A Neural Network Approach to Model-Based Recognition
Abstract
This project grew out of earlier work in which an approach to visual object recognition as an associative memory problem was pursued. The big problem with this earlier approach is that objects must be recognized regardless of changes in position, rotation, scaling, and a host of other deformations. One may store iconic patterns directly in a memory and design such invariances directly into the connection pattern, but the circuits quickly become complicated. We follow the approach taken in traditional computer vision systems and store relational models instead of iconic models into the circuit, and demand that the input data itself be organized into such a relational structure and optimally matched to the nearest model. Relational models are designed to automatically capture the desired invariances; the perceptual organization of input data into relational structure proceeds simultaneously with the matching process. Like some associative memory approaches, this one also uses objective functions for specification of the problem and design of the circuit. Early work on optical implementations was also continued in this grant. The original work focused on the use of optics for fast recognition of two dimensional geometrical patterns independent of scale, rotation, translation. By stages, the focus shifted toward implementation of neural nets to accomplish the same goal, and, by extension, toward the general problem of optical interconnects.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 26, 1990
- Accession Number
- ADA231835
Entities
People
- Gene R. Gindi
Organizations
- Yale University