We develop a two-stage, modular neural network classifier and apply it to an automatic target recognition problem. The data are features extracted from infrared and TV images. We discuss the problem of robust classification in terms of a family of decision surfaces, the members of which are functions of a set of global variables. The global variables characterize how the feature space changes from one image to the next. We obtain rapid training times and robust classification with this modular neural network approach.

Neural networks are used to analyze a complex simulated radar environment which contains noisy radar pulses generated by many different emitters. The neural network used is an energy-minimizing network. The limiting process contains the state vector within a set of limits, and this model is called the brain state in a box, or BSB model, which forms energy minima (attractors in the network dynamical system) based on learned input data. The system first determines how many emitters are present (the deinterleaving problem). Pulses from individual simulated emitters give rise to separate stable attractors in the network. Once individual emitters are characterized, it is possible to tentatively identify them based on their observed parameters. As a test of this idea, a neural network was used to form a small database that could potentially make emitter identification. There were three errors of classification. The number of iterations are required to reach an attractor state was very long, and some of the final states were not fully limited. These factors indicate the uncertainty of the neural network.>