Incremental Weighted Linear Discriminant Analysis for Face Recognition

Abstract

Face recognition has been a popular research area with several real-world applications. Linear discriminant analysis (LDA) is a well-known method for face recognition in literature. However, one of the requirements of LDA is the availability of all data samples available before training. In this paper, we have proposed a novel variant of LDA method based on incremental learning and is called incremental weighted linear discriminant analysis (IWLDA). In IWLDA, a weighted pairwise Fischer criterion is suggested to efficiently separate all class data homogeneously in the transformed subspace. IWLDA is developed in such a manner that the distance between nearby classes is increased and simultaneously the distance between farther classes is reduced and the overall distance is preserved. This results in improved classification accuracy. Experimental results on 5 publicly available datasets, viz. AR, CACD, YaleB, FERET, and ORL show that the proposed method outperforms the popular methods, i.e., principal component analysis (PCA), LDA, incremental principal component analysis (IPCA), and incremental linear discriminant analysis (ILDA) on all the datasets in terms of K-fold (K= 2, 3, 4, 5) cross-validation. Further, it is also found that the training time of IWLDA is better than the batch methods, i.e., PCA and LDA.