基于互信息梯度优化计算的信息判别特征提取  

Information Discriminant Feature Extraction Based on Mutual Information Gradient Optimal Computation

作　　者：谢文彪[1,2] 樊绍胜[1] 费洪晓[2] 樊晓平[2] 

机构地区：[1]长沙理工大学电气与信息工程学院,长沙410004 [2]中南大学信息科学与工程学院,长沙410083

出　　处：《电子与信息学报》2009年第12期2975-2979,共5页Journal of Electronics & Information Technology

基　　金：国家863计划项目(2008AA04Z214)资助课题

摘　　要：该文将互信息梯度优化引入特征提取矩阵求解,提出一种信息判别分析的特征提取方法。首先,分析了现有线性判别方法的特点和局限,建立了类条件分布参数模型下互信息最大化的信息判别模型。其次,证明了互信息判别的线性变换不变性和贝叶斯一致优化,构造了一个互信息梯度优化计算的特征提取算法。最后通过实际数据上试验验证了该方法的有效性。A linear feature extraction method is present with information discriminant analysis, it is based on a feasible computationally feature extraction matrix used mutual information gradient. Firstly, this paper analyzes the limitation for current linear discriminant, and constructs a information discriminant analysis model which facilitates the maximization of the mutual information under the parametric class-conditional PDF. Then, it is proved that the mutual information is linear transformation invariance and optimal in the sense of Bayes, and the algorithm is present for computing feature extraction matrix with mutual information gradient. Finally, the good performance of the method is proved on real-world data set.

关 键 词：特征提取 信息判别分析 线性变换 互信息梯度 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP274.3[自动化与计算机技术—计算机科学与技术]