摘要
张量算法克服了传统向量算法的维数灾难和小样本问题,在人脸识别中取得了较好的效果。尽管如此,现有张量算法容易导致邻近类别在低维空间中重叠,为此,提出了一种加权判别局部多线性嵌入算法。该算法设计了一种自适应加权的判别准则,结合类别信息,保持了同类人脸图像之间的局部流形结构,同时利用高斯基函数作为加权函数,根据人脸图像对其他类别的影响,自适应产生权重系数,加大了类间样本的区分度。此外,该算法采用张量形式表示图像数据,保留了图像的结构,继承了张量算法的优点,并且有效地解决了out-of-sample问题。人脸识别实验表明,对于光照,姿态和表情的变化,该算法具有较好的稳健性和较高的识别率。
In order to overcome the curse of dimensionality and small sample size problem,a large number of tensor algorithms are proposed and better performance is achieved in face recognition.However,the neighboring classes overlap easily in low dimensional space for existing tensor algorithms.Therefore,this paper proposes a weighted discriminative locally multi-linear embedding algorithm.Because the algorithm considers a face image as a high-order tensor,it contains the structure of the image,avoids the curse of dimensionality and relieves the sample size problem.Moreover the algorithm preserves the local manifold structure within the same class,and increases the separability between different classes using Gaussian Basis Function as the weighted function.The algorithm also solves the out-of-sample problem effectively.Face recognition experiments demonstrate that the algorithm proposed in this paper is robust for the variation of illumination,facial expression and poses,and achieves better performance compared with many popular face recognition algorithms.
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2011年第10期2248-2255,共8页
Chinese Journal of Scientific Instrument
基金
四川省科技厅支持项目(2010JQ0032
2011JY0077)资助
关键词
流形学习
判别分析
张量表示
高斯基函数
人脸识别
manifold learning
discriminative analysis
tensor representation
Gaussian basis function
face recognition
