Abstract

A new edge detection method is presented which borrows from recent research into primate vision biology, and offers improved noise performance over classical methods. Beginning with spatio-temporal shunting models for retinal cones, horizontal cells, bipolar cells, and retinal ganglions, a set of simplified steady-state solutions are developed which lend themselves to efficient computation on standard computer equipment. The retinal model output is found to be nominally equivalent to the classical edge detector, but is produced differently. A simplified model of the lateral geniculate nucleus (LGN) has been produced. Taking the output of the retinal model, the LGN simple cell and interneuron models perform noise reduction and segment completion. An orienting subsystem is used to adaptively infer segment strengths and orientations, throwing out spurious and foreshortened edges, while retaining and filling in the longer edges

Meeting Name

International Joint Conference on Neural Networks, 1999

Department(s)

Engineering Management and Systems Engineering

Keywords and Phrases

2D Pattern Recognition; Connectionist Model; Edge Detection; Feature Extraction; Image Recognition; Image Segmentation; Interneuron Models; Lateral Geniculate Nucleus; Neural Nets; Physiological Models; Retinal Model

Document Type

Article - Conference proceedings

Document Version

Final Version

File Type

text

Language(s)

English

Rights

© 1999 Institute of Electrical and Electronics Engineers (IEEE), All rights reserved.

Full Text Link

Share

 
COinS