Japanese / English

Detail of Publication

Text Language Japanese
Authors Masakazu Iwamura, Shin'ichiro Omachi, and Hirotomo Aso
Title A Modification Method of Eigenvalues to Avoid Recognition Accuracy Reduction Caused by Estimation Errors of Eigenvectors
Journal IEICE Technical Report
Vol. 99
No. 449
Presentation number PRMU99-146
Pages pp.93-100
Reviewed or not Not reviewed
Month & Year November 1999
Abstract Statistical pattern recognition methods such as Mahalanobis distance need to know true distributions of the patterns, but it is usually impossible. As a result, the covariance matrices estimated from training samples are used. If training samples are not enough, estimation errors in both eigenvalues and eigenvectors will occur and they will cause recognition accuracy reduction. Many researchers have studied on estimation errors of eigenvalues and have proposed various methods to modify eigenvalues. However, estimation errors of eigenvectors have not been considered enough. In this paper, we investigate estimation errors of eigenvectors and show the possibility that they make recognition performance down. We propose a new method estimating eigenvalues so as to reduce bad influence on estimation errors of eigenvectors.
Back to list