节点文献

人面部的代数特征与几何特征的提取及识别

Human Facial Algebraic and Geometrical Feature Extraction and Recognition

【作者】 王宏漫

【导师】 欧宗瑛;

【作者基本信息】 大连理工大学 , 机械设计及理论, 2003, 博士

【摘要】 基于生物特征的身份鉴定技术在社会生活中具有越来越重要的地位和作用。在多种生物特征中,基于人面部特征的识别和鉴定因为具有无侵害性、成本低、易于接受等优点,得到广泛的研究和应用。 本论文主要从人面部的代数特征和几何形状特征两个方面,探讨了人面部的特征提取和识别问题。 在代数特征方面,主要从主成分分析和独立成分分析两个角度出发,进行了探讨。主成分分析立足于图像的二阶统计特性,提取图像的灰度特征;它去除了各特征分量之间彼此的相关性,并且可以根据其对应的特征值的大小确定它在整个能量中占据的百分比,从而可以分离出信号中的主要成分和次要成分;独立成分分析从图像的高阶统计特性出发,提取图像的灰度特征。它的各特征分量之间彼此独立,由于不考虑主次关系,因此它难以清楚地反映信号各分量的主、次程度。考虑到这两种方法的不同特点,论文提出了主要独立成分分析方法,即在主成分分析的基础上,进一步做独立成分分析。通过这种方法得到的特征分量的均值、方差的分散性要优于仅做主成分分析的特征分量的均值、方差的分散性,因此聚类效果要更好一些。 对应于既定的特征选取,存在有一个最佳的相似性度量方法的选择问题。在考虑关于人脸的分类算法时,大多是评价分类算法对噪声样本的容错性,而忽视了特征空间维数的选取与相似性度量准则之间的关系。就特征空间选取问题,论文从理论上分析了最近邻法、支持向量机这两种方法的识别性能,由于支持向量机是通过类间超平面进行分类的,最近邻法是通过类间距离进行分类的,因此,论文认为,在一定程度上支持向量机更敏感于特征空间的变化,尤其是特征空间中的细节分量或噪声分量的变化,相比之下,在一定条件下,最近邻法对特征空间中的细节分量并不敏感。因此在论文所提出的人脸识别系统中,在较小的特征空间中支持向量机的识别效果要优于最近邻法。 在模式识别理论中,人脸识别是一种典型的多类判别问题。鉴于支持向量机是一种以两类判别为基本功能的分类函数,在用于多类判别时,本文提出的方案是以1-1判别策略为基础,根据各判别函数的分类间隔的差异对多个判别函数进行排序,对于“拒绝决策”的情形,则可利用判别函数间的冗余再判别来降低识别误差。这种基于支持向量机组的淘汰法对于每个判别函数的使用更加合理和具有选择性。 根据所选择的人脸的代数特征和支持向量机的分类性能,论文提出了一个脸像识别方法,即基于主成分分析对人脸图像进行降维,再用独立成分分析进行独立分量特征提取,最后基于SVM进行多类识别。该脸像识别方法在对ORL人脸图像库和自建的人脸图像库的测试实验中,识别率分别为97.5%、88.17%。 在几何形状特征方面,论文重点研究了人脸主要几何特征点的提取和脸部主要器官的连续形状的提取。 在人脸几何特征点提取问题中,首先需要准确定位人脸的左右边界。基于原有的灰度投影曲线定位算法,本论文进一步对原灰度图像做小波分解,然后仅取垂直分量进行重建,再沿垂直方向做投影,可以对人脸的左右边界进行更加精确的定位。 分形理论在图像处理中的应用表明,分维图像能够更加有效地抑制噪声,并能够充分地反映图像纹理的变化。因此,在提取人脸的几何特征点时,采用脸部图像的分形维数图像取代了传统的灰度图像或二值图像。 在提取面部轮廓及其主要器官的连续形状时,需要采用边缘检测与提取等手段。 在传统的微分算子的基础上,论文提出了一种广义的计算图像灰度梯度的方法一一自由差分运算。自由差分运算不仅是一些典型的微分算子的通用描述,而且它突破了传统微分算子的模板运算模式,可以同时从图像的全局区域、局部区域两个方面出发,沿若干个直线方向进行计算,以考察象素点的边缘特征。它把对图像这种二维信号的微分计算转化成一维方式进行,计算方向具有任意性,同时,便于采用多种手段对象素点的边缘性质进行分析,使得多种边缘检测方法不再彼此独立,为信息的二次融合提供了可能性、方便性。从对实际图像的测试效果来看,它也具有很好的抗噪性。把它与Snake模型相结合,论文提出了正交型Snake模型。 从传统的Snake模型的思想出发,正交型Snake模型利用自由差分运算的基本原理,通过在正交直线方向上计算灰度梯度,把轮廓变化的信息和图像灰度变化的信息更紧密地联系在一起。在曲线的演化过程中,以基本多边形为依托,以目标曲线的特征为指导,让轮廓的变化能够自适应于图像灰度变化,使得所提取的轮廓曲线的特征不断地逼近目标曲线的特征。在对多种图像的测试中,取得了比较理想的效果。传统的边缘提取算法是一个由底至上的过程;传统的Snake模型是一个由上至下的过程,在这两类方法中,信息处理的基本流程都是单向的;正交型Snake模型通过强调高层视觉和底层信息之间的相互联系在曲线演化过程中的作用,使得信息的流动过程处于闭环的状态。 基于正交型Snake模型,分别研究了脸部的轮廓、眼部和嘴部轮廓的提取。在处理眼的轮廓时,根据眼所具有的独特的类圆特点,把沿正交?

【Abstract】 Biometric technologies are becoming the foundation of an extensive array of highly secure identification and personal verification solutions. Among the features measured, facial feature identification and verification are gaining popularity and diverse applications for the reason that they are considered to be non-invasive, low cost, and natural biometric technologies.The paper discusses facial feature extraction and recognition from two aspects ?algebraic and geometrical features.The paper discusses the algebraic feature adopting principle component analysis (PCA) and independent component analysis (ICA) respectively. Principle component analysis extracts the image gray characteristics considering the second order moments, while independent component analysis accounts for higher order statistics and identifies the independent source component from their linear mixtures. Considering the differences between them, the paper proposes principle independent component analysis (PICA) that is doing ICA based on PCA. PICA can provide a more powerful classification data representation than PCA by comparing their means and variances.The similarity measurement should match selected features. Most face classification approaches pay more attention to dealing with noise samples, but ignore the influence of selected feature space. Support vector machine (SVM) performs classification by the hyperplane between classes, while the nearest neighbor method (NN) performs classification by the distance between classes. The paper discusses the difference between NN and SVM, finds SVM is more sensitive to the feature space variation than NN, especially for details or noise component. SVM can get better recognition rate only depending on smaller feature space or less approximation components for face recognition.Support vector machine is an advanced classifier, which has demonstrated high generalization capabilities; however, it was developed originally only for two-class classifying. The paper proposes a new multi-step approach to extend SVM capability dealing with the multi-class face recognition problem by incorporating with the elimination strategy. Based on the one-against-one strategy to classify, it sorts the discrimination functions according to their own Vapnik-Chervonenkis confidences and uses the redundancy among them to decrease the discrimination error for the rejecting decision case.The paper proposes a face recognition approach considering the facial algebraic features and the SVMs’s classification capability. It includes three basic parts: PCA ?for reducing dimensions, ICA ?for feature extraction and SVM ?for multi-classification. The face recognition experiments with ORL face databaseand a compound database. The recognition rates are 97.5% and 88.17% respectively.The paper discusses the geometric feature points extraction, face and its main parts contour extraction also.It is needed locating the left and the right boundaries of face when extracting geometric feature points. Based on the ordinary intensity level vertical projection curve locating algorithm, the paper employees the reconstructed image by the vertical component of wavelet decomposition of the original image to locating more accurately.The application of fractal theory in image processing shows that fractal image can suppress noise effectively, and reflect the variation of the texture enough. Therefore, the paper substitutes fractal image for intensity image or binary image.The paper proposes the free difference operation, which is a new method computing the variation of gray level, based on the classical differential operators. Free difference operation is not only a general description of some typical differential operators, but also breaks the template mode of classical differential operators. By computing grey level variation along line direction, it considers each pixel feature in the range of the global and local region of image, not the image block or subimage. It has the advantage of simplicity, operational flexibility and better

  • 【分类号】TP391.41
  • 【被引频次】15
  • 【下载频次】1080
  • 攻读期成果
节点文献中: 

本文链接的文献网络图示:

本文的引文网络