多分类器组合的投票表决规则

为了提高多分类器组合的效果 ,提出了一种改进的投票表决规则 .每个分类器对每个模式类别的识别能力不尽相同 ,通过对大量样本的统计 ,获得有关每个分类器识别性能的先验知识 ,将其作为投票表决的依据 .投票表决时每个类设置不同的表决阈值 ,使组合效果得以改善 .这些阈值可以通过训练获得 .在数字识别系统中的实验结果表明 ,该方法使多分类器组合的识别率和置信度明显提高

Voting Principle for Combination of Multiple Classifiers

Voting principle is usually utilized to combine the results of several individual classifiers.The recognition ability of each classifier to each class is not the same.A prior knowledge of each classifier was obtained by the statistic with a large number of samples,which acts as basis of the voting principle. Dif- ferent thresholds gained by training were employed for differentclasses,which can yield betterresults. Its application to numeral recognition shows thatthe presented voting principle makes the combination of mul- tiple classifiers achieve good performance for high recognition rate and high reliability.