Large Scale Face Recognition with Kernel Correlation Feature Analysis with Support Vector Machines

Recently, Direct Linear Discriminant Analysis (LDA) and GramSchmidt LDA methods have been proposed for face recognition. By utilizing the smallest eigenvalues in the within-class scatter matrix they exhibit better performance compared to Eigenfaces and Fisherfaces. However, these linear subspace methods may not discriminate faces well due to large nonlinear distortions in the face images. Redundant class dependence feature analysis (CFA) method exhibits superior performance compared to other methods by representing nonlinear features well. We show that with a proper choice of nonlinear features in the CFA, the performance is significantly improved. Evaluation is performed with PCA, KPCA, KDA, and KCFA using different distance measures on a large scale database from the Face Recognition Grand Challenge (FRGC). By incorporating the SVM for a new distance measure, the performance gain is significant regardless of which algorithm is used for feature extraction, with our proposed KCFA+SVM performing the best at 85% at 0.1% FAR where the baseline PCA gives only 12% at 0.1% FAR.