Abstarct: Face recognition is one of the most common biometric methods of authentication and many research has been done to improve the performance of face recognition systems. This biometric method is more popular than other factors such as iris, fingerprint, and voice, due to the ease of use of the face and also no need for human interaction. However, there are many challenges that researchers are struggling with. One of these challenges is the large size of the datasets, which as the number of images increases, due to the inability to extract discriminative features and also the increase in the number of comparisons, the usual methods fail and system performance decreases. The purpose of this article is to provide a robust face recognition method that performs well in large­scale datasets. The proposed method uses a simple and innovative clustering approach to hierarchically divide images into smaller clusters. Then, in each cluster, which the number of members is far less than the number of images in the entire dataset, face recognition is performed. Also, in the recognition phase, a combination of two Non­negative Matrix Factorization (NMF) features and Fast Retina Keypoint (FREAK) descxriptors is used to be able to simultaneously use local and global characteristics of the image. Some experiments were performed to evaluate the proposed method on the large­scale dataset FERET, as well as an extended image set that is collected from smaller standard datasets together. The proposed method achieved the accuracy of 98.36% and 94.54% on dataset FERET and the extended image set, respectively, which is about 2% better than the previous best work. The results of experiments show that the performance of the clustering approach is better than the conventional clustering methods. Also, it indicates the proposed method outperforms previous works in the face recognition task.