文档介绍：华侨大学
硕士学位论文
人脸识别系统的设计与实现
姓名:池静
申请学位级别:硕士
专业:计算机应用技术
指导教师:陈锻生
20071201
华侨大学硕士学位论文

摘要
论文在研究现有经典的人脸检测与识别算法的基础上,构建了一个计算机
视觉应用系统框架,并围绕着“人脸检测”与“人脸识别”两大任务,实现了该
框架的系统原型。
首先,对 Viola 等人提出的基于类 Haar 特征及 AdaBoost 的人脸检测算法进
行了改进,将原来的离散 AdaBoost 算法改用允许弱分类器具有连续置信度输出
的连续 AdaBoost 算法,加快了算法的收敛。相应的,将原来的单一阈值弱分类
器改为输出几乎可模拟任意概率分布的特征查找表(FST)弱分类器。
其次,研究了子空间降维的几种主要方法,包括线性的(PCA 和 LDA)和
非线性的(流形学****算法,以及在人脸识别中的应用,并给出了对比实验结果。
接着,介绍了两种分类器的设计,BP 神经网络和支持向量机,提出了一种
新的基于 Partial Hausdorff 距离(PHD)的图像匹配方法,并将该方法与 BP 神
经网络和支持向量机相结合,即:分类器输出的概率最大的前几个候选类别与待
识别人脸图像进行匹配,找到最佳的匹配作为最后的输出。
最后是人脸识别系统的一些实现细节和程序界面的设计。

关键字:人脸检测人脸识别 AdaBoost 人工神经网络支持向量机 Hausdorff
距离主分量分析线性判别分析流形学****br/>

II
华侨大学硕士学位论文
ABSTRACT

A framework of application system development puter vision is
proposed in this paper, after researching the currenty classical algorithms for
face detection and recognition. And then, a system prototype is implemented
centered by the tasks of "Face Detection" and "Face Recognition".
Firstly, the face detection method presented by Viola et al based on
Haar-like feature and AdaBoost algorithm is improved. The original discrete
AdaBoost algorithm is replaced with real AdaBoost one which allows the weak
classifiers have confidences in real values as outputs. It accelerates the
algorithm' s convergence. Accordingly, the original weak classifier with a
single threshold value is substituted by Feature Search Table (FST) whose
outputs can be suit for almost any distribution.
Secondly, several subspace methods for dimension reduction and its
applications in face recognition are discussed, including linear (PCA and LDA)
and non-linear (manifold learning) algorithms, and then, the experimental
results are shown.
Thirdly, the designs of two classifiers are described, including BP neural
network and support vector machine. And a novel method for image matching
based on Partial Hausdorff Distance (PHD) is