自适应通用学习框架改进FLDA的人脸识别

针对传统的Fisher线性判别分析(FLDA)算法在处理单训练样本人脸识别时由于类内散布矩阵为零而不能进行特征提取的问题,提出了一种基于自适应通用学习框架改进FLDA的人脸识别算法。首先选取一个合适的通用训练样本集,计算其类内散布矩阵和样本平均向量;然后,利用双线性表示算法预测单训练样本的类内、类间散布矩阵,巧妙地解决了单训练样本类内散布矩阵为零的问题;最后,利用Fisher线性判别分析进行特征提取,同时借助于最近邻分类器完成人脸的识别。在Yale及FERET两大通用人脸数据库上的实验验证了所提算法的有效性及可靠性,实验结果表明,相比其他几种较为先进的单样本人脸识别算法,所提算法取得了更好的识别效果。

FLDA Improved by Adaptive Generic Learning Framework for Face Recognition

For the issue that traditional Fisher linear discriminative analysis algorithm could not extract features due to its scattering matrix within class is zero in face recognition with single training sample per person, a face recognition algorithm based on FLDA improved by adaptive generic learning framework is proposed. Firstly, a suitable generic training sample set is selected and its scattering matrix within class and mean vectors are computed. Then, scattering matrix within-class and between classes are predicted by bilinear representation algorithm, which has settled the problem of its scattering matrix within class is zero. Finally, FLDA is used to extract features and nearest neighbour classifier is used to finish face recognition. The effectiveness of proposed algorithm has been verified by experiments on the two common databases Yale and FERET. Experimental results show that proposed algorithm has better recognition efficiency than several advanced single training sample face recognition algorithms.