Abstract

The cascaded model of neural networks proposed in this paper is a new approach to pattern recognition. The motivation comes mainly from the fact that the human brain is functionally subdivided into smaller parts. The model consists of some subnets that are functionally complementary and hierarchical, and therefore can enhance the performance of the entire system. Furthermore, the learning process of the cascaded model is carried out separately in each subnet, where the learning rule is much simpler. Therefore, the difficulty of subjecting the model to convergence on a local minimum, as is often the case in multi-layered neural networks, can be avoided. The design of a cascaded system is somewhat similar to that of an optical or electronic system composed of cascaded components. As an example, a 3-D target classifier for four kinds of aircraft with arbitrary spatial orientation is demonstrated based on the cascaded model. Computer simulation and experimental results have shown the model can correctly classify input patterns, both inside and outside the training sets, with rotational invariance.

© 1992 Optical Society of America