基于证据理论的多分类器融合方法研究

证据理论是建立在独立性假设基础上，理论和实际应用都需要突破这一限制。最近提出的一种相关证据模型认为，两个相关证据由一个相关源证据分别与两个独立源证据通过正交和合成得到，相关证据的合成可以归结为这三源证据的正交和，为此首先要由相关证据和相关源证据辩识独立源证据，这是证据理论中的反问题，其解是否有意义取决于相关源证据是否合适。该文给出一个充分条件，如果相关源证据满足此条件，反问题有唯一有意义的解，在此指导下，研究了字符识别中的多分类器融合问题，实验结果表明，识别性能优于传统证据理论方法。

A Study on Combining Multiple Classifiers Based on Evidence Theory

Traditional evidence theory is constructed on the assumption of independence, but both theory and practice require processing of dependent evidences. In a model of dependent evidences recently proposed by Sun and Yang, two dependent evidences are thought of as resulted from orthogonal sum of dependent original evidence and two independent original evidences, respectively. Combining the two dependent evidences can be reduced to an orthogonal sum of the three original independent evidences. For this, the independent original evidences must be identified from two dependent evidences and dependent original evidence. This is an inverse problem in evidence theory. A sufficient condition for the existence and uniqueness of the solution is given in this paper. Then a method for combining dependent multiple classifiers is proposed. Handwriting recognition experimental results indicate that the dependent evidence model is more appropriate for fusing multiple classifiers than traditional evidence theory, because it can efficiently utilize dependence between multiple classifiers.