一种正交局部鉴别嵌入的人脸识别算法

Orthogonal local discriminant embedding for face recognition

为了解决局部鉴别嵌入(LDE)算法的高维小样本泛化能力弱和分解致密矩阵计算量较大的问题,提出了一种基于谱回归的正交局部鉴别嵌入算法(SR-OLDE),采用谱回归理论与正交化技术相结合的方法,将投影函数的求解转化为回归问题的求解.该算法首先计算训练样本的特征向量;然后通过回归方法计算投影向量,得到测试数据集,从而将n×n维的致密矩阵的特征分解转化为m×m维矩阵的特征分解,n,m分别为人脸特征矩阵维数和人脸样本数;最后对投影向量进行Gram-Schmidt正交化,得到正交的投影矩阵,从而可准确估计高维数据的内在维数,提高了样本的泛化能力.实验结果表明,该算法在降低人脸特征矩阵维数和提高人脸识别率的同时,缩短了计算时间.

The spectral regression-based orthogonal local discriminant embedding （SR-OLDE）algo-rithm is proposed to improve the generalization performance of high-dimensional small samples and the efficiency of decomposing dense matrix in the local discriminant embedding （LDE）algorithm. The projection function is transformed into the regression problem by using the spectral regression theory and orthogonalization technology.First,the eigen vector of the training samples is calculated. And then in order to obtain the test data sets,the projection vector is calculated through the regres-sion method.Thereby the eigen decomposition of n ×n dimensional dense matrix is transferred into that of m ×m dimensional matrix,where n is the dimension of eigenface matrix and m is the number of face samples.Finally,the projection vector is orthogonalized by Gram-Schmidt method to obtain the orthogonal projection matrix,which can accurately estimate the intrinsic dimension of high-di-mensional data and improve the generalization performance of the sample.The experiments show that the SR-OLDE algorithm has better performance in reducing dimensions of eigenface matrix and recognition rate than the LDE algorithm,and its computation time is decreased.