Knowledge-Based Systems Papers

 
Knowledge-Based Systems for Aerial Photo-Interpretation

For several years the MAPSLab has been exploring the use of knowledge-based systems for the interpretation of complex, high-resolution aerial photography. Our research has used generic task-domain knowledge with low and intermediate level scene segmentations to generate individual object interpretations and recognize consistent collections of objects.

SPAM, a production system architecture for the interpretation of aerial imagery, has been the focus and vehicle for most of our research in this area. SPAM integrates spatial, generic, and site-specific knowledge and coordinates scene segmentation tasks to produce a detailed description of a given scene in an aerial image. The SPAM system architecture interprets a two-dimensional scene segmentation as a collection of three-dimensional real-world objects. One of the first rule-based vision systems, SPAM contains over 600 OPS5 productions encoding spatial and geometric constraint knowledge about airports and suburban housing scenes. SPAM generates interpretations bottom-up, identifies mutually consistent groups, and builds scene models from these groups.

Current research involves the development of tools for knowledge acquisition and refinement, and the development of computational environments to support and exploit parallelism inherent to the scene interpretation process.

• A large knowledge-based system requires tools for knowledge-base management and development. Specialized tools are needed to maintain and extend the knowledge-base while preserving its consistency. Our research in this area includes the development of knowledge representations, knowledge compilation technology, interactive and automatic knowledge acquisition systems, and performance evaluation methods.

• As the knowledge base has grown in SPAM, its computational requirements can grow geometrically. We have pioneered the use of task-level parallelism in the context of knowledge-based systems. Task-level parallelism is obtained by decomposing the system at a high level, thereby exploiting the inherent parallelism in the task. In previous work done with the Production System Machine group at CMU, we have shown that using task-level parallelism can result in near-linear speed-ups on shared memory multiprocessors with between 16 and 27 processing elements. Together with the MIDWAY project at CMU, we have successfully implemented the task-level parallelism model on a distributed shared memory programming environment running on eight workstations using an ATM switch for communication.