Abstract

Processing images by a neural network means performing a repeated sequence of operations on the images. The sequence consists of a general linear transformation and a nonlinear mapping of pixel intensities. The general (shift variant) linear transformation is time consuming for large images if done with a serial computer. A shift invariant linear transformation can be implemented much easier by fast Fourier transform or optically, but the shift invariant transform has fewer degrees of freedom because the coupling matrix is Toeplitz. We present a neural convolution network with shift invariant coupling that nevertheless exhibits autoassociative restoration of distorted images. Besides the simple implementation, the network has one more advantage: associative recall does not depend on object position.

© 1990 Optical Society of America

Three-dimensional imaging: a space invariant model for space variant systems

David N. Sitter and William T. Rhodes
Appl. Opt. 29(26) 3789-3794 (1990)

Parallel distributed processing model with local space-invariant interconnections and its optical architecture

Wei Zhang, Kazuyoshi Itoh, Jun Tanida, and Yoshiki Ichioka
Appl. Opt. 29(32) 4790-4797 (1990)

Shift-invariant interpattern association neural network

Chii-Maw Uang, Shizhuo Yin, P. Andres, Wade Reeser, and Francis T. S. Yu
Appl. Opt. 33(11) 2147-2151 (1994)

Space invariant multiple imaging for hypercube interconnections

Yunlong Sheng
Appl. Opt. 29(8) 1101-1105 (1990)

Fabrication of space invariant fanout components in dichromated gelatin

Brian Robertson, Mohammad R. Taghizadeh, Jari Turunen, and Antti Vasara
Appl. Opt. 29(8) 1134-1141 (1990)