基于概率线性判别分析的可扩展似然公式化人脸识别

针对概率线性判别分析(PLDA)方法在训练及似然计算过程中矩阵大小随着标志类采样数量呈平方增长的问题,提出了一种基于概率线性判别分析的可扩展似然公式化方法。首先通过简单变换变量对角化PLDA模型;然后,利用贝叶斯准则和最大期望算法估算潜在变量一阶矩、二阶矩,从而对变换后的PLDA模型进行可扩展训练;最后,通过结合Woodbury矩阵特征存储模型信息,从而将大矩阵转换成低维向量或标量。在FLW及Multi-PIE两大通用人脸数据库上的实验验证了所提方法的有效性及可靠性,实验结果表明,相比其它几种较为先进的同类人脸识别方法,所提方法不仅取得了更高的识别率、更低的半错误率,还大大地降低了训练、似然计算复杂度。

Scalable Likelihood Formulation Based on Probabilistic Linear Discriminant Analysis for Face Recognition

For the issue that size of matrix has the square growth with sampling number of mark classes in the training and likelihood calculation process of probabilistic linear discriminant analysis (PLDA), scalable formulation based on probabilistic linear discriminant analysis is proposed. Firstly, PLDA model is diagonalized by simply transforming variable. Then, Bayesian criterion and expectation maximum algorithm is used to estimate first rank, second rank matrix of potential variable so that transformed PLDA model can be extensible trained. Finally, big matrix is transferred to low dimensional vector or scalar by combining Woodbury matrix feature to saving model information. The effectiveness and reliability of proposed method has been verified by experiments on the two common face databases FLW and Multi-PIE. Experimental results show that proposed method has higher recognition accuracy, low half total error rate and clearly reduces computing complexity of training and likelihood comparing with several other advanced similar face recognition approaches.