Object recognition by a massively parallel 2-D neural architecture

Abstract

Listen

Translate article

The use of a massively parallel neural array for multiple 2-D object recognition is explored. The array architecture has a parallel modular form with each module being trained over a specific object class. One test bed is developed using alphabetic characters which have been subjected to a scale factor and rotational operations. This test bed provides a simultaneous measure of geometric invariance and of character recognition. The performance of the modular design is benchmarked against a backprop-trained multilayer perceptron network of equivalent generality. A second test of the modular array is conducted using TV and FLIR images. This second evaluation assesses the ability to extract obejct signatures from a clutter background.