RECOGNIZING BUILDING PATTERNS USING MATCHED-FILTERS AND GENETIC SEARCH

This paper is concerned with recognising buildings in aerial images. W e abstract the images in terms of relational graphs. Specifically, we use Delaunay triangulations to represent the arrangement of located bu ildings. Localisation is realised using matched filters. The filters a re trained by drawing upon the duality between convolution in the imag e domain and multiplication in the Fourier domain. The matched filters prove to be remarkably stable. We match Delaunay graphs representing image pairs using genetic search with a Bayesian relational consistenc y criterion as fitness function. The use of genetic search allows us t o perform the optimisation without the traditional problems of sensiti vity to initial conditions and convergence to local optima. (C) 1998 E lsevier Science B.V. All rights reserved.