STRUCTURED NEURAL NETWORKS FOR SIGNAL CLASSIFICATION

In this paper, artificial neural networks are considered as an emergen t alternative to the classical 'model-based approach' to the design of signal-processing algorithms. After briefly examining the pros and co ns of the neural-network approach, we propose the application of struc tured neural networks (SNNs) for the classification of signals charact erized by different 'information sources', such as multisensor signals or signals described by features computed in different domains. The m ain purpose of such neural networks is to overcome the drawbacks of cl assical neural classifiers due to the lack of general criteria for 'ar chitecture definition' and to the difficulty with interpreting the 'ne twork behaviour'. Our structured neural networks are based on multilay er perceptrons with hierarchical sparse architectures that take into a ccount explicitly the 'multisource' characteristics of input signals a nd make it possible to understand and validate the operation of the im plemented classification algorithm. In particular, the interpretation of the SNN operation can be used to identify which information sources and which related components are negligible in the classification pro cess. SNNs are compared with both commonly used fully connected multil ayer perceptrons and the k-nearest neighbour statistical classifier. E xperiments on two multisource data sets related to magnetic-resonance and remote-sensing images are reported and discussed. (C) 1998 Elsevie r Science B.V. All rights reserved.