基于矩阵分解的DLDA特征抽取方法分析

ANALYSING FEATURE EXTRACTION METHOD OF DLDA BASED ON MATRIX DECOMPOSITION

提出两种基于矩阵分解的DLDA特征抽取算法。通过引入QR分解和谱分解(SF)两种矩阵分析方法,在DLDA鉴别准则下,对散布矩阵实现降维,从而得到描述人脸图像样本更有效和稳定的分类信息。该方法通过对两种矩阵分解过程的分析,证明在传统Fisher鉴别分析方法中,矩阵分解同样可以模拟PCA过程对样本进行降维,从而克服了小样本问题。在ORL人脸数据库上的实验结果验证了该算法的有效性。

In this paper,two new DLDA feature extraction algorithms based on matrix decomposition are proposed.The algorithms import two matrix analysing methods of the QR decomposition and the spectra factorization（SF） respectively,under the direction of DLDA discriminant criterion to reduce the dimension of scatter matrices,so as to obtain more efficient and stable classification information of the sample which describes a whole set of human face images.These algorithms prove by analysing two matrices decomposition processes that in conventional method of Fisher discriminant analysis the matrix decomposition can also simulate the process of PCA to reduce sample＇s dimension,therefore the limitation of small sample size problem has been overcome.Simulation results on ORL face image databases verified the validity of these algorithms.