Cluster-based LDA for single sample problem in face recognition

The extreme case of the curse of dimensionality is that only single sample is available for each class, which is often true for face recognition. Consequently, Fisher linear discriminant analysis (LDA) cannot work due to disappearance of within-class scatter matrix. To tackle this problem, we propose to cluster the training samples firstly. Then within cluster scatter matrix can be computed. Substituting the within-class scatter matrix with the within-cluster matrix, we get a variant of the original LDA. Experimental results on FERET face databases show that the proposed method is promising.