基于局部线性嵌入(LLE)非线性降维的多流形学习

Multi-manifold learning using locally linear embedding(LLE) nonlinear dimensionality reduction

为了研究多人脸多表情数据集的多流形学习问题,提出了一种基于局部线性嵌入（LLE）算法的多流形学习方法。对于分布在不同流形上的高维数据,该方法在降维的同时首先对数据集进行非监督的聚类,然后分析每一类数据的低维流形的本质维数以及流形空间的构成,聚类及流形空间的确定是通过对LLE降维的结果进行分析而完成的,计算复杂度小。在Cohn-Kanade人脸表情数据库上的表情识别实验表明,该方法在多人脸多表情流形的学习中优于基本的LLE算法,表情的识别率提高了20%~40%。

A variant of the locally linear embedding （LLE） technique is used to learn multiple low-dimensional facial expression manifolds formed by multiple subjects with multiple expressions. The algorithm first separates the multivariate expression data distributed on several disjoint manifolds into different groups and then analyzes the intrinsic dimensionality and the low-dimensional manifold representation for each group of data. The simultaneous data grouping and the intrinsic dimension detection are both automatic, with reasonable computational loads. Recognition tests using the Cohn-Kanade facial expression database show that the algorithm is superior to the original LLE in terms of the multi-manifold subspace learning, increasing the expression recognition rate from 20% to 40%.