基于局部约束字典学习的非线性降维人脸识别

针对现有的非线性降维(NLDR)算法复杂度高而不能很好地处理现实世界大规模数据集的问题,提出了基于局部约束字典学习的非线性降维(LCDL-NLDR)方法。首先通过一些潜在的标志点重构极小的内在流形,并将训练数据和未知数据自然地嵌入到内部流形中；然后利用局部约束字典学习(LCDL)算法在非线性流形中学习由标志点组成的紧密原子集；最后利用最近邻分类器完成人脸的识别。在扩展的YaleB及CMU PIE两大人脸数据库上的实验验证了所提方法的有效性及鲁棒性,通过与几种先进的字典学习算法比较表明,所提算法提高了嵌入质量,取得了更高的识别率,同时也大大地降低了NLDR算法的复杂度。

Face Recognition with Nonlinear Dimensionality Reduction Based on Locality Constrained Dictionary Learning

For the problem that existed nonlinear dimensionality reduction (NLDR) algorithms can not be applied in large-scale data sets in the real world due to its high complexity, nonlinear dimensionality reduction based on locality constrained dictionary learning (LCDL-NLDR) is proposed. Firstly, training and unknown dataset is naturally embedded to tiny inner manifold reconstructed by some potential landmarks. Then, closely atomic sets consist of mark points will be learned effectively by locality constrained dictionary learning (LCDL) algorithm in nonlinear manifold. Finally, nearest neighbor classifier is applied to finish face recognition. The effectiveness and robustness of proposed method has been verified by experiments on extended YaleB and CMU PIE face databases. Comparison with several latest dictionary learning algorithms shows that proposed algorithm has improving embedded quality, getting higher recognition accuracy, and reducing the complexity of NLDR algorithm clearly.