Abstract
The neocognitron is a multilayered feedforward hierarchical neural network model for visual pattern recognition that was first proposed by Fukushima.1 As opposed to some of the neural-network models (e.g., associative memory, perceptron, and adline), the neocognitron is a featu re-extraction-based system that offers the unique advantage of deformation and shift invariance. The primary difference between a neocognitron and other neural networks is that the neocognitron recognizes and extracts simple features from an input pattern in its beginning layer. Extracted features are then grouped for recognition in a deeper layer. This feature-extraction mechanism provides the foundation for deformation invariance. We have devised an optoelectronic system for the implementation of the neocognitron. Each generic layer of the neocognitron consists of a multichannel correlator and a two-dimensional electronicschip array. The function of the correlator is to perform parallel feature extraction with shift invariance. The electronics chip performs realtime thresholded correlation detection. The basic layer is used iteratively to realize a multilayer functionality. In this paper, the system issues in implementing this neocognitron will be discussed. Preliminary experimental results will also be presented.
© 1990 Optical Society of America
PDF ArticleMore Like This
Guogang Mu, Ying Sun, Yanxin Zhang, and Xiangping Yang
TuD2 OSA Annual Meeting (FIO) 1992
Tien-Hsin Chao
MD3 OSA Annual Meeting (FIO) 1992
Claude Lejeune, Young Sheng, and Henri H. Arsenault
MII2 OSA Annual Meeting (FIO) 1991