Investigation of Classifier Algorithms (CAPIR).
Abstract
The overall objective of this effort is to develop and study an optimal classifier so that future ATR designers can concentrate on feature extraction. The specific tasks of this project are: 1) Perform feature extraction and analysis on ERIM-supplied segmentor outputs; 2) Develop probability density estimates and figure-of-merit computations for kernel-based classifies; 3) Develop estimates for confidence in the probability density estimates; and 4) Evaluate and demonstrate classifier using above items. This report is divided into nine sections. Section 2 details the feature extraction task and includes information about the exchange of data with ERIM. Section 3 describes the classifiers used in this study as well as the window-width optimization problem in producing density estimates with the kernel-based classifier. Section 4 describes the confidence measures used to attack task C. Section 5 presents the results of th study on the development set of images, while Section 6 reports results for the characterization set of images and comparisons of these results to the development set results. Section 7 gives the results of our window-width optimization study while Section 8 presents confidence measure experiments. Finally, Section 9 gives the conclusions of this study.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 1987
- Accession Number
- ADA178410