A New Manifold-Based Feature Extraction Method

摘要

Many traditional feature extraction methods takes the global or the local characteristics of training samples into consideration during the process of feature extraction. How to fully utilize the global or the local characteristics to improve the feature extraction efficiencies is worthy of research. In view of this, a new Manifold-based Feature Extraction Method (MFEM) is proposed. MFEM takes both the advantage of Linear Discriminant Analysis (LDA) in keeping the global characteristics and the advantage of Locality Preserving Projections (LPP) in keeping the local characteristics into consideration. In MFEM, Within-Class Scatter based on Manifold (WCSM) and Between-Class Scatter based on Manifold (BCSM) are introduced and the optimal projection can be obtained based on the Fisher criterion. Compared with LDA and LPP, MFEM considers the global information and local structure and improves the feature extraction efficiency.