Abstract: To solve the problem that too much features have great effects on the instantaneity and accuracy of face recognition, a method named combined facial feature selection based on ReliefF-SVM RFE is proposed. The proposed method uses DCT extract facial feature and ReliefF select feature to reduce the?feature?dimension space initially, then uses improved SVM RFE to select optimal feature. This method solves the problem that the SVM REF feature selection consums long time because of training much features repeatedly. Finally, it uses leave-one-out method to select optimal feature subset from feature ranking table, and classification by SVM. Experiments are performed on UMIST facial database, accuracy of 98.84% is achieved when facial features are 52, recognition time only needs 0.037 s.

Key words: face recognition, Support Vector Machine Recursive Feature Elimination（SVM RFE）, ReliefF, discrete cosine transform, feature selection

摘要： 针对人脸识别中因特征个数较多对识别的实时性和准确性影响较大的问题，提出了ReliefF-SVM RFE组合式特征选择的人脸识别方法。利用离散余弦变换提取特征和ReliefF对人脸图像特征集做特征初选，降低特征维数空间，再用改进的SVM RFE（Support Vector Machine Recursive Feature Elimination）选择最优特征，解决了利用SVM RFE特征选择时因特征数多而算法需多次训练耗时长的问题。对训练得到的特征排序表采用交叉留一验证方法选取最优子集，再由SVM分类识别。在UMIST人脸库上实验证明，可以在特征数为52时，达到98.84%的识别率，识别时间仅需0.037 s。

关键词: 人脸识别, 支持向量机回归特征消除（SVM RFE）, ReliefF, 离散余弦变换, 特征选择