A feature extractor for curvilinear patterns: a design suggested by the mammalian visual system

Summary

A feature extractor for a pattern recognizer which can effectively process curvilinear drawings has been synthesized and simulated on a digital computer.

The design of the network was suggested by the visual system of higher animals — especially the structure of the receptive fields of cortical neurons. This feature extractor is a multilayered parallel network composed of “analog threshold elements”. It consists of six layers in cascade. The first layer is a two-dimensional array of photoreceptors. The second layer is a contrast-detecting layer, each element of which has an “on”-center-type receptive field. The third one is a line-detecting layer. An element of this layer corresponds to a “simple” cortical cell, and responds to lines whose orientation is proper for the element. Each element has a receptive field consisting of an elongated excitatory region flanked on either side by inhibitory regions. The fourth layer is also a line-detecting layer, but each element, which corresponds to a “complex” cell, is not sensitive to the exact position of the line. An element of the fifth layer, which may correspond to a “hypercomplex” cell, responds when the line detected in the preceeding layer is curved. In the final layer, the curvature of the line is detected regardless of the orientation of the line, that is, an element of this layer gives an output approximately proportional to the curvature of the line presented in its receptive field.