Abstract: A novel intra-class sample selection method named kernel subclass convex hull sample selection algorithm is proposed and used for SVM. The algorithm is an iterative procedure based on kernel trick. At each step, only one sample furthest to the convex hull spanned by chosen samples is picked out in the feature space. Experiments show that a significant amount of training data can be removed without sacrificing the performance of SVM, while the memory requirements and the computation time of the classifiers are reduced significantly.

Key words: sample selection, convex hull, support vector machine, kernel function, face recognition

摘要： 提出一种基于核函数方法的类内训练样本选择方法——核子类凸包样本选择法，并将其用于支持向量机。该样本选择方法通过迭代方法，逐一选择了那些经映射后“距离已选样本”，并将其映射、生成“凸包最远的样本”。实验结果表明，该方法选择的少量样本使支持向量机获得了较高的识别比率，减少了存储需求，提高了分类速度。

关键词: 样本选择, 凸包, 支持向量机, 核函数, 人脸识别

CLC Number: 

• TP391