CABALA-Collaborative architectures based on biometric adaptable layers and activities

  • Authors:

  • Maria De Marsico;Michele Nappi;Daniel Riccio

  • Affiliations:

  • Department of Computer Science, Sapienza University of Rome, Rome, Italy;Department of Civil Engineering, University of Salerno, Salerno, Italy;Department of Civil Engineering, University of Salerno, Salerno, Italy

  • Venue:
  • Pattern Recognition
  • Year:
  • 2012

Quantified Score

Hi-index 0.01

Visualization

Abstract

The lack of communication and of dynamic adaptation to working settings often hinder stable performances of subsystems of present multibiometric architectures. The calibration phase often uses a specific training set, so that (sub)systems are tuned with respect to well determined conditions. In this work we investigate the modular construction of systems according to CABALA (Collaborative Architectures based on Biometric Adaptable Layers and Activities) approach. Different levels of flexibility and collaboration are supported. The computation of system reliability (SRR), for each single response of each single subsystem, allows to address temporary decrease of accuracy due to adverse conditions (light, dirty sensors, etc.), by possibly refusing a poorly reliable response or by asking for a new recognition operation. Subsystems can collaborate at a twofold level, both in returning a jointly determined answer, and in co-evolving to tune to changing conditions. At the first level, single-biometric subsystems implement the N-Cross Testing Protocol: they work in parallel, but exchange information to reach the final response. At an higher level of interdependency, parameters of each subsystem can be dynamically optimized according to the behavior of their companions. To this aim, an additional Supervisor Module analyzes the single results and, in our present implementation, modifies the degree of reliability required from each subsystem to accept its future responses. The paper explores different combinations of these novel strategies. We demonstrate that as component collaboration increases, the same happens to both the overall system accuracy and to the ability to identify unstable subsystems.