基于支持向量机的人脸姿态判定

Face pose discrimination with support vector machines

对于多姿态人脸检测中的姿态判定问题,提出了一种基于支持向量机(SVM)的人脸姿态判定算法。将人脸姿态划分成6个类别,从一个多姿态人脸库中手工标定出1800幅人脸图像作为训练样本集,分别训练基于支持向量分类(SVC)和基于支持向量回归(SVR)2种姿态分类器。另外标定出300幅人脸图像作为测试样本。SVC方法和SVR方法分别取得了1.67%和3.33%的分类错误率。其中SVC方法的分类效果明显优于在传统方法中效果最好的人工神经元网络(ANN)方法(分类错误率为3.33%)。对比实验结果表明,SVM方法对于解决姿态判定问题是很有效的。

An approach for face pose discrimination using support vector machines (SVM) is proposed for the multiview face detection. Six different types of poses were defined with 1 800 images from a multiview face database as the training set and another 300 images as the test set. Classifiers based on support vector classification (SVC) had a 1.67% error rate while those based on support vector regression had a 3.33% error rate. The SVC performance was superior to an artificial neural network classifier (3.33% error), which performed best among traditional pattern classifiers. Thus the experiments demonstrated that SVM is a feasible approach for face pose discrimination.