摘要
在人脸识别中,传统有效的鉴别分析方法需要更多样本评估类内散度信息。由于人脸的单样本问题,导致某些经典的方法如Fisherface和Eigenface等失效,解决的方法通常是生成各种虚拟样本来扩充训练集以实施这些算法。针对这个问题,根据人脸的对称相似理论,人脸样本的相关变化信息可以从它的对称脸上提取,提出组合原始训练样本及它的虚拟平均脸、对称脸作为训练样本集,应用稀疏理论进行加权积分融合,分两步进行识别的方法,并在ORL和FERET人脸数据库做了对比实验。实验结果表明,该方法比现有一些突出效果人脸识别方法更好,并具有一定的鲁棒性。
In face recognition,due to the single sample problem,some traditional discriminant analysis fail to work,such as Fisherface,Eigenface,etc.,since they need more than one sample to estimate the within-class scatter for discrimination on future test samples. In order to overcome this deficiency,some new virtual sample would be generated to perform these methods usually. According to the facial symmetry theory,some relevant information of possible change could be extract to adapt to the future samples. The proposed method applied sparse representation and proposed a score level fusion method which combined the original train sample,symmetrical virtual face and average virtual face to a new train sample set to perform a two-step classification. The efficiency experiments show that the proposed method outperforms some new face recognition methods and has certain robustness.
出处
《计算机应用研究》
CSCD
北大核心
2015年第5期1527-1531,共5页
Application Research of Computers
基金
广东省学科共建育苗工程项目(2013LYM0055)
韩山师范学院一般项目(LY201301)
关键词
模式识别
人脸识别
稀疏表示方法
人脸单样本问题
pattern recognition
face recognition
sparse representation method
single sample per person(SSPP)