Machine Vision and Applications, vol. 4, no. 2, pp. 59-87, Spring 1991.
Abstract
In this paper, we propose a unified optimization framework for feature extraction that lets us simultaneously take into account image data and semantic knowledge: We model objects using a language that specifies both photometric and geometric constraints and define an information-theoretic objective function that measures the fit of the models to the data. We then treat the problem of finding objects as one of generating the optimal description of the image in terms of this language. We have validated our framework by performing extensive experiments on detecting objects in aerial imagery described by simple geometric constraints, and have developed two algorithms for generating optimal descriptions. The first starts with a rough sketch of a polygonal object and deforms the initial contour to maximize the objective function, thus finding object outlines. The second automatically extracts complex rectilinear buildings from complex aerial images.
