Fusion of multiple matchers

With reference to the four Scenarios described in the Bibliography Section:

  1. Fusion of multiple matching algorithms (multi-algorithmic systems): two or more matching algorithms are applied to a unique biometric impression.
  2. Fusion of multiple impressions: the same matching algorithm is applied to two or more impressions of the same biometric (or different biometrics).
  3. Fusion of multiple sensors: the same matching algorithms is applied to two or more images provided by two or more fingerprint sensors.
  4. Fusion of fingerprints and other biometrics: the score provided by the fingerprint matcher is combined with that provided by other biometric matchers (e.g. a face matcher).

The contribution of PRA group on Scenario 1 is related to fingerprint and face matchers. In particular:

  • a first experimental comparison of multiple matchers by several score-level fusion rules has been made for fingerprint systems;
  • a novel approach based on single-layer perceptron has been tested on fingerprint verification systems;
  • the score-level fusion of PCA and LDA algorithms for face recognition and verification systems.

The third scenario has been proposed only by the PRA group people. It is well-known that the information obtained by different acquisition sources can be complementary. Therefore, combining this information can provide a more reliable matching score and thus a better classification of the subject. 
The contribution of PRA group to Scenario 4 has been the proposal of a novel approach to design the serial combination of multiple biometrics. This model, proposed for the case of two biometrics, have shown good prediction properties, and is currently under extension to more than two biometrics.
Finally, the combination of the so-called soft biometrics has been recently proposed. Soft biometrics are physiological or behavioural human characteristics which are poor discrimination properties, as the hair colour, the ethnicity, the height. However, they can help in improving the performance of biometric system. In this field, PRA group people proposed a novel model for combining soft biometric and main biometrics, based on Bayesian networks. The idea behind this work come from the observation that, due to their low discriminant power, it can be argued that a certain soft biometric can be useful only for a limited set of users. For example, Latin people rarely exhibit blond hair. The model has been applied to a case-study related to face as main biometric and hair colour and ethnicity as soft biometrics. Experimental results on two benchmark data sets have shown the effectiveness of the proposed approach.

 

People working on these topics:

  • Luca Didaci
  • Giorgio Giacinto
  • Gian Luca Marcialis
  • Daniele Muntoni
  • Fabio Roli
  • Roberto Tronci