Multimodal biometrics refers to automatic authentication methods that depend on multiple modalities of measurable physical characteristics. It alleviates most of the restrictions of single biometrics. To combine the multimodal biometrics scores, three different categories of fusion approaches including rule based, classification based and density based approaches are available. When choosing an approach, one has to consider not only the fusion performance, but also system requirements and other circumstances. In the context of verification, classification errors arise from samples in the overlapping region (or non- confidence region) between genuine users and impostors. In score space, a further separation of the samples outside the non-confidence region does not result in further verification improvements. Therefore, information contained in the non-confidence region might be useful for improving the fusion process. Up to this point, no attempts are reported in the literature that tries to enhance the fusion process using this additional information. In this work, the use of this information is explored in rule based and density based approaches mentioned above. The first approach proposes to use the non-confidence region width as a weighting parameter for the Weighted Sum fusion rule. By doing so, the non-confidence region of the multimodal biometrics score space can be minimised. This effectively leads to a better generalisation performance than commonly used Weighted Sum rules. Furthermore, it achieves fusion performances comparable to the more complicated training based approaches. These performances are not only achieved in a wide range of bimodal biometrics experiments, but also in higher dimensional multibiometrics fusion. This method also eliminates the need for score normalization, which is required by other rule based fusion methods. The second approach proposes a new Gaussian Mixture Model based likelihood ratio fusion method. This approach suggests the application of this density based fusion to the non-confidence region only and directly reject or accept the samples in the confidence region. By applying Gaussian Mixture Model to the non-confidence ii region, a smaller and more informative region, the impact of an inaccurately chosen component number on the fusion performance can be reduced. Without tuning or using any component searching algorithm, this proposed approach achieves comparable performance to the one using specific component number searching algorithm. This successful demonstration means less resource is required whilst comparable performance can be achieved and processing time is also significantly reduced.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:629227 |
| Date | January 2011 |
| Creators | Chaw Poh, C. |
| Publisher | Nottingham Trent University |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://irep.ntu.ac.uk/id/eprint/361/ |
Page generated in 0.0018 seconds