针对人脸识别问题,提出了最小距离鉴别投影算法,其与经典的线性鉴别分析不同,它是一种流形学习降维算法。该算法首先定义样本的类内相似度与类间相似度:前者能够度量样本与类内中心的距离关系,后者不仅能够反映样本与类间中心的距离关系而且能够反映样本类间距与类内距的大小关系;然后将高维数据映射到低维特征空间,使得样本到类内中心距离最小同时到类间中心距离最大。最后,在ORL、FERET及AR人脸库上的实验结果表明所提算法识别性能要优于其他算法。

A minimum-distance discriminant projection (MDP)algorithm is proposed to address face recognition problem. Different from the classical linear discriminant analysis (LDA), the MDP is a manifold learning based dimensionality reduction algorithm. MDP first defines the intra-class similarity, weight, and the inter-class weight of each sample. The former one can measure the distance between each data point and the intra-class center, while the latter one does not only characterize the distance between the data point and the inter-class center but also can reflect the relation between the between-class distance and the within-class distance. Then, the high-dimensional data is mapped into a low-dimension space such that the points to within-class center distances are minimized while the points to between-class center distances are maximized simultaneously. At last, experiments on the ORL, FERET, and AR face databases show that the proposed algorithm can outperform other algorithms.