Publications Details
One-class classifier networks for target recognition applications
Target recognition requires the ability to distinguish targets from non-targets, a capability called one-class generalization. Many neural network pattern classifiers fail as one-class classifiers because they use open decision boundaries. To function as one-class classifier, a neural network must have three types of generalization: within-class, between-class, and out-of-class. We discuss these three types of generalization and identify neural network architectures that meet these requirements. We have applied our one-class classifier ideas to the problem of automatic target recognition in synthetic aperture radar. We have compared three neural network algorithms: Carpenter and Grossberg`s algorithmic version of the Adaptive Resonance Theory (ART-2A), Kohonen`s Learning Vector Quantization (LVQ), and Reilly and Cooper`s Restricted Coulomb Energy network (RCE). The ART 2-A neural network gives the best results, with 100% within-class, between-class, and out-of-class generalization. Experiments show that the network`s performance is sensitive to vigilance and number of training set presentations.