Neural networks for electronic eyes

It is important to integrate neural network's capability into traditional vision techniques so that systems display the flexibility and reliability closer to what is inherent in biological vision systems. Neural networks are useful at various levels in electronic eye applications, including low-level processing (feature extraction, active sensing, object detection) and high-level processing (object recognition, scene analysis, decision, control, and system integration). Neural networks are amenable to systems with a large number of processing cells enhanced by hierarchically structured interconnection. Their effective application-specific implementation hinges upon a thorough understanding of hierarchical network structure, training efficiency, real-time retrieving, and parallel processing technology. Particularly promising is a recently proposed decision-based Neural network (DBNN). We shall describe a face recognition system based on DBNN, developed jointly by Siemens Corporate Research and Princeton University. The system has yielded very high recognition accuracies based on experiments on public and in-house face databases. Issues on implementation and technology integration will be addressed. In terms of recognition accuracies, processing speeds, and parallel processing, the hierarchical DBNN perform far superior to that of the conventional multilayer perceptron (MLP).