Abstract

A three-layer optoelectronic neural network for recognizing multiplex 3-D objects from arbitrary perspective views is demonstrated. Every kind of object is transformed into its feature codes and then classified by an associative memory of the codes. The experimental system is composed of two stages of different neural networks. The first network is implemented optically, and a bank of SDF filters are used as the interconnection weights. The optical system performs a hetero-association that encodes the projective images into a set of codes. Four kinds of aircraft are chosen as the examples, and 252 training images, which are obtained from different perspective views, are selected for each object. The four kinds of aircraft include airliner, fighter, bomber, and rocket. The experimental results show that the system can recognize correctly most of the projective images, including those outside the training set and partially hidden.

© 1992 Optical Society of America