A Study On The Fuzzy Measure For Multi-classifier Fusion

Fuzzy Integral is an aggregation tool in multi-classifier fusion, which can improve the accuracy of classification and the robustness of systems. In multi-classifier fusion based on fuzzy integrals, fuzzy measures have much influence on the performance of fusion systems. If the fuzzy measures are well defined, the accuracy of classification can be improved distinctly. However, if the fuzzy measures are badly defined, it is possible that most classifier's accuracy of classification is higher than the fusion system's accuracy. In this thesis we analyze what influence fuzzy measures have on the classification of the fusion system with given classifiers. It is discovered that multi-classifier fusion systems based on fuzzy integrals have some ability to correct the misclassification of classifiers. It is possible that the classification of fusion system is correct even if all classifiers misclassified one sample. And we propose the fuzzy measure function such that fuzzy measures can variety according to input sample to reflect the importance of each classifier and the interaction among them. And, a method to find a fuzzy measure is given. Additionally, we have found that the monotony of fuzzy measures can be ignored when using Sugeno integral as fusion operator.