基于补集零空间与最近空间距离的人脸识别

Face recognition based on complement null-space and nearest space distance

针对人脸识别中在分类器判别时没有充分利用类间差异的问题,提出一种补集零空间(CNS)算法,并进一步提出结合CNS算法与最近空间距离的人脸识别算法——补集零空间与最近空间距离算法(CNSD)。首先,在训练样本中,对每一种类别的人脸样本,构建其子空间并计算其补集的零空间;其次,计算测试样本与所有子空间和补集零空间的距离,找到最小的子空间距离与最大的补集零空间距离对应的类别,将其判别为测试样本的类别。算法在ORL与AR人脸数据集上进行了测试,当训练样本数较小时,CNS算法与CNSD算法识别率远高于最近邻分类器(NN)算法、最近空间距离(NS)算法、最近最远空间距离(NFS)算法;训练样本数较大时,CNS算法与CNSD算法识别率也略高于NN算法、NS算法、NFS算法。实验结果表明,所提算法能充分利用图像的类间差异,提高人脸识别的成功率。

In order to solve the problem that elassifiers do not make full use of the differences between different types of face samples in face reeagnition, an effective method for face recognition was proposed, namely Complement Null-Space （CNS） algorithm; and further more, another method which combined CNS and nearest spaee Distance （CNSD） was proposed. Firstly, subspaees and eomplement null-spaces of all types of training images were eonstrueted. Secondly, the distances between the test image and all types of subspaces as well as the distances between the test image and all types of eomplement null-spaces were calculated. Finally, the test image was elassifiod into the type whieh has the minimum subspaee distance and the maximum complement null-space distance. On ORL and AR faee databases, the reeagnition rates of CNS and CNSD are much higher than those of Nearest Neighbor （NN）, Nearest Space （NS） method and Nearest-Farthest Subspaee （NFS） method when the number of training samples is small; and it is a little higher than that of NN, NS and NFS when dealing with large samples. Simulation results show that the proposed algorithm ean make full use of the differences between different types of images and has good recognition ability.