Dynamic Laplacian Principle Component Analysis on Objective Space
Shaoe Xue, Shuqin Rao, Wenxin Yang, Jina Wang, Jian Ping Yin · 2008
Laplacian PCA tries to maximize the intra-class covariance across all samples while preserving the local manifold information.However, the static geometry center is incompetent to express the data's manifold structure, and easily influenced by the unique noise point. Moreover, Laplacian PCA also neglects the upgraded information in the objective space. In this paper, we propose the Dynamic Laplacian PCA method, which introduces the gradient based Laplacian center to precisely illustrate the local manifold,besides, iterative Laplacian PCA is employed to optimize the feature vectors in the objective space. The experimental results on a face database and a virtual database show the promise of our method.