Robust Training of the Quadratic Classifier
Abstract
The quadratic classifier is one of the most applied parametric classifiers used in pattern recognition. To use this classifier, one trains it by estimating the center and the dispersion of the different classes from the data. These estimates are usually made using sample means and sample covariances. If the data errors are normal, this is the optimal procedure. However, in practical situations where the data are not normal or contain outliers, the training can fail because the estimation procedure is not robust. This technical memorandum describes a robust method of estimating these parameters. This estimation method is much more resistant to outliers and perturbations from the assumed normal distribution than existing methods.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 02, 1994
- Accession Number
- ADA640494
Entities
People
- Paul R. Kersten
Organizations
- Naval Undersea Warfare Center