基于重叠片排列的流形学习算法

Manifold Learning Algorithm Based on Overlapped Patch Alignment

针对线性数据降维算法对处理非线性结构数据的降维效果不是很好,提出一种基于重叠片排列的流形学习算法,该算法根据局部的线性贴片处在非线性流形中的特性,将流形划分为线性互相重叠的局部区域贴片,且利用主成分分析方法得到局部区域贴片的低维表示,然后排列且对齐其低维坐标,以获得整体数据的低维坐标。通过仿真结果证明,基于重叠片排列的流形学习算法在应用于人脸识别和分类问题时以及在识别准确率方面要优于其他经典的流形学习算法。

Linear data dimensionality reduction algorithms have difficulty in handling the data with non-linear structure. A manifold learning algorithm based on overlapped patch alignment is proposed. The algorithm uses local linear characters of nonlinear manifold to decompose the manifold into the overlapping linear local area, and acquires local low-dimensional coordinates by principal component analysis algorithm. Then the coordinate alignment technology is used to align each local coordinate, which obtains the entire low-dimensional represen- tation of the data set. The simulation results show that the face recognition and classification accuracy of the proposed manifold learning algorithm based on overlapped patch alignment is superior to other popular manifold learning algorithms.