L1-Graph Based Neighborhood Preserving Embedding with an Application to Face Recognition

Neighborhood Preserving Embedding (NPE) is a famous graph-oriented dimension deduction algorithm, which has got lots of successful applications in computer vision field. Just as all the graph-oriented algorithms, the effectiveness of the NPE greatly relies on whether a suitable graph can be constructed. While the traditional graph construction method has some intrinsic weaknesses, especially when the face appearance exist large changes resulting from illumination, expression and pose. L1-graph method is a parameter-free unsupervised graph construction method and is more robust to noise and outliers than the traditional graph construction method. Combining the merits of L1-graph and NPE, this paper proposes a L1-graph based Neighborhood Preserving Embedding (L1GNPE) method which can promote the robustness and discriminative ability of NPE. Extensive experiments conducted on several popular face databases show that the L1GNPE method outperforms the traditional NPE method under unsupervised conditions.