【作者】 李伟红；

【导师】 陈伟民；

【作者基本信息】 重庆大学 ， 仪器科学与技术， 2006， 博士

【摘要】 人脸识别实质是稀疏超高维空间、典型的小样本模式识别问题。解决这类问题的关键在于如何获得对分类有意义的特征。特征选择的评价将取决于其分类识别的能力。一种好的特征选择方法应该使所选择的特征子集即使采用最普通的分类器也可以获得较好的分类结果。特征选择需要解决两个问题,一个是特征选择的判据或评估判据问题;另一个是特征搜索策略问题。本论文首先对支持向量机(Support Vector Machine,SVM)的基本理论及相关的核心技术-最优化理论、核函数特征空间、泛化性理论进行了分析、研究。其次,对特征选择基本方法-嵌入式(Embedded)、过滤式(Filter)和封装式(Wrapper)特征选择模型的特性、特征选择评估判据以及启发式特征搜索方法进行了研究。针对人脸识别问题,提出将小波变换及核主元分析(KPCA)作为Filter特征选择模型,将SVM作为Wrapper特征选择模型。并基于Filter和Wrapper特征选择模型提出了复合的人脸特征选择及识别新框架以及技术解决方案。本论文在提出的新方案中,对以SVM为核心的Wrapper特征选择模型中的特征评估判据问题进行了重点研究。(1)提出将支持向量对分类的贡献取决于SVM训练中具有最小间隔的支持向量原则作为特征评估判据首次用于人脸特征选择。这样,通过GSFS特征搜索策略,从而在一定分类误差阈值下得到优化的人脸特征子集。在ORL、IITL及UMIST人脸图像库上的实验表明,无论在特征维数的降低、分类识别率的性能及分类识别速度方面都有不同程度的改善。(2)提出将半径/间隔R 2 || w ||2作为Wrapper特征选择模型的特征评估判据首次用于人脸特征选择,在基于SVM RFE的特征搜索策略中引入缩放因子并用梯度法代替贪婪搜索法,这种方法将有利于提高人脸识别的泛化能力。这种方法在UMIST人脸图像库上的实验表明,同样可以在保持分类识别率不变的情况下,大幅度降低特征维数,提高分类识别的速度。(3)提出将最小化正则化风险( Rr eg[ f ])作为特征评估判据,通过SVM RFE搜索策略获得优化的特征子集,并首次用于人脸特征选择。这种方法可以有效的抑制特征选择及识别过程中“过学习”问题,调和学习能力与推广性能力间的矛盾,对解决人脸识别的小样本问题更有效、合理。这种方法分别在两个FERET人脸图像子库上进行了相应的实验,结果表明:与传统的特征选择方法相比,均能大幅度降低特征维数,提高分类识别速度,同时分类识别率也有所提高。(4)根据Sample-margin和Hypothesis-margin的定义分别对相关特性进行了分更多还原

【Abstract】 Face recognition is essentially a typical small-sample pattern recognition problem in sparse hyper-high dimensional space. The key to solve the problem is how to obtain the significant features for classification. The criterion to evaluate the performance of feature selection lies on the ability of classification. A perfect feature selection method can make sure the selected feature subset obtain the best classification performance even adopt common classifiers. The aim of feature selection is to solve two problems. One is the criterion of feature selection, and the other is the strategy of feature search. Firstly, the dissertation gave a detailed analysis on the fundamental theory of SVM and corresponding core technologies-optimal theories, kernel function feature space and generalization theory. Secondly, the dissertation did research on the basic feature selection methods’charactristics (including Embedded, Filter and Wrapper feature seletion models), the features selection criterion and heuristic feature search method. The dissertation proposed to treat wavelet transform and KPCA as Filter feature selection model, and treat SVM as Wrapper feature selection model. Also, the dissertation proposed a whole new frame and corresponding solution on multiplex face feature selection and recognition method.The new solution proposed in the dissertation did a deep research on feature evaluation criterion of Wrapper feature selection model based on SVM.(1) The dissertation proposed a feature evaluation criterion that support vector contributions to classification depend on the support vector with margin minimization in SVM training, and applied it to face feature selection for the first time. So, we can obtain the optimized face feature subset in a certain classification threshold by GSFS feature search strategy. The experimental results on ORL, IITL and UMIST face databases demonstrate that the proposed algorithm has improved the performance on the reduction of feature dimensionality, classification recognition rate and the speed.(2) The dissertation proposed to treat radius/margin R 2 || w ||2 as evaluation criterion of Wrapper feature selection model, and applied it to face feature selection for the first time. The dissertation introduced feature selection via adaptive scaling base on SVM RFE, and this method would be useful to enhance the generalization of face recognition. The experimental results on UMIST face database demonstrate that the proposed algorithm could reduce feature dimensionality and enhance the classification更多还原