摘要： 在给定概率分布条件下对贝叶斯分类器进行改进，提出一种基于数据库的小本征值阈值重置的贝叶斯分类器。用一个阈值替代类协方差矩阵小于阈值的本征值，使给定数据库的分类错误率最小，是一种优于零子空间法的分类方法。通过在MNIST 6×104个手写体数字数据库的测试，识别率大于96%。对小字集手写体汉字进行的实验表明，识别率大于99%。

关键词: Bayes分类器, Bayes距离, MNIST手写体字库

Abstract: This paper propose a novel small eigenvalues resetting Bayesian classifier based on database. This method is an improved Bayes classifier under normal distribution by using a threshold value to substitute the eigenvalues of each class covariance which is smaller than the threshold value. The threshold substitution is selected to minimize the error rate for a given database. The novel scheme performs better than null sub-space classifiers. Experiments on 60 000 digit MNIST handwriting database prove the effectiveness of such method and the recognition ratio is over 96%, and experiments on 3 000 Chinese characters handwriting indicate the recognition ratio over 99%.

Key words: Bayes classifier, Bayesian distance, MNIST handwriting database

中图分类号: 

• TP311