面向酉子空间的二维判别保局投影的人脸识别

Unitary-subspace based 2D discriminant locality preserving projection for face recognition

保局投影算法(LPP)在人脸识别中具有较好的识别性能,但它是一种非监督学习,并且在具体实现时需要把图像转换为向量,破坏了图像的像素结构,这显然不利于模式识别。针对这些问题,提出基于酉子空间的二维判别保局算法,不仅在判别保局算法的基础上增加了类别信息,而且直接在灰度矩阵上进行水平和垂直方向上的二维保局投影。该方法构造酉空间上的复向量后再运用线性判别分析提取特征。在ORL、Yale和XJTU人脸库中验证了算法的正确性和有效性,其识别率比传统的2DLDA和2DLPP等方法提高4～5个百分点。

LPP has good performance in face recognition,but it is an unsupervised algorithm and has to convert the matrix to a vector which will destroy the image pixel structure,both of which are not beneficial to pattern recognition.To address these problems,proposed an algorithm based on unitary-subspace 2D discriminant locality preserving projection.Based on LPP,the proposed algorithm took the class information into account,performed also 2DLPP in the horizontal and vertical direction to obtain two feature matrixes,and then combined them to form a complex matrix in the unitary subspace,and lastly extracted the feature according to the linear discriminant analysis.Experiments based on ORL,Yale and XJTU face database demonstrate the correctness and effectiveness of the new algorithm,and the recognition rate increased by approximately 5% compared with traditional methods,such as 2DLDA,2DLPP.