Abstract

An optoneural system, which consists of an optical correlator and a neural network, is developed for invariant pattern recognition. The correlator uses Fourier-Mellin spatial filters (FMF) for feature extraction. The impulse response of a FMF is equal to the kernel function of the circular-Fourier and radial-Mellin transform. The filter itself contains no object information and yields an unique output for each input object. The features used as input to the neural network are the geometrical parameters of the 2-D pattern of the output local peaks. The neural network used is a multilayer feed forward net with a back propagation learning rule. The advantages of this approach are that a FMF may be used for all input objects without the need for training or updating the filter, and that the number of the extracted features is small, making it possible to use a small neural network. This Fourier-Mellin optoneural system shows multiple object recognition, which is invariant not only to rotation and scale changes but also to translations of the input objects.

© 1991 Optical Society of America