人脸特征选择中的SVM泛化误差估计

根据统计学习理论，特征选择可以通过有效的特征搜索策略最小化某个预测泛化误差及其它相关性能来实现。本文研究通过递归特征排除法（Recursive Feature Elimination,RFE）最小化SVM VC留一法(Leave-One-Out, LOO)误差或支持向量span误差估计选择优化特征子集问题，并将最小化VC LOO误差或支持向量span误差估计作为Wrapper特征选择模型的选择判据。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。将特征选择与分类器设计结合，理论上优于传统的特征提取或特征选择方法。为此，本论文将WT和KPCA作为过滤模型（Filter），最小化SVM泛化误差估计作为封装模型（Wrapper），结合这两种模型的优势提出人脸特征选择及识别的新框架。并在UMIST人脸数据库上进行了相应的实验，结果显示提出的特征选择方法和特征搜索策略及人脸特征选择构架有效可行。

SVM Generalization Error Estimation for Facial Features Selection

According to statistical learning theory, feature selection is realizable by valid heuristic search stage, which minimizes an estimated generalization error or some other related performance measure of SVM. In this paper, we introduce a facial features selection method. The optimal features subset is selected by minimizing VC leave-one-out (LOO) error or span error estimate of support vectors, which are regarded as the feature selection criterion of wrapper approach, through recursive feature elimination (REF). Face recognition is essentially a pattern recognition problem with typical small-sample size in sparse hyper-high dimensional space. The basic or most important part is how to obtain the significant features for classification. Theoretically, combing feature selection with classification model design outperforms traditional feature extraction or feature selection methods. Therefore, we propose a novel framework of the facial features selection based on filter (WT+KPCA) and wrapper (minimizing generalization error estimation) approaches. Experimental results on UMIST face database indicates that the proposed feature selection framework is time efficiency and has a significant improvement on the classification accuracy.