CASIA OpenIR  > 毕业生  > 硕士学位论文
三维人脸识别关键问题研究
Alternative TitleResearch on key problems of 3D face recognition
黄永刚
Subtype工学硕士
Thesis Advisor王蕴红 ; 谭铁牛
2007-06-04
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword三维人脸识别 多模态人脸识别 人脸特征提取 深度图像 融合 关联特征 3d Face Recognition Multimodal Face Recognition Depth Image Fusion Correlative Feature
Abstract利用人脸的空间形状结构信息进行分类的三维人脸识别是人脸识别领域新的研究热点,人们寄希望于它能解决传统二维人脸识别的光照、姿态、表情变化的瓶颈问题,目前正受到广泛的关注,初步的研究成果令人鼓舞。 三维人脸识别系统目前研究热点主要包括数据获取、特征点检测、建模、去噪、特征提取、 分类器设计、与二维人脸信息的融合等关键性问题。本文针对其中的特征提取、分类器设计、 多模态融合等三个问题进行了研究。文中主要工作与贡献包括: 1. 在特征提取方面,提出了一种基于几何特征和关联特征统计量的特征提取方法。一方面从全局统计的角度分析比较了传统的几何特征,如深度、曲率等;另一方面提出关联特征来描述人脸的三维结构,并提出了一种三维局部二值模式描述子(3DLBP)来描述三维人脸的关联特征;最后融合几何特征和关联特征统计量形成一种可分性好的三维人脸特征。 2. 在分类识别方面,提出了一种基于类内类间深度差直方图统计量的识别方法。人脸的深度图像的象素值直接反映了人脸上每点的三维深度信息,他们的相减差就直接反映了两幅人脸结构特征差异。与灰度差异会受到光照变化的强烈影响不同,这种深度差异只受预处理配准误差和表情变化的影响。我们采用计算局部最小差异绝对值来代替两幅深度图像的直接相减,并提出三种框架来融合局部匹配和全局匹配,以减小配准误差和表情变化的影响。实验结果表明经过两步改进处理,两者的影响小于不同人脸的差异影响,所提出的的方法取得很好的识别效果。 3. 在多模态融合方面,研究分析了目前融合二维和三维信息的人脸识别的研究现状,并做了一些尝试性探索。我们将局部GABOR二值模式算子应用于三维人脸识别,并比较分析了一些常见的融合方法的融合性能。 总的说来,本文对三维人脸识别的特征提取、分类识别、多模态融合等环节做了一些初步尝试和探索。
Other AbstractHuman face is intrinsically 3D deformable object with texture. The 3D shape information should not be ignored in face recognition since they provide another type of distinct feature to distinguish different faces. 3D face recognition based on 3D facial information is believed to own the potential to solve the bottleneck in 2D face recognition. Its research is getting more and more attention, and primary research shows promising effects. The research points in 3D face recognition mainly include feature points detection, mesh modeling, noise removing, feature extraction, classifier design and fusion with 2D face. In this thesis, based on the depth image, we focus on three of the problems: feature extraction, classifier design and fusion. Here lists the main work and our contributions in this thesis: 1. We review almost all appeared algorithms on 3D face recognition and 2D+3D multi-modal face recognition. This review can help us realize the state of art in this research area, and provide a base for us to do further research. 2. We propose a new expressive feature for 3D face recognition which bases on the combination of global statistics of geometrical features and local statistics of correlative features. Correlative feature is introduced and analyzed detailedly, and 3DLBP descriptor is proposed to encode correlative features. We believe correlative feature and traditional geometrical feature can be complementary to describe 3D face, and the combination of both features is demonstrated to own great discriminating power. 3. We propose a simple but effective method to discriminating 3D faces.the method is based on the statistics of depth image differences.After analyzing the physical meanings of facial depth images, we believe the difference from depth image substraction can represent the difference between 3D faces. By using histogram proportion, andmaking two improvements to solve the problems caused by registration errors and expression variation, we finally achieve a new recognition method. It is simple, straightforward but works well. In a word, in this thesis, we have made some fruitful attempts and significant progresses on 3D face recognition.
shelfnumXWLW1100
Other Identifier200428014628043
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/7410
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
黄永刚. 三维人脸识别关键问题研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2007.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20042801462804(3298KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[黄永刚]'s Articles
Baidu academic
Similar articles in Baidu academic
[黄永刚]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[黄永刚]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.