结合可信度和DS证据理论的ECOC多类分类研究

针对多类目标识别问题，利用三符号纠错输出编码作为将多类分解为若干个二类问题的结构框架，用改进的证据理论作为融合策略，将每一个二分器的输出作为证据之一进行融合；同时对分类器的可信度进行估计，并将得到的信任度量值融入证据中，从而提高多类分类的正确率。实验中分别对UCI数据集和三种一维距离像数据集进行测试，结果表明本文提出的基于分类器可信度的多类目标识别方法能有效地提高复杂环境下多传感器目标识别的正确率。

Error-correcting Output Codes Based on Classifiers’ Confidence for Multi-class Classification

To the problem of multi-class target recognition,the ternary ECOC is introduced to reduce the multi-class to the binary as the decomposing framework.The outputs of binary classifiers are used as evidence to fuse under the decoding strategy of DS theory.The classifiers’ confidence is estimated,which is fused into the DS’s evidence to enhance the correct classification rates.The experiments results based on UCI datasets and HRRPS indicate the multi-class target recognition based on classifiers’ confidence can improve the performance of target recognition under complex circumstance especially.