Abstract: In order to solve the feature extraction problem of nonlinear small sample sizes present in the traditional Fisher discriminant analysis method, a matrix transform is proposed on the basis of kernel linear subspace theory, by which a new kernel symmetrical linear subspace of within-class scatter matrix is constructed. Two kernel solution spaces derived from the within-class scatter matrix and its corresponding symmetrical subspace are respectively utilized to obtain the efficient discriminatory information of the samples. Experimental results conduct on the NUST603 and ORL face databases demonstrate the effectiveness of the proposed method.

Key words: feature extraction, Linear Discriminant Analysis(LDA), symmetrical subspace, Small Sample Size Problem(SSSP)

摘要： 为解决传统Fisher鉴别分析方法中非线性小样本的特征抽取问题，从核线性子空间角度出发，构造一种矩阵变换，得到核空间中类内散布矩阵的另一个对称核子空间，通过对2个核子空间分别求解，从而得到样本的有效鉴别信息。在NUST603和ORL人脸数据库上的实验结果验证了该算法的有效性。

关键词: 特征抽取, 线性鉴别分析, 对称子空间, 小样本问题

CLC Number: 

• TP391.41