There are various face recognition techniques in literature, which are faced with challenges such as occlusion, pose variation, illumination, and facial expressions. Existing methods often perform well when their database is small, or multiple samples per person exist. However, face recognition methods with just one reference sample per person may not work well, especially on a large database. To address this problem, this paper proposes a scheme to extract features from facial images. Using Non-negative Matrix Factorization (NMF), basic features are extracted from the face structure. The matrix of images is decomposed into basis matrix (W) and weight matrix (H). The basis matrix contains several versions of mouths, noses and other facial parts, where the various versions are in different locations or forms. Hence, to recognize a facial image in the database, searching is done on the weight matrices feature set. In this research, to more precisely form the structural elements, a separate basis matrix is constructed for the upper and lower parts of the facial images from the database. Also the images are enhanced using pre-processing techniques including histogram equalization, image intensity, and contrast limited adaptive histogram. The FERET database with 990 single images per person was used to evaluate the proposed method. Experimental results show that the proposed method can achieve a recognition rate close to 93%.

文献中存在多种面部识别技术，这些技术面临着诸如遮挡，姿势变化，照明和面部表情等挑战。现有方法在其数据库较小或每人存在多个样本时通常表现良好。但是，人均只有一个参考样本的人脸识别方法可能效果不佳，尤其是在大型数据库上。为了解决这个问题，本文提出了一种从面部图像中提取特征的方案。使用非负矩阵分解（NMF），从面部结构中提取基本特征。图像矩阵被分解为基本矩阵（W）和权重矩阵（H）。基本矩阵包含口，鼻和其他面部部位的几种版本，其中各种版本位于不同的位置或形式。因此，为了识别数据库中的面部图像，需要对权重矩阵功能集进行搜索。在这项研究中，为了更精确地形成结构元素，从数据库中为面部图像的上部和下部构造了单独的基础矩阵。同样，使用包括直方图均衡化，图像强度和对比度受限的自适应直方图在内的预处理技术来增强图像。使用FERET数据库（每人990张单幅图像）来评估所提出的方法。实验结果表明，该方法可以达到接近93％的识别率。同样，使用包括直方图均衡，图像强度和对比度受限的自适应直方图在内的预处理技术来增强图像。使用FERET数据库（每人990张单幅图像）来评估所提出的方法。实验结果表明，该方法可以达到接近93％的识别率。同样，使用包括直方图均衡，图像强度和对比度受限的自适应直方图在内的预处理技术来增强图像。使用FERET数据库（每人990张单幅图像）来评估所提出的方法。实验结果表明，该方法可以达到接近93％的识别率。