Locality Preserving Discriminant Projections

摘要

A new manifold learning algorithm called locality preserving discriminant projections (LPDP) is proposed by adding between-class scatter matrix and within-class scatter matrix into locality preserving projections (LPP). LPDP can preserve locality and utilize label information in the projection. It is shown that the LPDP can successfully find the subspace which has better discrimination between different pattern classes. The subspace obtained by LPDP has more discriminant power than LPP, and is more suitable for recognition tasks. The proposed method was applied to USPS handwriting database and compared with LPP. Experimental results show the effectiveness of the proposed algorithm.