Face recognition based on LDA in manifold subspace

Abstract

Although LDA has many successes in dimensionality reduction and data separation, it also has disadvantages, especially the small sample size problem in training data because the "within-class scatter" matrix may not be accurately estimated. Moreover, this algorithm can only operate correctly with labeled data in supervised learning. In practice, data collection is very huge and labeling data requires high-cost, thus the combination of a part of labeled data and unlabeled data for this algorithm in Manifold subspace is a novelty research. This paper reports a study that propose a semi-supervised method called DSLM, which aims at overcoming all these limitations. The proposed method ensures that the discriminative information of labeled data and the intrinsic geometric structure of data are mapped to new optimal subspace. Results are obtained from the experiments and compared to several related methods showing the effectiveness of our proposed method.