【作者】 李春晓；

【导师】 任获荣；

【作者基本信息】 西安电子科技大学 ， 测试计量技术及仪器， 2011， 硕士

【摘要】 人脸识别是模式识别和机器视觉领域最富挑战性的研究课题之一,它在人机交互,安全监视,身份认证等方面有着非常广阔的应用前景。在目前的特征提取算法中,基于Fisher准则的线性判别分析(LDA)算法是一种较成功的特征提取算法。但是LDA用于人脸识别时,遇到三个问题：(1)小样本问题；(2)LDA利用同一个准则函数把所有样本投影到同一个特征空间,忽略了不同类的样本分布的特征差异；(3)LDA对于复合模型分类能力较低。本文针对问题一,利用最大散度差线性鉴别分析来解决；对问题二,利用类依赖线性判别分析方法,对每一类样本根据一定的准则函数,建立一个反应本身类别特性的投影矩阵；对问题三,利用增强LDA,通过两个不同的加权矩阵对类间离散度矩阵和类内离散度矩阵进行重新定义,改进的Fisher准则可以在降维后的子空间中保存局部信息,使距离较近的类别投影后不会混在一起。最后本文在这些方法的基础上提出了一种改进方法,类依赖增强LDA,充分发挥了增强LDA和类依赖的优点,最后在ORL和YaleB人脸库上进行实验,实验结果表明了该方法的有效性。更多还原

【Abstract】 Face recognition is one of the most challenging problems in the fields of pattern recognition and machine vision. It has a wide range of promising application, such as human-computer interaction, security surveillance, identity authentication, and so on. Linear Discriminate Analysis algorithm (LDA) based on the Fisher criterion is a successful algorithm among the feature extraction algorithms.However, there are three problems with LDA for face recognition. First one is the small training sample size(SSS) problem, we introduce the maximum margin criterion(MMC) to solve this problem;The second one is LDA projects samples from all classes to one single subspace obtained by the same criterion function. It ignores the fact that a class has its own optimal feature vector discriminating itself from the other classes; we introduce the class-dependent LDA (CDLDA).In the scheme of CDLDA, a class has its own optimal feature vector to discriminate itself from the other classes. The third one is it would collapse the samples from different classes into one single cluster when the class distributions are multimodal. An enhanced LDA method is proposed. The between-class and the within-class scatter are reformulated by introducing two different weighted matrices in respective, which can preserve the local structure of different class in the reduced subspace. Then it can avoid collapsing samples of different classes into one single cluster.In the end an improved method based on these above methods is presented, called class-dependent and enhanced LDA (CD-ELDA).Experiments show encouraging performance of the proposed algorithm.更多还原