关键词: 人脸识别, 特征提取, 多流形, 非监督线性差分投影（ULDP）

Abstract: To overcome the drawbacks of existing unsupervised linear differential projection (ULDP), this paper proposes a novel algorithm called multi-manifold unsupervised linear differential projection (MULDP) for face recognition. Firstly, multi-manifold local neighborhood graph and the largest global variance are constructed. Nextly, a low-  dimensional manifold embedded in high-dimensional space is calculated through the multi-objective optimization. This mapping can represent not only the global structure but also the local structure. MULDP can get the low-dimensional manifolds embedded in a high-dimensional space and maintain the local and global structural information effectively. The experimental results on the ORL, Yale and AR face databases demonstrate the superiority of the proposed algorithm.

Key words: face recognition, feature extraction, multi-manifold, unsupervised linear differential projection (ULDP)