The Design of Automatic Pattern Classifiers
Abstract
This document exploits relational graphs and local training of piecewise linear decision surfaces to achieve improved procedures and insights in the design of automatic pattern classifiers. It uses the orientations of the linear segments of these classifiers as guides in a branch-and-bound search for optimal subsets of features. The author develops procedures for determining near-optimal sequences of features for piecewise linear classifiers. He develop new procedures for selecting small subsets of features from large initial sets for use in automatic pattern classifiers. His interactive relational graph design technique for the design of piecewise linear two-class classifiers is extended to the design of multiple-class classifiers. This research led to solutions of several fundamental problems associated with the use of automatic classifiers of detected candidate military targets appearing in aerial and terrestrial images. These problems fall in three areas; data analysis, classifier design, and classifier architectures. The author developed a highly effective means of discovering clusters in multiple-dimensional data, using the CAMPA (Clustering and Mapping for Pattern Analysis) interactive software in conjunction with modern computer graphics. Keywords: Target classification; Infrared images.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 06, 1988
- Accession Number
- ADA200920
Entities
People
- Jack Sklansky
Organizations
- University of California, Irvine