Regularized kernel discriminant analysis with a robust kernel for face recognition and verification.

S Zafeiriou,T Stathaki,G Tzimiropoulos,M Petrou

doi:10.1109/tnnls.2011.2182058

Abstract

We propose a robust approach to discriminant kernel-based feature extraction for face recognition and verification. We show, for the first time, how to perform the eigen analysis of the within-class scatter matrix directly in the feature space. This eigen analysis provides the eigenspectrum of its range space and the corresponding eigenvectors as well as the eigenvectors spanning its null space. Based on our analysis, we propose a kernel discriminant analysis (KDA) which combines eigenspectrum regularization with a feature-level scheme (ER-KDA). Finally, we combine the proposed ER-KDA with a nonlinear robust kernel particularly suitable for face recognition/verification applications which require robustness against outliers caused by occlusions and illumination changes. We applied the proposed framework to several popular databases (Yale, AR, XM2VTS) and achieved state-of-the-art performance for most of our experiments.

Highlights

Learning algorithms for face recognition/verification have generated a wealth of scientific research within the computer vision community for more than two decades [1]
The proposed nonlinear kernel is suitable for face recognition/verification applications which require robustness against outliers caused by occlusions and illumination changes
We propose a direct complete kernel discriminant analysis (KDA) method based on the complete eigen analysis of the within-class scatter matrix SHw

Summary

Introduction

Learning algorithms for face recognition/verification have generated a wealth of scientific research within the computer vision community for more than two decades [1]. This research has primarily revolved around providing efficient solutions to the following problem: given samples of a highdimensional space (facial images of our training set), estimate a low-dimensional face space that preserves the intrinsic structure of the input data [2]–[15]. Classification is typically performed by projecting the probe face onto this low-dimensional space and applying off-the-shelf classifiers. Efficient estimation of the face space is usually hindered by two factors. In a practical setting, very few samples (usually 1–5 for each class) are available for training.

Methods

Results

Conclusion