A novel improved linear discrimnant analysis (ILDA)method is presented. Comparing with LDA, under the condition of d ≪ c ȡ2;1, d and c are the dimensionality of feature subspace and the number of classes respectively, ILDA uniformly preserves the class distances of classpairs by rearranging the contribution of each class-pair to the generalized between-class scatter matrix after whitening within-class scatter matrix. Experiment results based on simulating data and measured radar data both show that, under the condition of d ≪ c ȡ2;1, the features extracted by ILDA are more efficient for multi-class classification than those extracted by LDA.
展开▼
