CLASSIFICATION OF MIXED-FONT ALPHABETICS BY CHARACTERISTIC LOCI.
Abstract
An alphabetic pattern is reduced to a vector of many dimensions (pattern features), each of which is the area of a locus (region) in the pattern's background. Each locus is the set of the points of the background that share, in a certain sense, a particular view of the pattern's line. A corrective learning process, executed by computer, linearly separates the vector space into alphabetic categories so that every pattern, of a large selection of patterns of nine different fonts, is found in its proper category, without regard for font and despite some noise. When the coefficients of the separating hyperplanes are then used to classify patterns not included in the learning sample, the error rate decreases sharply as the learning sample is increased. An implementation is described that would associate a digital computer with a special device that would include a matrix of simple, identical, intercommunicating microcells capable of determining the pattern's features. A method is described for designing by computer a binary decision tree that, using the coefficients determined by the learning process, would speed up the classification of patterns. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 1969
- Accession Number
- AD0694487
Entities
People
- Herbert A. Glucksman
Organizations
- Air Force Cambridge Research Laboratories