2D Weighted Linear Discriminant Analysis and the Application for Feature Extraction
Xueyong Li, Changhou Lu · 2010
A new technique named 2D Weighted Linear Discriminant Analysis (2DWLDA) is proposed for feature extraction. The main idea of 2DWLDA is to reconstruct the between-class scatter matrix and set weighted coefficients to the covariance matrixes of class pairs. The method for computing the project vectors and the structure of the weighted function are also inferred. The proposed method has effectively overcome the suboptimal problem of traditional Fisher criterion under multiple classes. Experiments on the Pressed Characters database and the ORL database show that the proposed method has significantly improved the recognition rate while reduce the discriminant vector's number when got the maximum recognition rate.