Task driven image understanding: lisp programming for vision research

The artificial intelligence aspects of computer vision make LISP appropriate for some vision re search needs. At the University of Washington an experimental image understanding system has been implemented in MACLISP. Since MACLISP supports arrays and compiles into reasonably efficient code, it has been convenient for the project. The major subsystems are the following: preprocessing, region growing, feature computation, and control. A novel aspect of the system is its control structure which consists of a "task agenda" which is processed by procedures each based on heuristics relevant to image analysis. Processing one task may create other tasks and the system is designed to suggest alternatives during the course of analyzing an image. Tasks are ordered by their heuristically-computed "priority" values. The system has been applied to the problem of segmenting aerial photographs of an airport. Design issues and possible extensions for further work are discussed.