摘要
训练样本选择是支持向量机应用研究领域的重要课题之一。为此提出了一种类内模式选择新方法。该方法从选择集子空间逼近原类别样本子空间的思想出发,通过迭代,逐一选择那些到已选样本集所在子空间距离最远的样本。在MIT-CBCL人脸识别数据库training-synthetic子库上的同其他方法的比较识别实验中,表明该文方法在选样比率、选样时间以及SVM测试时间等方面均取得了较为明显的优势。
Sample selection is an important topic for SVM.To attack it,a novel intra-class method based on subspace approximation of training class dataset is proposed in this paper.In one class,the subspace of the chosen set is used to approximate that of the original set.An iterative algorithm is employed to realize this process.The furthest sample to the subspace of the chosen set is selected at each step.The comparative experiments on the training-synthetic set of the MIT-CBCL face recognition database show that much lower selection ratio,much less sampling time and much faster test speed has been obtained by this approach combined with linear SVM without a loss of accuracy.
出处
《计算机工程与应用》
CSCD
北大核心
2007年第20期14-17,共4页
Computer Engineering and Applications
基金
国家自然科学基金(the National Natural Science Foundation of China under Grant No.60472060
No.60632050)
关键词
样本选择
子空间
支持向量机
人脸识别
模式分类
sample selection
subspace
Support Vector Machine （SVM）
face recognition
pattern classification
