Volume 14, Number 3
PATTERN RECOGNITION LETTERS
heterogeneous methodologies for intelligent solution of pattern recognition problems and some of the design prin- ciples we have found useful. In that paper and in Raghavan and Kanal (1992), we describe some proof-of-concept systems developed by my associates at LNK Corporation. These include a multilevel hybrid of several types of neural networks for classification of Radar Cross Sections (Figure 2); and an integrated platform for intelligent fusing of 3 components-sensors, spatial databases, and maps, using neural networks, an expert system,
Figure 3. A hybrid system for feature extraction using sensor fusion.
and a fuzzy logic controller (Figure 3). From our experience, as outlined in the above mentioned papers, hybrid systems provide a key to achieving significantly better performance than individual component methodologies, in solving complex pattern recognition problems. A recently published book (Kandel and Langholz (1992)) appears to contain an excellent collection of papers on hybrid architectures and applications for intelligent systems.
A significant program based on hybrid pattern recognition methodologies, including neural nets, expert systems, and decision trees, is the Intelligent Data Management (IDM) project at NASA/Goddard Space Flight Center. The research being done for this project is described in a series of papers (see Campbell et al. (1989), Cromp (1991)). The Earth Observing System (EOS) is expected to generate massive quantities of satellite imagery and other data from a variety of instruments. While being rapidly archived, this data needs to be pro- cessed by fast algorithms which can characterize the data in such a way as to allow efficient querying and retrieval from object-oriented databases. An Intelligent Information Fusion System (IIFS) for the IDM is illustrated in Figure 4. Applications such as the NASA IDM point to the problem of scalability which remains one of the basic concerns for employing various pattern recognition, parallel processing, and machine intelligence tools on real-world problems. Hierarchically organized hybrid systems are being explored as one alternative to meeting the challenge of scalability, i.e., having such systems work well on problems involving very large data sets and tough real time constraints.