A Study On Multi-classifiers Fusion Based On Fuzzy Integrals

In the model of multi-classifier fusion based on fuzzy integrals respect to lambda-fuzzy measures, the fusion results are heavily dependent on fuzzy densities which represent the importance of individual classifier to the fusion results. So determining fuzzy densities is very important to the fusion systems. Some existing methods only take account of accuracy in determining of fuzzy densities but take little concern about other uncertainty which can also reflects the performance of a classifier. In this thesis, three types of uncertainties of classifiers are studied: accuracy, fuzziness and ambiguity. Fuzziness and ambiguity are particularly discussed in detail. Firstly, the conditions that a function used to measure the fuzziness must satisfy are given. It presents a function satisfied these conditions. Secondly, the similarity is introduced and used to measure the ambiguity. Finally, a new method for determining fuzzy densities is proposed through incorporating accuracy, the uncertainty of fuzziness and ambiguity of the classifiers. This method is verified to be practical by experiments.