The exploitation of results provided by Automatic Feature Extraction (AFE) and Automatic Target Recognition (ATR) algorithms can be greatly facilitated by analysis of the spatial relationships between the features and/or targets. This type of analysis can, when appropriate, automatically group objects and targets to produce higher-level descriptions about the content of any scene. We call this type of higher-level analysis Automated Scene Description (ASD).

One of the most powerful tools to describe the relative position of one object to another is the histogram of forces. This quantitative representation of the relative position between two objects encapsulates structural information about the objects as well as information about their spatial relationships. It is sensitive to the shape of the objects, their orientation, their size, and the distance between them. It lends itself, with great flexibility, to the modeling of spatial relationships for ASD applications.

Spatial relation models can use the force histograms as input into a fuzzy logic rule base that then automatically produces linguistic scene descriptions of the content of an image. The objects in the image/scene can be identified using AFE and/or ATR methods as appropriate. Spatial rules and meta-rules can be constructed to group the features/objects together into meaningful higher-level descriptions. In addition, each linguistic description has a natural confidence value associated with it. This process is illustrated below.

An example of the combination of ATR algorithms with automated scene description is shown below. Here, the MSNN ATR algorithm has been used to simultaneously detect and classify the targets in the scene. The final (maximum confidence) linguistic description generated by the fuzzy rule system is then output. The underlined terms surround, convoy, SAM site, etc. are generated automatically by the linguistic scene methodology.

We are presently conducting research to extend our ASD research to work with AFE inputs and more complex scenes from high-resolution satellite imagery over urban areas. In addition, we are pursuing research to develop a methodology to incorporate other intelligence information (human, signals, etc.) into the ASD methodology. We are investigating the development of appropriate "intelligence" meta-rules into the scene understanding rule base.

We believe that a common fuzzy logic framework will allow us to build a natural ontology for representing this non-image based intelligence. For example, human intelligence and other intelligence sources often contain fuzzy spatial descriptions (nearby, in the general vicinity, parked around, etc.) that must be fused with geospatial data to present a complete intelligence picture.