Fusing horizontal and vertical components of face images for identity verification

This paper presents an empirical investigation of two sparse random projections which correspond to extraction of vertical and horizontal features from a face image for identity verification. In order to enhance the performance of each projection, the matching scores of both directional features are fused via a total error rate minimization. The BERC face database is used for evaluating the effectiveness of the proposed method. Our empirical results show that the proposed vertical projection outperforms the commonly used PCA and a Random Projection algorithm in terms of the Equal Error Rate (EER) measure. The result of fusion shows an even better EER performance than that from each individual projection.