CASIA OpenIR  > 毕业生  > 博士学位论文
基于视觉词典的三维人脸识别和分类
其他题名3D Face Recognition and Categorization Based On Learned Visual Codebook
钟诚
学位类型工学博士
导师谭铁牛
2009-05-31
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词三维人脸识别和分类 三维人脸特征表达 视觉词典 3d Face Recognition And Categorization 3d Facial Feature Representation Learned Visual Codebook
摘要随着社会的进步和计算机网络技术的迅速发展,基于生物特征的身份鉴别算法由于其出色的安全性能受到了广泛的关注。在所有的生物特征识别算法中,人脸识别技术由于其易接受性、采集方便等优点一直是该领域的研究热点。但是由于人脸本身是一个非刚性的三维几何体,因此传统的二维人脸识别技术必然会受到光照、姿态和表情等因素的影响,造成识别信息的丢失。为了解决这些问题,三维人脸识别技术应运而生。三维人脸识别和分类就是赋予计算机类似于人一样的判断能力,可以通过分析由三维扫描仪采集得到的三维人脸数据进行身份识别和分类。本文针对如何准确而鲁棒的进行三维人脸特征表达进行了深入的研究,主要的工作和贡献包括: 1.实现了三维鼻尖区域的定位算法,并以此为基准区域对不同的三维人脸数据进行配准。这些工作为三维人脸识别平台的建立奠定了基础。 2.提出了基于视觉词典的三维人脸特征并将其成功应用于三维人脸识别系统中。视觉词典算法首先通过学习得到三维人脸最常见的纹理基元,然后以这些基元作为纹理直方图的基础,通过将原三维人脸图像向这些纹理基元映射得到视觉词典直方图向量,以此作为三维人脸特征表达。实验结果表明,视觉词典特征实现了泛化能力、识别性能和鲁棒性的统一。 3.改进了基于视觉词典的三维人脸框架中的各个步骤(滤波器选择、聚类算法设计和匹配距离选择),进一步提升了该算法的识别性能。在FRGC2.0和CASIA三维人脸数据库的实验中,都取得了很好的识别效果。 4.提出了鲁棒局部Log-Gabor直方图特征以克服在非可控环境下三维人脸识别遇到的困难。该算法在FRGC2.0三维人脸数据库的大表情数据子集和CASIA三维人脸数据库表情子集的实验中,取得了很好的识别效果。 5.利用视觉词典特征提出了模糊三维人脸种族分类。利用视觉词典特征学习得到东方人种视觉基元和西方人种视觉基元,并以此为基础设计了模糊隶属度函数,实现了模糊三维人脸种族分类的目的。 总的说来,本文在基于视觉词典的三维人脸特征表达方向做了深入的研究,并取得了一些初步成果,希望本文的工作可以对进一步的三维人脸识别研究提供帮助。
其他摘要Biometric identification has received much attention due to the increasing demand on reliably characterizing individuals. Of all the biometric features, face recognition remains one of the most active research issue because of its advantages, such as most accessible and easier to collect. However, since the human face is a 3D deformable object with textures in nature, traditional intensity face recognition must be influenced by illuminations, poses and expressions, which results in the loss of some discriminative information. To solve these problems, 3D face modality has attracted more and more attention in recent years. 3D face recognition and categorization tries to make it possible that the computers can make the recognition and categorization decisions similar to humans, which can be achieved by accurate analysis of the 3D facial data collected from some 3D scanners. In this thesis, we have a comprehensive study on the features which can efficiently and robustly characterize the 3D facial data based on the Learned Visual Codebook. The main contributions are as follows: 1.We accurately locate the nose area in 3D facial data, based on which we also achieve face registration between different individuals. 2.We propose Learned Visual Codebook (LVC) features to efficiently represent the 3D facial data, which can be successfully applied into the 3D face recognition system. Some 3D facial textons are first learned using clustering. Then these learned textons are adopted as the bases of the histogram. And a histogram vector can be obtained by mapping the 3D face into these learned textons, which is the final representation of the original 3D facial data. Experimental results illustrate that LVC can combine the generality, efficiency and robustness together. 3.We also modify the LVC algorithm in each procedure of the recognition framework, including filter responses, clustering and matching distances selection, which significantly improve the recognition performance in both the FRGC2.0 and CASIA 3D Face Database. 4.We propose the Robust Local Log-Gabor Histogram (RLLGH) features to overcome the problems encountered in un-controlled environments, which achieve promising performance for both the expression situations in CASIA 3D Face Database and the large expression situations in FRGC2.0 3D Face Database. 5.We introduce the fuzzy 3D face categorization based on the LVC features. First the eastern codes and western codes are learned from the predefined traini...
馆藏号XWLW1329
其他标识符200518014628078
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/6198
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
钟诚. 基于视觉词典的三维人脸识别和分类[D]. 中国科学院自动化研究所. 中国科学院研究生院,2009.
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
CASIA_20051801462807(13120KB) 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
[钟诚]的文章
百度学术
百度学术中相似的文章
[钟诚]的文章
必应学术
必应学术中相似的文章
[钟诚]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。