Laplacian Transductive Optimal Design for Face Classification
Rong Rong Zhang, Yue Hua Li, Jian Qiao Wang · Applied Mechanics and Materials · 2014
The problem of face classification is essentially a nonlinear subspace classification problem. The features of different face samples lie on different nonlinear subspaces. If the most representative samples of each subspace are selected as the training set, we can enhance the reliability of the classification, as well as reduce the computation. In this paper, a manifold based active learning algorithm, called Laplacian Transductive Optimal Design (LTOD), is presented to select the most representative samples. LTOD reconstructs each sample by the graph Laplacian matrix to make sure that the reconstructed samples and the original samples share the same manifold structure. Then, the samples which can be used to best reconstruct the whole sample set are selected as the training set. The experimental results on Yale face database have demonstrated the effectiveness of our method.