基于核的Fisher非线性最佳鉴别分析在人脸识别中的应用

Face Recognition Based on Kernel Fisher Nonlinear Optimal Discriminant Analysis

抽取最佳鉴别特征是人脸识别中的重要一步。对小样本的高维人脸图像样本,由于各种抽取非线性鉴别特征的方法均存在各自的问题,为此提出了一种求解核的Fisher非线性最佳鉴别特征的新方法,该方法首先在特征空间用类间散度阵和类内散度阵作为Fisher准则,来得到最佳非线性鉴别特征,然后针对此方法存在的病态问题,进一步在类内散度阵的零空间中求解最佳非线性鉴别矢量。基于ORL人脸数据库的实验表明,该新方法抽取的非线性最佳鉴别特征明显优于Fisher线性鉴别分析(FLDA)的线性特征和广义鉴别分析(GDA)的非线性特征。

Extracting the most discriminatory features is important in face recognition tasks. In the case of a small number of face samples, as the existed methods for extracting nonlinear most discriminatory face features encounter various problems. So a new method for extracting fisher nonlinear most discriminatory features is proposed in this paper. The fisher criterion is formulated using between-class scatter matrix and within-class scatter matrix based on kernel method. Thus nonlinear most discriminatory features are obtained. However, this method causes ill-problem. To solve this problem, we search optimal discriminant vectors in null space of within-class scatter matrix. Repeated experimental results on ORL database indicate that the proposed method significantly outperforms the Fisher linear discriminant analysis（FLDA） and generalized discriminant analysis（GDA）.