Efficient Kernel Discriminate Spectral Regression for 3D face recognition
Yue Ming, Qiuqi Ruan, Xiaoli Li, Meiru Mu · 2010
In this paper, a novel framework for 3D face recognition based on depth information, is proposed. The core of our framework is Spectral Regression Kernel Discriminate Analysis (SRKDA), a method for utilizing a reproducing kernel Hubert space (RKHS) into which data points are mapped. In order to overcome facial expression variation, we first utilize curvature information projected onto the moving least-squares (MLS) surface to segment a face rigid area, which is insensitive to expression variation. Then we make use of SRKDA to extract discrimination features for a depth image obtained by use of a 3D face mesh model, thus avoiding an eigen-decomposition of a kernel matrix. This effectively merges 3D facial shape information; then a nearest neighbor classifier is used for recognition. A non-linear kernel trick solves the high dimensional small sample size problem, and enhances feature extraction from the local non-linear structures of a face. Our experiments, using the CASIA 3D face database, show our framework performs more effectively and efficiently than many commonly used methods. SRKDA decreased the complexity from cubic-time to quadratic-time resulting in a very significant reduction in computation time. In addition, recognition accuracy, based on face rigid areas, improved accuracy significantly when compared to accuracy before segmentation.