Local similarity preserving projections for face recognition

In graph embedding based learning algorithms, how to construct the local neighborhood graphs in applications is a difficult but important problem. In this paper, we propose a novel supervised subspace learning method called local similarity preserving projections (LSPP) for linear dimensionality reduction (DR). LSPP seeks to project the original high-dimensional data into a subspace, which preserves the local neighborhood structure of the data in a certain sense. Compared with most existing DR algorithms, such as locality preserving projections (LPP) which is unsupervised in nature and predefines the neighborhood parameters, LSPP takes special consideration of class information to guide the procedure of graph construction, which effectively avoids the difficulty of neighborhood parameter selection and shows more valuable discriminatory information for classification tasks. To evaluate the performance of LSPP, we conduct extensive experiments on three face databases, i.e. Yale, FERET and AR face datasets. The results corroborate that LSPP delivered promising performance compared with other competing methods such as PCA, LDA, LPP, Supervised LPP, LDP, SLPP and MFA.