A multinetwork architecture for classification
Paolo Stofella, S Zecca, FRANCESCO GARDIN, Giancarlo Mauri · BOA (University of Milano-Bicocca) · 1992
This paper describes a methodology for building hierarchical structures of neural networks, in order to deal with very complex image classification problems, with hundreds or thousands of objects to be recognized. The hierarchy can be automatically built up or imposed by the user. In the learning phase, training cycles on the high level classes are used in training all the depending subclasses, leading to dramatic computational effort saving. The overall architecture has been implemented on a parallel computer system, exploiting parallelism both on the networks structure and on the hierarchical architecture. A set of experimental results on selected image classification problems is reported.