Many biometrics offer the advantage to require quite short enrollment/testing times, yet with a recognition rate that, though satisfying in controlled settings, quickly deteriorate in the presence of expression or lighting variations or of occlusions. On the contrary a recognition system based on fingerprints/iris/DNA provides a high degree of reliability, the counterbalance of which is an execution time growing with the number of users in the database; as a result, such system is particularly suited for off-line systems, while the computational cost is hardly affordable in an on-line one. In this case several biometrics are suitably combined in an on-line system providing recognition rates comparable to that of the more reliable biometric, though execution times are kept comparable to those required by fastest one.
N-Cross Testing Protocol
The system is composed of N different subsystems working in parallel and exchanging information in well defined points. Each subsystem could either exploit a different biometry or implement a different technique for the same biometry exploited by another subsystem. Moreover, in a first phase they all work in identification mode, while in a second phase they work in pseudo-verification mode. We will clarify this afterward. The Figure shows a sketch of system operation and then we provide further details on the recognition process.
As in the N-cross-testing protocol all subsystem work in parallel, execution time of the overall system equals the one of the slowest subsystem. The solution resides in substituting the testing protocol, adopting a hierarchical scheme.
In this further proposal the subsystems Face and Ear work in parallel and independently, each producing a list of subjects retrieved from its database; each list is ordered by similarity with the input subject. Such lists have the same structure: an ID identifying the database subject and a numeric value expressing similarity degree. The two lists are fused in a single one by a fusion module. The resulting list sets up the input for the Fingerprint subsystem, that performs a verification process for each subject in it, stopping as soon as a subject with a score higher than a fixed threshold e is found. The main advantage of this kind of architecture is to reduce the time required by the Fingerprint subsystem, though preserving the same recognition rate.