SVD用于人脸识别存在的问题及解决方法  

The Problem Existed in Face Recognition Using SVD and Its Solution

作　　者：高全学[1] 梁彦[2] 潘泉[2] 陈玉春[2] 张洪才[2] 

机构地区：[1]西安电子科技大学通信工程学院,西安710071 [2]西北工业大学自动化学院,西安710072

出　　处：《中国图象图形学报》2006年第12期1784-1791,共8页Journal of Image and Graphics

基　　金：国家自然科学基金项目(60372085;60172037);航空科学基金项目(03D53032)

摘　　要：通过对人脸图像奇异值的分析,证实了图像奇异值是图像在特定基空间分解得到的,这个基空间是由图像本身决定的。进一步研究发现,导致基于奇异值向量人脸识别算法识别率低的根本原因是:不同人脸图像对应的奇异值向量所在的基空间不一致、奇异值向量与人脸图像之间并不存在一一对应关系、奇异值向量具有不可分割性。最后提出了类估计基空间识别算法。在ORL、ORL-NWPU1以及ORL-NWPU2数据库进行仿真,实验结果证实了分析和所提算法的正确性。The singular value vector of face image is analyzed and its nature is revealed, i.e. , the singular values of face image are obtained by decomposing the face in the special basis space that is determined by the face image. Based on this result, the intrinsic reasons that singular value vector-based recognition methods have the low recognition accauracy are follows： singular value vectors of arbitrary two face images have not the equivalent basis space in general; there is no oneto-one correspondence between singular value vector and face image; singular value vectors are not separated. Finally, a novel face recognition method coined class estimated basis space method is presented. ORL, ORL-NWPU1 and ORL- NWPU2 database are used to test, and the experimental results show the correctness of analysis and the proposed method.

关 键 词：奇异值向量 人脸识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]