中国法分类号tp39中国图象图形学报.docVIP

中国法分类号：TP391.41 文献识别码：A 文章编号: 格拉斯曼流形上的半监督判别分析 姜伟1, 陆瑶1, 杨炳儒2 （1.辽宁师范大学 数学学院，辽宁 大连 116029；2.北京科技大学 计算机与通信工程学院, 北京 100083） 摘 要： 最近研究表明，将图像集合表示为格拉斯曼流形上的点能够获得更好的识别性能。然而，传统格拉斯曼流形上的判别分析方法仅考虑了带标签样本的统计信息，忽略了无标签样本。鉴于此，，标准上的实验表明了的有效性。Yang Bingru2 (1.School of Mathematics, Liaoning Normal University, Dalian 116029, China; 2. School of Computer Communication Engineering, University of Science and Technology Beijing, Beijing 100083, China) Abstract ：Recent research has shown that better recognition performance can be attained through representing image sets as points on Grassmannian manifolds. However, the conventional discriminant analysis methods based on such manifolds take into account only the statistical information of labeled samples and suffer from ignoring unlabeled samples. To address this issue，based on manifold regularization, a novel method, called Semi-supervised Discriminant Analysis on Grassmannian Manifold(SDAGM), is presented and applied to image sets recognition problem. In SDAGM, a nearest neighbor graph is constructed to capture the local geometrical structure of all samples on Grassmannian manifold and incorporated into the objective function of Discriminant Analysis on Grassmannian manifold as a regularization term. Not only does the proposed algorithm consider the label information, but also utilize a consistency assumption. The feasibility and effectiveness of SDAGM are verified on several standard data sets with promising results. Key words: Grassmann Manifold; Image Sets; Discriminant Analysis; Semi-supervervised 0 引言 在传统的图像识别领域，通常是以单幅静态图像作为研究目标，在可控场景的情况下取得了良好的性能[1-2]。然而，基于单幅静态图像学习算法在很多实际应用场景中效果不佳[3-4]。随着数字成像和海量存储设备等硬件设备快速发展以及互联网资源的触手可及，人们能够获取到图像涵盖了物体的各种表观变化，如姿态、表情、光照角度等，从而为基于图像集合[4-6]学习问题提供了丰富的数据。在这类问题中，每个图像集合通常包含同一类别的多个图像样本，利用了图像样本综合信息提取特征，提升算法的性能，因此，基于图像集合的识别问题受到越来越广泛的关注。 研究表明，将图像集合建模为子空间[4,7]，可以达到更好的性能。Yamaguchi[8]提出了针对人脸识别问题的互子空间方法（Mutual Subspace Method，MSM），开创了利用子空间对图像集合非参数建模方法研究的先河，引起各领域科学研究者的广泛关注，衍生出一系列改进算法。Fukui[9]提出约束互