Survey of image registration methods

Abstract

Survey of image registration methods roger d. eastman, nathan s. netanyahu, and jacqueline le moigne 3.1 Introduction Automatic image registration, bringing two images into alignment by computing a moderately small set of transformation parameters, might seem a well-defined, limited problem that should have a clear, universal solution. Unfortunately, this is far from the state of the art. With a wide spectrum of applications to diverse categories of data, image registration has evolved into a complex and challenging problem that admits many solution strategies. The growing availability of digital imagery in remote sensing, medicine, and numerous other areas has driven a substantial increase in research in image registration over the past 20 years. This growth in research stems from both this increasing diversity in image sources, as image registration is applied to new instruments like hyperspectral sensors in remote sensing and medical imaging scanners in medicine, and new algorithmic principles, as researchers have applied techniques such as wavelet-based features, information theoretic metrics and stochastic numeric optimization. This chapter surveys the diversity of image registration strategies applied to remote sensing. The objectives of the survey are to explain basic concepts used in the literature, review selected algorithms, give an overall framework to categorize and compare algorithms, and point the reader to the literature for more detailed explanations. While manual and semi-manual approaches are still important in remote sensing, our primary intent is to review research approaches for building fully automatic and operational registration systems. Following the survey article by Brown (1992), we review an algorithm by considering the basic principles from which it is constructed. These principles include, among others, the measure of similarity used to compare images and the optimization algorithm used to optimize the measure. Most image registration algorithms in the literature and those used in practice are based on variations on these basic elements and their combination. Indeed, a reader familiar with the basic, major principles and their various combi-nations, can easily gain a good understanding of a new image registration technique 35 P1: SFK Trim: 174mm × 247mm Top: 0.553in Gutter: 0.747in CUUK1136-03 cuuk1136/Le-Moigne ISBN: 978 0 521 51611 2 or system. A survey of the scientific literature on image registration specifically for remote sensing is provided by Fonseca and Manjunath (1996). Previous surveys of the general image registration literature include Brown (1992) and Zitová and Flusser (2003), while general books on registration include Modersitzki (2004) and Goshtasby (2005). Surveys limited to medical applications, with specific focus on mutual information, include Maintz and Viergever (1998) and Pluim et al. (2003). Within the wide spectrum of image registration principles and applications, this chapter focuses on techniques relevant to automatic registration of regular two-dimensional image data from Earth satellite instruments used for remote sensing, primarily those instruments sensing in the visible or near-visible spectra. We do not treat extensively methods for instruments that use radar, those that directly produce range information, or sensors borne aloft by airplanes and balloons. These forms of imagery may differ from satellite imagery in perspective and other characteristics. However, since new image registration techniques are often imported into remote sensing from research undertaken in other fields, we also review some methods from those fields as appropriate. Other fields with active research in image registration include medical imaging, video analysis for multimedia, and robotics. These fields differ from remote sensing in their requirements for image registration. In contrast to remote sensing, medical image registration works with a large number of imaging modalities from whole body to retinal scans, and involves tissues that can deform or change drastically. Similarly, image registration for multimedia video analysis and robotics works on short-range imagery of complex 3D scenes, rather than long-range imagery of planetary surfaces. We specifically do not include in our review articles on elastic or nonrigid registration, 3D volumetric registration, and 3D range registration. Readers interested in the latter topics are referred to Maintz and Viergever (1998), Lester and Arridge (1999), Goshtasby (2005), and Salvi et al. (2006). Regardless of the field of application, the articles reviewed in this chapter originate roughly from three overlapping communities. The image processing community focuses on developing methodologies and techniques, and a research contribution presents usually a novel technique or an evaluation method. Such contributions include mostly proof-of-concept demonstrations on limited image sets so the techniques described can be considered promising prototypes. A sec-ond set of contributions comes from the community of ground support satellite teams. This community specializes in the design of effective, practical algorithms used to create orthorectified image products for a limited set of instruments. The techniques they describe serve as a basis for successful operational systems run on extensive datasets. Operational satellite teams have an end-to-end understanding of data processing used in their satellite system, with inside knowledge of sensor and satellite engineering. They use image registration to update orbital and navigational