A generalized orthonormal discriminant vector method for pattern recognition
Yoshihiko Hamamoto, Shingo Tomita · Electronics and Communications in Japan (Part III Fundamental Electronic Science) · 1994
Abstract Discriminant analysis on one of the well‐known feature extraction methods. Unfortunately, discriminant analysis has a serious shortcoming in that the maximum number of features is limited by the number of pattern classes. To solve the forementioned shortcoming, Okada and Tomita proposed the orthonormal discriminant vector (ODV) method based on the Fisher criterion. However, no feature is provided by the ODV method when the mean vectors are the same. First, this paper proposes a new ODV method based on a generalized Fisher criterion which takes into consideration the difference of covariance matrices. Second, a theoretical comparison is made between the proposed method and the other feature extraction methods. Finally, experiments are conducted to show the usefulness of the proposed method.