一种可调控制参数的LDA人脸识别算法  

A LDA-based Face Recognizing Algorithm of Regularization Parameters

作　　者：曾岳[1,2] 冯大政[2] 

机构地区：[1]江西财经职业学院信息工程系,江西九江332000 [2]西安电子科技大学雷达信号国家重点研究所,陕西西安710071

出　　处：《计算机与现代化》2010年第12期61-64,共4页Computer and Modernization

基　　金：国家自然科学基金资助项目(60372049);江西省科技计划基金资助项目(GJJ09412)

摘　　要：线性判别分析(LDA)方法进行高维的人脸识别时,经常会遇到小样本问题(SSS)和边缘类重叠问题。本文提出一种新的LDA方法,重新定义类内离散度矩阵,利用参数来权衡其特征值估计的偏差和方差,以解决小样本问题;对类间离散矩阵加权,让边缘类均匀分布,防止边缘类的重叠,以提高识别率。大量的实验已经证明该方法能根据小样本问题的严重度调控参数以达到最高识别率,比传统的方法更优。It is well-known that the application of Line Discriminant Analysis（LDA） to high-dimensional face recognition often suffers from the so-called the problem of Small Sample Size（SSS） and close to class overlap.A new LDA method is proposed in this paper.The SSS problem is resolved by defining within-class scatter matrices for balancing bias and variance estimate of eigenvalues;weighting between-class scatter matrices for preventing edge class from being overlapped;introducing a regularized Fisher＇s criterion for increasing the stability of the null space B of within-class scatter matrices to well project the samples into the optimal space.Extensive experimental results show the new LDA algorithm can solve the above problems and outperforms traditional methods by controlling the parameters according to the degree of the SSS problem.

关 键 词：线性判别分析 小样本 偏差 方差 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]