一种加权的分块邻域保持嵌入人脸识别算法

摘要

在人脸识别中,传统的子空间识别算法将整幅人脸图像作为输入模式.但是,当人脸图像的光照、姿态和表情变化较大时,仅考虑图像的全局特征,识别的效果不够理想.为此,论文提出一种加权的分块邻域保持嵌入人脸识别算法.算法先对人脸图像进行分块,对分块得到的子图像利用邻域保持嵌入算法分别提取特征信息,并利用Geman-McClure函数和类标信息计算各分块子图像的权重.该算法能够增强分类效果,对表情、光照、姿势等变化具有鲁棒性,在ORL和Yale人脸数据库上对论文的提出的算法进行了验证.%In face recognition,the whole facial image is considered as the input mode in the traditional subspace recognition al?gorithm. However,under the circumstances of big changes in illuminations,poses and expressions in facial images,facial recogni?tion effects are not satisfactory if only the global features of images are considered. In this paper,a weighted sub-pattern neighbor?hood preserving embedding algorithm is proposed for face recognition. Firstly,the original whole face pattern is partitioned into a set of equally-sized non-overlapped sub-patterns. Then,the neighborhood preserving embedding algorithm is directly performed on each sub-pattern to extract its features. And the weight of each sub-pattern is calculated with the Geman-McClure function com?bined with the label information. Experimental results on ORL and Yale human face databases showed that the advocated algorithm was more robust for face recognition under the circumstances of big changes in illuminations,poses and expressions in face images.