Possibilistic modeling of iris system for high recognition reliability
Résumé
The human iris is attracting a lot of attention leading to an increased demand for a robust and reliable iris-based recognition system. However, typical systems have random performances which can be efficient for some contexts and not for others. Generally, in real-world applications, iris data imperfections affect the performance of the recognition algorithms. Possibilistic modeling is a powerful tool to handle data imperfections or redundancy without being affected by data variability. Therefore, in this paper, we propose a new method for biometric recognition of human irises based on possibilistic modeling approach. The proposed approach is based on possibility theory concepts for modeling iris features in a possibilistic space. Therefore, a set of iris features, extracted from image samples, is transformed into possibility distributions for possibilistic iris representation. Features in the possibilistic space allow better recognition rates over conventional features. Validation of the proposed method was done on four challenging CASIA iris image databases namely Thousand, Interval, Twins and Synthetic. For illustration, we consider a typical iris system, from the literature. Such a system leads to an Accuracy Recognition Rate (ARR) of 72.27% on the CASIA Thousand iris database, an ARR of 78.06% on the CASIA Interval database, an ARR of 71.88% on the CASIA twins and an ARR of 77.35% on the CASIA synthetic database. Possibilistic modeling of features leads to an improvement of the ARR which is respectively of 99.92%, 99.94%, 99.91% and 99.83% for the considered databases. Possibilistic modeling leads as well to a significant reduction of the Error Equal Error Rate (EER). Indeed the typical system leads to an EER of 27.72%, 21.93%, 28.11% and 22.64%, respectively on the CASIA Thousands, Interval, Twins and synthetic databases. whereas our proposed system leads to EER of 0.04%, 0.05%, 0.08% and 0.16%, respectively for the above mentioned databases. Besides, the proposed method improves the system performance with respect to False Acceptance Rate (FAR), False rejection Rate (FRR), Area Under the receiver operating characteristics Curve (AUC).