Abstract: Under the condition of small sample size，the average of all training samples used in the traditional principal component analysis algorithm is not always the scatter center of the samples.To address the problem，a new two dimension principal component analysis method based on the sample median is proposed.This algorithm is called Median Two Dimension Principal Component Analysis（M2DPCA），in which the median of training samples is substituted for the average.To demonstrate the effectiveness of the method，extensive experiments are performed on two popular face databases，such as ORL and FERET.Experiment results indicate that the proposed method is better than traditional PCA and 2DPCA.

Key words: face recognition, Two-Dimension Principal Component Analysis（2DPCA）, sample median, feature extraction

摘要： 在小样本情况下，传统的2DPCA算法中采用的训练样本的平均值不一定就是训练样本分布的中心，为了解决这个问题，提出了一种基于样本中间值的2DPCA人脸识别算法（M2DPCA），该算法采用训练样本的中间值代替训练样本的平均值，以此重建总体散布矩阵。在ORL和FERET人脸数据库上的实验结果证明，新方法可以有效改善识别性能，优于传统的PCA和2DPCA方法。

关键词: 人脸识别, 二维主成分分析, 样本中间值, 特征提取