One of the most important conclusions after observing the results in Table 2 is that in our particular application fusion always improves the system performances beyond those of even the best single expert. These results have to be seen in their correct perspective, taking into account the very limited database. In any case, these performances are much better than those of verification systems using only one of the presented modalities. With respect to the use of statistical considerations in the design of multi-modal identity verification systems, the following two comments can be made: 1. Statistical analysis a priori is important to increase the supplementary information input coming from additional verification experts. This can be done by analyzing the correlation of the different experts. An interesting offspring of this correlation analysis is the fact that sometimes - in a personalized approach -it might be interesting to analyze the extreme values quite carefully, before discarding them. 2. Statistical analysis a posteriori is important to choose a fusion method which is statistically significant the best (or which belongs to the best group). The question of which fusion method should be chosen, is indeed a difficult one to answer. A lot depends on the application. To be able to choose a number of potentially powerful fusion paradigms, it is important to have a (large) representative database of your application, which can be used for training purposes. It also helps if one is able to visualize the different populations (clients versus impostors), because in that specific case the choice of the fusion methods could be guided by the shape of the separation frontier between the two populations one wants to obtain.
展开▼
