摘要： 改进的局部切空间排列(ILTSA)算法解决了当样本稀疏、分布不均匀或数据流密度曲率变化较大时,局部 切空间排列算法不能揭示流形结构的问题,用于人脸识别能提取更好的低维特征,但不能有效处理不断增加的数 据集的问题。为此,提出一种可泛化的ILTSA(GILTSA)算法。结合类别信息定义样本间的距离实现各样本的近 邻集选择,基于ILTSA 算法求解训练样本集的低维流形,对每个新样本寻找其在训练样本集中的最近邻,然后根据ILTSA 算法原理求得其近似低维流形。在ORL、Yale 和埃塞克斯大学人脸库上的实验结果表明,与主成分分析算法和线性局部切空间排列算法等相比,GILTSA 算法具有更好的识别率。

关键词: 流形学习, 局部切空间排列, 泛化, 特征提取, 人脸识别

Abstract: The Improved Local Tangent Space Alignment(ILTSA) can obtain better low dimension feature for face recognition because it can efficiently recover the problem that the Local Tangent Space Alignment(LISA) fails to reveal the manifold structure in the case when data are sparse or non-uniformly distribute or when the data manifold has large curvatures. To solve the problem that the ILTSA cannot efficiently handle ever-increasing data set,this paper presents a Generalization method for the ILTSA(GILTSA). The nearest neighborhood set is obtained based on the distance defined according to the classes of the samples,then the low manifold of the training set is implemented using the ILTSA. Through finding the nearest sample in the training set,and the low manifold of a new sample is approximately calculated by the projection of its nearest sample. Experimental results on the ORL,the Yale and the University of Essex face image database indicate that the proposed GILTSA method increases the overall accuracy compared with Principal Component Analysis(PCA) and Linear Local Tangent Space Alignment(LLTSA) algorithm etc.

Key words: manifold learning, Local Tangent Space Alignment(LTSA), generalization, feature extraction, face recognition

中图分类号: 

• TP18