摘要
针对目标识别中多类分类的难点问题,提出了一种C-DSECOC多目标识别方法。该方法采用二符号纠错输出编码(binary Error-Correcting Output Codes)作为分解框架,采用DS证据理论作为解码策略,并结合ECOC结构特点对传统的DS证据理论进行重新构造。在确定DS基本概率赋函数值时引入损失函数,使BPA的获取除与二分器的输出有关外,还由其对不同类别样本的正确分类能力决定,从而提高ECOC集成的分类性能和泛化性能。实验中分别对UCI数据集和3种一维距离像数据集进行测试。结果表明,提出的C-DSECOC方法能有效地提高多类目标识别的正确率。
It is an open issue how to recognize the multi-class target in pattern recognition.This paper proposed a novel approach,called C-DSECOC.The binary ECOC is used as the decomposing frame to reduce multi-class to binary,and the DS evidence theory as decoding strategy.The confidence of base classifiers is estimated by loss function and fused into BPA,when considering the effect of difference in classification performance on target recognition,thus promoting the classification accuracy and generalization ability.Meanwhile,the mass function is reconstructed based on the structure features of ECOC.The experimental results based on UCI datasets and three kinds of HRRPS datasets with support vector machine（SVM） as the binary classifiers indicate that our approach can provide a better performance of classification.
出处
《计算机科学》
CSCD
北大核心
2012年第12期245-248,共4页
Computer Science
基金
国家自然科学基金(60975026)资助
