Quaternion Fisher Discriminant Analysis for Bimodal Multi-feature Fusion

Aiming at the accuracy and security of pattern recognition system, this paper proposes a quaternion based multi-modal recognition algorithm that is more accurate and safe than unimodal, and fuses more features than most existing methods. Our algorithm fuses four features that involve two linear features and two non-linear features of two kinds of modalities. We fuse features into quaternion and the process of recognition is dealt in quaternion field. The equal error rate (EER) and DET curves given by the experiment we did on Yale face database and PolyU palm print database show that the quaternion based algorithm we proposed improves the recognition rate observably.