Multi-manifolds Discriminative Canonical Correlation Analysis for Image Set-Based Face Recognition

In this paper, multi-manifolds discriminative canonical correlation analysis (MMDCCA) is presented for solving face recognition problem with using different image sets. We adopt Linearity-Constrained Hierarchical Agglomerative Clustering algorithm for dividing all image sets into a range of local clusters. Then MMDCCA is proposed to find multiple orthogonal projection functions for maximizing the margins of manifolds with different persons. In order to obtain gains in discrimination accuracy, we enforce a constraint that each person-specific manifold is orthogonal to those of all other manifolds after linear transformation. An efficient sequential iterative learning algorithm is used for finding the discriminative features. Extensive experiments confirm the effectiveness of our model.