关键词: 判别近邻嵌入（DNE）, 邻接图, 局部结构, 人脸识别

Abstract: Discriminant neighborhood embedding (DNE) is one of methods for dimensionality reduction. By constructing an adjacency graph to keep the local structure of original data in the subspace, DNE can effectively find an optimal discriminant direction. However, there are two shortcomings. On the one hand, it cannot identify the detail location of neighbors, and thus cannot keep the neighborhood structure well. On the other hand, the adjacency relationship would be imbalanced, which may not achieve the goal of minimizing the inter-class scatter and maximizing the intra-class scatter. It is not useful for classification. In order to overcome the shortcomings of DNE, this paper proposes a novel supervised subspace learning method, called locality-balanced DNE (LBDNE). In LBDNE, the homogenous and heterogeneous adjacency graphs are constructed. By adjusting the value of control parameter, the intra-class and inter-class relations can be balanced effectively. This paper compares LBDNE with the other state-of-art methods for dimensionality reduction techniques on artificial dataset and real face datasets. The experimental results show the feasibility and effectiveness of LBDNE.

Key words: discriminant neighborhood embedding (DNE), adjacency graph, local structure, face recognition