A Regularized Margin Fisher Analysis Method For Face Recognition

Xiaoyu Xue,Xiaohu Ma,Yuxin Gu,Xiao Sun,Zhiwen Ni

DOI: https://doi.org/10.1007/978-3-319-70136-3_45

2017-01-01

Abstract:Margin Fisher Analysis is a typical graph-based dimensionality reduction technique and has been successfully applied to face recognition. However, it always suffers from the over-fitting, noise, and singular matrix problems. Common preprocessing methods such as PCA lose certain discriminant information in data, which leads the poor classification rate. We propose a novel method called Regularized Margin Fisher Analysis, which decomposes the inter-class similarity matrix into three subspace: principal space, noise space and null space. Then, we regularize the three subspaces in different ways to deal with the noise and over-fitting problems. Moreover, we use twice standard eigendecompositions instead of single generalized eigendecomposition which avoids the singular matrix problem. The experiments on Extended YaleB, CMU PIE and FERET face databases demonstrates that the proposed method is effective and can improve the classification ability.