CASIA OpenIR  > 毕业生  > 博士学位论文
基于表示学习的人脸数据分析
王晓波1,2
学位类型工学博士
导师李子青 研究员
2018-05-24
学位授予单位中国科学院研究生院
学位授予地点北京
关键词人脸数据分析 人脸聚类 人脸识别 子空间聚类 字典学习 深度学习
其他摘要

人脸聚类和人脸识别是人脸数据分析领域中的关键技术。由于采集的人脸数据容易受光照、姿态、模糊、低分辨率、遮挡等因素的影响,人脸数据往往包含着多种噪声,并由此带来较大的类内差异和类间相似性,给聚类和识别的性能造成了重大影响。

    本论文集中于解决人脸数据分析中存在的难点和问题,尝试从算法角度研究解决这些问题的方法。具体来说,对基于自表达的人脸聚类、基于判别字典学习的人脸识别和基于深度学习的人脸识别算法进行改进并提出新的方法,主要贡献点可归纳为以下几点:

    基于自表达的人脸聚类。首先在单视角聚类算法中,提出了利用数据间的高阶信息(协同参考信息)来构建相似度矩阵。如果两个数据点在自我表示时,共享的参考样本越集中,越类似,则说明两个数据点间的相似度是越高的,否则,相似度越小。该方法构建的高阶相似度矩阵,鲁棒性较强,在人脸聚类数据库上取得了不错的性能。在多视角聚类算法中,主要的目标是融合多视角的信息,为考虑视角间的补充信息,提出了位置感知的排外约束项,可有效地避免视角间的尺度差异问题。同时,将自表达学习和后续的谱聚类算法融入到一个框架中,在多视角人脸聚类问题上取得了不错的性能。最后,由于每个视角的信息都是片面,不完整的,考虑从这些片面的信息中学习出完整的空间,在完整的空间上构建相似度矩阵,可有效提升聚类算法的性能。

    基于判别字典学习的人脸识别。首先是监督的判别字典学习,从理论上完善了基于pair-wise约束算法的收敛性,给出了相关的理论证明。设计的约束项使得同一对人脸在字典表示空间尽可能地接近,而不同人脸对在字典表示空间则尽可能地远。该算法在人脸数据库上取得了不错的性能。为充分地利用无标签数据,提出了半监督字典学习算法,利用分类器对无标签数据进行置信度的判定,选择置信度较高的无标签样本进行再学习,以学习判别的字典,用于人脸识别任务。

    基于深度学习的人脸识别。针对现有深度学习的人脸识别常用Softmax损失函数进行训练,但Softmax训练出的人脸特征判别性不强的缺点,提出了软间隔的Softmax函数,使得不同类别之间的人脸特征具有一定的间隔。同时由于单个分类器一般是弱分类器,集成多个弱分类器以构建强分类器,用于人脸识别。

    总结而言,本论文主要贡献在于,提出多个人脸数据分析的算法,对现有的工作在理论和算法上提出创新和改进,提升了人脸数据分析的性能。

;

    Face clustering and face recognition are key technologies in the field of face data analysis. Due to the influence of lighting, pose, blurring, low resolution, occlusion and other factors, face data often contains multiple types of noise and results in large intra-class difference and inter-class similarity. Therefore, face clustering and face recognition are difficult tasks in data analysis.
In this thesis, we focus on solving the difficulties and problems in face data analysis and try to solve them from the perspective of algorithm. Specifically, this thesis improves face clustering based on self-expression learning, face recognition based on discriminative dictionary learning, and face recognition based on deep learning, and proposes new methods for both face clustering and face recognition. Detailedly, the contributions can be summarized as follows:

    Firstly, in the single-view face clustering based on self-expression algorithm, we have proposed to use the high-order information (co-reference) between data points to construct an informative similarity matrix. Specifically, if two data points share similar reference samples to represent them, it means that the similarity between these two points is higher, otherwise, their similarity is smaller. The constructed similarity matrix is more robust than directly using the linear self-expressing representation and achieves better performance on the face clustering databases. On multi-view face clustering, the goal is to fuse information from multiple views. To consider the supplementary information between different views, the position-aware exclusivity is designed, which can effectively avoid the scale issue between different views. Meanwhile, we integrate the self-expression learning and the subsequent spectral clustering into one framework and achieve good performance on multi-view face clustering. Finally, note that the information of each view is incomplete, it is beneficial to learn an intact space from these partial information and construct a similarity matrix in an intact space.

    Face recognition based on discriminative dictionary learning is widely used in previous years. However, the convergence of the pair-wise constraint algorithm is not guaranteed, and without giving the theoretical proofs. In this thesis, we finish the theoretical part and enforce the same face pair to be as close as possible while different face pairs are as far apart as possible. To make full use of the unlabeled data, we have proposed a semi-supervised dictionary learning algorithm. The classifier was used to determine the confidence level of the unlabeled data, and the high confidence points were selected to learn the discriminative dictionary.

    To face recognition based on deep learning, Softmax loss function is commonly used. However, the deep features learned by Softmax loss are usually not strong. Therefore, a soft-margin Softmax function is proposed to explicitly enforce the margin between different categories. Meanwhile, since a single classifier is usually a weak classifier, multiple weak classifiers are assembled to build a strong classifier for face recognition.

    In summary, in this thesis, we have achieved a great number of improvements for face clustering and face recognition, in both theoretical and algorithmic aspects.

 
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/20998
专题毕业生_博士学位论文
作者单位1.中国科学院自动化研究所模式识别国家重点实验室
2.中国科学院大学
推荐引用方式
GB/T 7714
王晓波. 基于表示学习的人脸数据分析[D]. 北京. 中国科学院研究生院,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
Thesis_Xiaobo_签名_Fin(9710KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[王晓波]的文章
百度学术
百度学术中相似的文章
[王晓波]的文章
必应学术
必应学术中相似的文章
[王晓波]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。