Abstract
Person authentication is the process of confirming or determining a person's identity. Its purpose is to ensure that a system can only be accessed by authorised users. The Biometric method uses a person's physical or behavioural characteristics. The use of biometric characteristics is increasingly more popular as it makes unauthorised access more difficult. The use of a single biometric characteristic has some limitations: intra-class variation - differences in the captured biometric characteristics from the same user; lack of distinctiveness - similarities in biometric characteristics from different users; and nonuniversality - not all users being able to provide a particular biometric characteristic. These limitations can be overcome through the use of two or more biometric characteristics. Systems using multiple biometrics give use to the problem of fusion addressed in this thesis. In this thesis two novel methods for quality based fusion are presented. (1) Quality information is included in fusion as a feature to the input of a fusion classifier. This is achieved by weighting similarity measures with the quality measures before fusing the experts. We investigate and compare different ways of including the quality information and present A priori and A posteriori results when combining six face experts and one speech expert. We also present results for all possible combination of experts using a box plot. While current quality dependent fusion algorithms are restricted to the particular fusion classifier or algorithm reported in the literature, our proposed method offers the flexibility of being used with several fusion classifiers. (2) Quality information is used to group data, allowing different parameters/fusion classifiers to be used for each group. A priori and A posteriori results were presented for all possible combination of six face and one speech experts. We also investigated the affect on system performance when the data was split into different numbers of groups. This method utilises the favourable attributes fixed fusion rules. Both quality based fusion methods deliver significant gains in accuracy over combining expert outputs (scores) without quality measures and over other fusion methods using quality information. This thesis also investigates ways of dealing with incomplete samples for fusion. We introduce variants of the popular k-NN imputation method, where we initial predict the class of the sample to ensure that the estimated data is computed only from the predicted class. We also introduced a client specific variant, that allows missing data to only be estimated from the claimed identity of the sample. Both these variants delivered improvement in system accuracy.