This paper proposes a new approach to estimating the locations of a set of landmarks for a large image ensemble, using manually annotated landmarks for only a small number of images in the ensemble.
Multi-modal biometric fusion is more accurate and reliable compared to recognition using a single biometric modality; however, most existing fusion approaches neglect the influence of the qualities of the biometric samples in information fusion. The authors’ goal in the current project is to advance the state-of-the-art in biometric fusion technology by providing a more universal and more accurate solution for personal identification and verification with predictive quality metrics. In this work, the authors developed score-level multi-modal fusion algorithms based on predictive quality metrics and employed them for the task of face and fingerprint biometric fusion. The causal relationships in the context of the fusion scenario are modeled by Bayesian Networks. The recognition/verification decision is then made through probabilistic inference. The authors’ experiments demonstrated that the proposed score-level fusion algorithms significantly improve the verification performance over the methods based on the raw match score of a single modality (face or fingerprint). Furthermore, the fusion framework with both face and fingerprint image qualities achieves the best verification performance and outperforms all other baseline fusion algorithms tested, including other straightforward quality-based fusion methods. (Publisher abstract provided)