Spectral image registration using multivariate mutual information.

摘要

Image registration is the operation of aligning images that, in general, are of the same scene but collected under different viewing conditions and/or using separate imaging devices. The cross-correlation between images and functions related to cross-correlation have long been used as measures for registration. In recent years, it has been recognized in the medical imaging and computer vision communities that statistical dependence is a more powerful measure of the relationship between two images and that measures of dependence used in information theory, such as mutual information, can be used to register imagery more reliably than correlation-based methods.; Remote sensing imagery typically has characteristics that differ significantly from those of medical imagery or imagery from computer vision applications. In particular, remote sensing imagery is often characterized either by subtle, very low-frequency contrast variations or rapid, high-frequency contrast variations that challenge any registration algorithm. A specific class of remote sensing, multispectral/hyperspectral imaging, can require the simultaneous processing of tens to hundreds of images that differ by wavelength and whose contrast variations are often related nonlinearly to one another. This class of imagery creates special problems in the performance as well as the efficiency of the registration algorithms. The logical approach is to perform an in-depth analysis of information theoretic techniques in order to determine their performance for remote sensing image registration and explore the efficiency of registering large image sets.; In this dissertation, a theoretical foundation is developed for mutual information image registration and validated experimentally on a variety of image types. A simple and powerful model of the mutual information registration function is presented and compared with experimental data. This model provides a method for predicting the performance of mutual information registration for a variety of applications and verifies the relationship between mutual information and the frequency content of the imagery.; Finally, an efficient method is presented for simultaneously registering a finite number of images using a multivariate extension of mutual information. This method permits the information available in all of the images to be exploited simultaneously and provides a practical method for registering multispectral and hyperspectral imagery.