摘要： 提出了一种新的基于LDA 的人脸识别算法。该方法重新定义了样本的类间散布矩阵，在原始的定义基础上增加了一种径向基函数（RBF）调节类间距离，使得在选择投影方向时能更好地分开各类样本；同时该方法在类间散布矩阵与类内散布矩阵的特征分解的基础上，通过变换求出符合Fisher 准则的最优投影方向，可以证明这样得到的投影方向同时具有正交性与统计不相关性。通过ORL 人脸数据库的数值实验，表明了该算法比传统算法有更好的性能。

关键词: 线性判别分析；样本类间离散度；样本类内离散度；特征提取；人脸识别

Abstract: This paper introduces a new approach of improved-LDA to overcome the drawbacks existing in the traditional PCA and LDA methods. It redefines the between-class scatter matrix by adding a radical basis function(RBF). Therefore, it can work better than the traditional methods. At the same time, a optimal set of uncorrelated discriminant vectors have been founded on the basis of the eigen decomposition of between-class scatter matrix and within-class scatter matrix. The numerical experiments on facial database of ORL show this method achieves better performance of face recognition than traditional methods.

Key words: Linear discriminant analysis(LDA); Between-class scatter; Within-class scatter; Feature extraction; Face recognition