Abstarct: Over the past two decades, face recognition (FR) has become a popular area of research in computer vision and one of the most successful application of image analysis and understanding. Face recognition is used in a variety of areas, including security, image processing, video surveillance and criminal identification. Face recognition challenges such as, occlusion, pose variation, illumination, facial exxpressions, large databaxse and a single sample per person. Most of the methods perform well, when their databaxse is small, or multiple samples per person are existed. In this thesis, the main focus is on three challenges of databaxse capability, single sample per person and speed a test image retrieval. In this research a method is proposed to extract features from facial images. By using Non-negative Matrix Factorization (NMF), the basic features of the face structure is extracted. The matrix of images is decomposed to basis matrix (W) and weight matrix (H). The basis images contain several versions of mouths, noses and other facial parts, where the various versions are in different locations or forms. The variability of a whole face is generated by combining the element of matrix W and matrix H. Hence, to recognize a facial image in the databaxse, searching is on the feature vectors of them. In this research, in order to more precisely form structural elements, a separate a basis matrix is formed for the upper and lower half of the set of facial images. Also three pre-processing methods is used to improve to quality of the images, e.g. histogram equalization, Image intensity and contrast limited adaptive histogram. The FERET databaxse contains 990 images and only one image per person. It is used to evaluate the proposed method. Experimental results show that the retrieval speed of a test image is faster than the other methods, however the recognition rate is closed to 93%.