多流形判别分析在人脸识别中的研究

RESEARCH ON MULTI-MANIFOLD DISCRIMINANT ANALYSIS IN FACE RECOGNITION

局部保持投影LPP(Locality Preserving Projection)是一种有效的非线性降维方法,能够使投影降维后的数据与原输入空间中的相似局部结构保持一致,但是该方法没有充分利用类间样本点的权重等重要信息。为了解决这个问题,提出基于Fisher准则的多流形判别分析FMMDA(Fisher Multi-Manifold Discriminant Analysis)方法。结合Fisher准则训练样本类内拉普拉斯图和样本均值类间拉普拉斯图,既保持了原样本的相似局部结构,又充分地利用了不同类别之间的权重。在ORL及Yale人脸库上验证了该方法的有效性。与其他几种最先进的方法相比,FMMDA方法取得了更好的识别效果。

Locality preserving projection （LPP） is an effective nonlinear dimensionality reduction method which could preserve the similar local structure of the dimensionality-reduced data after projection in accord with that in original input space. However, it fails to take full advantage of the important information of the weights of between-class sample points. In order to address this issue, a new multi-manifold discriminant analysis method based on Fisher criterion is proposed, which combines the within-class Laplacian graph of training samples and the between-class Laplacian graph of sample average in Fisher criterion, while preserving the similar local structure of original sample, it also brings the weights between different classes into full play. The effectiveness of the method has been validated on ORL and Yale face database. Comparing with other state-of-the-art methods, the proposed FMMDA method achieves better recognition effect.