Coastline Registration: Efficient Optimization in Large Dimensions Using Genetic Algorithms

Registration of remote imagery underpins a number of important applications in 'remote sensing' including mosaicing, sensor fusion, temporal change detection and integration with GIS databases. Accurate and automatic methods of establishing dense correspondences between images are vital given the increasing volume of high resolution imagery available. An explicit 'structural matching' framework is presented derived from the experience in the wider computer vision community of matching in a variety of domains including stereopsis, motion and object recognition. Match constraints can be combined within an optimization framework enabling the exploitation of a variety of minimization techniques. In this work, a multi-resolution representation of coastline contours is used to enable matching to handle very large scale difference between images, and to reduce the dimensionality of the match problem. Nonetheless, the resultant optimization problem remains huge and highly non-convex necessitating the use of 'genetic algorithms' to recover accurate correspondences.