Abstract: The stability and discrimination of low-dimensional feature extracting?are the key problems in the study of the pattern recognition. In terms of the practical application, this paper puts forward the discriminate criterion based on factors analysis and the in-depth study of Fisher discriminate criterion. It aims to solve the problem of the suboptimal definition for Fisher criterion with between-class scatter matrix and within-class scatter matrix. The experimental comparison between the synthetic data and real face data sets demonstrates that the presented algorithm possesses better robustness to deal with the data centralized edge and facial variations of expressions or poses than Fisher discriminate criterion.

Key words: feature extraction, linear dimension reduction, Fisher criterion, factors analysis, face recognition

摘要： 提取稳定且具有判别性的低维特征是模式识别研究中的关键问题。在深入研究Fisher判别准则的基础上，从因子分析的实际角度考虑，提出基于因子分析的判别准则，解决Fisher判别准则类内和类间散布矩阵非最优定义问题。通过在合成数据集和真实人脸数据集上进行实验比较表明，该方法在解决数据集中的边缘类和人脸的表情、姿态变化等问题上比Fisher判别准则更优。

关键词: 特征提取, 线性降维, Fisher判别准则, 因子分析, 人脸识别