Multiple classifier systems incorporating uncertainty

Abstract

Multiple classifier systems (MCS) unite the answers of separately-trained powerful base-classifiers to obtain the right classification for the sample at hand. In practical applications, a sample is not associated with exactly one class, but belongs fuzzily to multiple ones. How uncertain class information can be incorporated into multiple classifier systems is detailed in this work. On the theoretical side, it is described how existing approaches to modelling uncertainty like Bayesian probability, Dempster-Shafer theory, fuzzy logic or fuzzy sets, and also unfamiliar ones like the distribution of opinions, are able to support and deal with the core notions of uncertainty in classification: vagueness, imprecision and certainty. In the larger practical part, the use of uncertainty is detailed for every stage of the MCS. The most suitable classifiers are identified, and some well-known schemes extended to deal and answer with uncertain class information. Notably every aspect of the newly proposed and award-winning (KES 2007) fuzzy-input fuzzy-output support vector machines is explained. How the certainty of a classifier answer can be quantified is explored as well as which fusion scheme to use to come to a final classification. The steps for applying the techniques above to real-world problems are shown exemplarily for two applications, the recognition of emotions in facial expression videos, and land cover mapping from satellite images (a winner of the IEEE DFTC Contest 2008). Homepage des Autors: http://www.christianthiel.com/dissertation.html