【作者】 董冀媛； 穆志纯； 欧阳定恒；

【Author】 DONG Jiyuan,MU Zhichun,OUYANG Dingheng School of Information Engineering,University of Science and Technology Beijing,Beijing,100083,P.R.China

【机构】 北京科技大学信息工程学院；

【摘要】 局部切空间排列算法(LTSA)是一种学习高维观测数据集所在低维流形的有效的非线性降维工具,然而存在的一些问题影响了它的进一步应用。本文分析了LTSA用于图像识别时需要解决的问题,有针对性地提出了一种基于最大线性片划分的局部切空间排列算法(MLPPLTSA)。首先提出最大线性片(MLP)的概念及划分策略,把数据集划分成多个相互重叠的最大线性片,然后对这些线性子集作局部主成分分析得到局部坐标,再通过最大线性片之间相互重叠的较少量的样本点引导建立所有样本点的整体坐标。这样不但避免了LTSA在求取整体坐标时大规模特征值分解的问题,新来的样本点也可以直接通过线性变换得到整体低维嵌入坐标。本文将其用于多姿态人耳图像的特征提取。实验结果表明这种算法更加适用于图像识别的任务。更多还原

【Abstract】 As an effective nonlinear dimension reduction tool,Local Tangent Space Alignment algorithm(LTSA)can construct the global low-dimensional embedded coordinates of the data points sampled in high-dimensional space.However,there have been some technical problems unresolved as yet for its further applications.In order to solve these problems,an improved Local Tangent Space Algorithm based on Max Linear Patch Partition(MLPPLTSA)is proposed in this paper.First,the sample space is divided into several overlapping patches using a max linear patch partition strategy.Then local principal component analysis is performed on each patch to get the local low-dimensional coordinates of the data points in it.At last the overlapping points in the patches are used to get the global low-dimensional embedded manifold by local affine transformations.The MLPPLTSA can avoid large-scale matrix eigenvalue decomposition when samples are in large quantities and provide more simple transformations for new coming data to calculate the global embedded coordinates.This improved LTSA is applied to the multi-pose ear recognition in this paper.Experimental result shows its validity in image recognition tasks.更多还原