sensors covering visible, near infra red and microwave regions of the electromagnetic spectrum. The image registration experiments suggest that both the similarity metrics have the capability of successfully registering the images with high accuracy and efficiency. In general, Mutual Information has yielded more accurate results than Cluster Reward Algorithm. The limitations of intensity based techniques have also been highlighted..