摘要： 对多类问题中的“基于平均马氏距离特征提取”、“基于马氏距离特征提取”、和“基于最小错误概率上界特征提取”方法进行了分析和比较。证实了“基于最小错误概率上界特征提取”是目前可以用迭代算法完成的最好的特征提取, 而“基于马氏距离特征提取”在错误概率上界上升很小的代价下,简化了计算量,也是一种可行的方法。文中使用实际科研项目“小字集手写体特征提取”中的数据进行实验,当特征数从4下降到2时,基于平均马氏距离特征提取的错误概率上界要比其它两种方法高10%左右。

关键词: 特征提取, 错误概率上界, 马氏距离, 巴氏距离

Abstract: Several feather selection methods based on averaged Maharanobis distance, Maharanobis distance and Bhattacharyya distance is presented and compared in this paper, and the conclusion is confirmed with the experiment. The theoretical analysis and experimental results show that the result of the first one has bug error with optimal results, and the second one is superior to the former, and the third one is best, which can get the optimal result based on least upper bound of error probability. It can be concluded that both theoretical normally data set and practical data set accord with our conclusion.

中图分类号: 

• TP391.4