Confidence evaluation for combining diverse classifiers

For combining classifiers at measurement level, thediverse outputs of classifiers should be transformed touniform measures that represent the confidence ofdecision, hopefully, the class probability or likelihood.This paper presents our experimental results of classifiercombination using confidence evaluation. We test threetypes of confidences: log-likelihood, exponential andsigmoid. For re-scaling the classifier outputs, we usethree scaling functions based on global normalizationand Gaussian density estimation. Experimental results inhandwritten digit recognition show that via confidenceevaluation, superior classification performance can beobtained using simple combination rules.