| We describe and evaluate a system for segmenting and matching multisensor aerial images including synthetic aperture radar (SAR), infrared (IR), and electro-optical (EO) images. This system can be an important aid to aerial navigation. Multisensor aerial images are difficult to match because of great differences in distortions and artifacts produced by diverse sensors, as well as variability in points of view, terrain, weather, and illumination. We introduce a class-scale space concept for segmenting these images and describe a system that exploits this concept.; Our segmentation technique consists of three stages: multiscale feature extractor, multiclass pattern classifier, and class-scale logic. In the multiclass pattern classifier, an array of genetic algorithms selects a subset of features for classification---one genetic algorithm for each class-scale pair. A second array of genetic algorithms optimizes the initial weights of an array of neural classifiers. After training, the array of neural classifiers produces an array of segmented images, one image for each class-scale pair. Class-scale logic combines these images in a manner that models human visual interpretation. This results in a final segmented image that combines several classes of coarsely detected regions with finely detected curves and points. We describe applications of these techniques to segmenting and matching SAR, IR, and EO images. |
