基于流形学习的泛化改进的LTSA算法

发布时间：2018-01-20 12:23

本文关键词： 改进的局部切空间排列(ILTSA) 人脸识别流形学习可泛化　出处：《计算机工程与应用》2017年03期 　论文类型：期刊论文

【摘要】：在数据稀疏、数据非均匀分布和数据流形具有较大曲率的情况下,传统的局部切空间方法不能够有效地揭示流形结构。提出了一种泛化的ILTSA(GILTSA)流形学习方法,该方法以改进的局部切空间排列算法(ILTSA)为基础,在解决流形结构问题的同时,不仅能够获得用于人脸识别更好的低维特征,而且能有效地处理日益增加的数据集的问题。该方法首先基于样品间距离选择近邻集,实现训练集的低维流形,为每个新样本寻找最近的样本训练集。然后结合ILTSA算法,根据其最近样本投影距离计算低维流形。在ORL的人脸图像数据库的实验、Swiss roll和手书的"2"等实验结果表明,与局部线性嵌入和局部切空间排列算法等相比,GILTSA方法增加了整体精度。
[Abstract]:In the case of sparse data, non-uniform distribution of data and large curvature of data stream shape. The traditional local tangent space method can not reveal the manifold structure effectively. A generalized ILTSAG ILTSA manifold learning method is proposed. Based on the improved local tangent space arrangement algorithm (ILTSA), this method can not only solve the problem of manifold structure, but also obtain better low-dimensional features for face recognition. And it can effectively deal with the problem of increasing data sets. Firstly, this method selects nearest neighbor sets based on distance between samples to realize the low dimensional manifold of training set. To find the nearest sample training set for each new sample, and then combine the ILTSA algorithm, calculate the low-dimensional manifold according to its nearest sample projection distance. Experiment in the face image database of ORL. The experimental results of Swiss roll and "2" in calligraphy show that compared with the local linear embedding and local tangent space arrangement algorithms, the global precision of the GILTSA method is improved.
【作者单位】：哈尔滨理工大学计算机科学与技术学院;
【基金】：黑龙江省自然科学基金(No.F201302)
【分类号】：TP391.41;TP181
【正文快照】： 1引言随着电脑和网络的普及,人们可以很容易地访问大量信息,使表示信息的数据集维数急剧增加,非结构化的特点更加突出[1]。流形学习可以根据数据分析和降维的方法揭示高维数据集的内在规律,利用局部的几何性质更好地了解数据集,以便实现数据降维并且因此,已经在机器学习和认知

【相似文献】