ABSTRACT
A general approach to the application of genetic algorithms (GA) to automatic scene labeling task is presented. The remote sensing application discussed involves labeling the oceanic features of the North Atlantic Gulf into known classes from presegmented images. Every segment in the image is an instance of some class and inherits all the predicates associated with it. Each solution evolved by the GAs represents a complete labeling of the image. Fitness of such a solution is computed as a bottom-up measure of the satisfaction of domain constraints. GAs have been applied to several aspects of image understanding. Ignoring the considerable work concerning GAs and neural nets, image applications of GAs may be divided into two categories—those that determine settings for real-valued parameters and those that solve for combinatorial aspects of the problem. The Building Block Hypothesis suggests that the power of GAs is derived from their ability to combine smaller building blocks of solutions together into better solutions.
