Score-level biometric information fusion with generalized power mean

Abstract

To overcome the fundamental shortcomings of single-trait biometric systems, multimodal solutions have gained considerable interest. In this work, a score-level fusion scheme for biometric authentication is introduced, where information from multiple modalities is combined using conventional mean operators such as arithmetic, harmonic, geometric, and quadratic means, with particular attention given to the power mean formulation. The proposed framework increases system robustness while preserving low computational complexity and requiring no training phase. Performance is assessed on three well-known public datasets: National Institute of Standards and Technology (NIST)-fingerprint, NIST-face, and XM2VTS, using standard score normalization methods and commonly employed evaluation metrics. The experimental analysis shows that the quadratic mean attains a genuine acceptance rate (GAR) of 91.50% on the NIST-fingerprint dataset, while the power mean with α = 5 achieves 82.40% on NIST-face. Furthermore, the half total error rate (HTER) on XM2VTS is reduced to 0.059. In comparison with learning-based fusion techniques, the proposed approach provides a more straightforward, computationally efficient, and dependable alternative for real-world biometric applications.