A knowledge-based system for recognizing man-made objects in aerial images

A description is given of a knowledge-based vision system for recognizing and classifying man-made objects in aerial images. Images are interpreted and image object descriptors are created, based on model-driven high-level vision processing. Knowledge of various low-level vision techniques, and their applicability to generic applications are used to dynamically select specific low-level vision techniques for image segmentation. Terrain map and feature information is used as an adjunct by the low-level vision to assist in image segmentation, and by high-level vision for image interpretation. The authors have used a unique software architecture based on standard knowledge-based approaches in which knowledge is represented explicitly and is separated from program control. Off-the-shelf tools including LISP and the ART language from Inference Corporation, running on a Symbolics 3675, were used. The current system has been tested both with low-resolution forward-looking infrared (FLIR) images for target cueing and higher-resolution airport scenes for scene analysis