基于证据理论的多分类器集成方法研究

提出了一种基于证据理论的多分类器集成方法。该方法首先对分类器输出的抽象层信息进行不确定性表征，即对输出证据的基本概率分配进行赋值，然后，利用证据组合规则对多分类器进行集成，并针对语气组合中“焦元爆炸”的问题，推导了一种快速算法。使信息复杂度和计算复杂度大大降低低。该方法将对往设计一个性能优良的高维输入分为器的问题转化为设计多个性能较优的低维分类器，较好的解决了高维特征空间的判分问题和高可靠性分类器的设计问题，最后将此算法应用于字符识别，结果令人满意。

A method of combining multiple classifiers based on evidence theory

The paper proposes a method of combining multiple classifiers based on evidence reasoning. First, the uncertainty of each classifier's output in modeled. Then, all the classification results are integrated by the use of D-S combination rule. Furthermore, a fast algorithm is derived to deal with the complex computation in evidence combination. In this method, we use multiple classifiers with moderate performance and low-dimension instead of a classifier with high-dimension input space and/or high performance. Thus, it gives a good resolution of designing a classifier with high recongnition rate for high-dimension input prob- lem. Finally, the method is applied to handwritten digit recognition system where the results demnostrate its effectiveness.