SCALE, TRANSLATION, AND ROTATION-INVARIANT ORTHONORMALIZED OPTICAL OPTOELECTRONIC NEURAL NETWORKS

We use a higher-dimensional version of the one-dimensional scale, translation, and in-plane rotation invariant transforms of Fang and Hausler [Appl. Opt. 29, 704-708 (1990)] in conjunction with an orthonormalization technique in an optical or optoelectronic resonator neural network. The system is tested by computer simulations that use a number of realistic stored and input images. Type-I (in-class discrimination) and type-II (out-of-class discrimination) false-alarm rates for several distortion types as well as results for individual examples of distorted images are presented. Our results indicate that the two-dimensional transforms exhibit considerably lower type-I false-alarm rates than the one-dimensional ones. They also show that such a configuration is capable of identifying a set of diverse inputs with cluttered and noisy backgrounds.