Fusion of Gabor Feature Based Classifiers for Face Verification

We present a fusion of Gabor feature based support vector machine (SVM) classifiers for face verification. 40 wavelets are used in parallel to extract features for face representation. These 40 feature extracted vectors are first projected onto the corresponding Principal Component Analysis (PCA) subspaces, and then fed into 40 SVMs for classification and fusion. No downsample is used. A publicly available FRAV2D face database with 4 different kinds of tests, each with 4 images per person, has been used to test our algorithm, considering frontal views, images with gestures, occlusions and changes of illumination. Compared to three baseline methods developed in literature, i.e. PCA, feature-based Gabor PCA and downsampled Gabor PCA, the proposed algorithm achieved the best results in the neutral expression and occlusion experiments. Compared to a downsampled Gabor PCA method, our algorithm also obtained similar error rates with a lower feature dimension.