Abstarct: Face recognition is one of common methods of identifying individuals which has many applications such as personal device login, tracking offenders, human interaction with computers and etc. The objective of face recognition is to identify a person or persons from an image using databaxse. There are many challenges for face recognition such as illumination variation, facial exxpression variations, occlusion, age variations, low number of training samples, high number of subjects and etc. In this thesis, we propose a robust method against illumination variations, facial exxpression and age variations. In this thesis, after face detection local features such as HOG, LBP and STFT are extracted. To achieve best performance, we use preprocessing and tuning feature’s parameters for each feature. Then, by concating the vectors of optimal features, a final feature vector is constructed and classifying is done by using the nearest neighborhood classifier and Manhattan's distance criteria. In order to achieve even better performance, we produce virtual samples using image blurring, sharpening and rotation for each train sample and then we add these virtual samples to training set. To evaluate the proposed method, we have used the Feret databaxse, which contains about 1,000 people images, and there is only one face image per person in the training set. It should be noted that most of the methods presented in Face Recognition have addressed one of the existing challenges and have been tested on a databaxse of fewer than 200 people. Experimental results on the Feret databaxse have a recognition rate of 96.96% using virtual samples and 95.95% without using virtual samples with combination of HOG and STFT feature on the fb series with 158 milliseconds per image of the test set.