Regularization of LDA for Face Recognition: A Post-processing Approach
When applied to high-dimensional classification task such as face recognition, linear discriminant analysis (LDA) can extract two kinds of discriminant vectors, those in the null space (irregular) and those in the range space (regular) of the within-class scatter matrix. Recently, regularization techniques, which alleviate the over-fitting to the training set, have been used to further improve the recognition performance of LDA. Most current regularization techniques, however, are pre-processing approaches and can’t be used to regularize irregular discriminant vectors. This paper proposes a post-processing method, 2D-Gaussian filtering, for regularizing both regular and irregular discriminant vectors. This method can also be combined with other regularization techniques. We present two LDA methods, regularization of subspace LDA (RSLD) and regularization of complete Fisher discriminant framework (RCFD) and test them on the FERET face database. Post-processing is shown to improve the recognition accuracy in face recognition.
KeywordsFace Recognition Linear Discriminant Analysis Recognition Rate Regularization Technique Average Recognition Rate
Unable to display preview. Download preview PDF.
- 4.Torkkola, K.: Linear discriminant analysis in document classification. In: Proc. IEEE ICDM Workshop Text Mining (2001)Google Scholar
- 14.Zhao, W., Chellappa, R., Krishnaswamy, A.: Discriminant analysis of principal components for face recognition. In: Proc. Int’l Conf. Automatic Face and Gesture Recognition, pp. 336–341 (1998)Google Scholar
- 16.Liu, C., Wechsler, H.: Enhanced Fisher linear discriminant models for face recognition. In: Proc. 14th Int’l Conf. Pattern Recognition, vol. 2, pp. 1368–1372 (1998)Google Scholar