Abstarct: Face recognition system is one of the important research areas that is used to identify the identity of people. This system is used in various fields such as security, image processing, video surveillance. With the advancement of digital cameras, the development of facial recognition system has increased. The face recognition system faces challenges such as covering part of the face, head angle, and lighting changes. Each of the existing methods in this field have investigated and solved one or more challenges. One of the drawbacks of most existing methods is that a person's face must be in front of the camera for the system to work well. However, due to the fact that closed circuit cameras are usually installed in high positions in open environments, they cannot take pictures of the entire face of a person, but the pictures are taken from different angles and most of the facial areas are not visible. Therefore, changing the angle of the person's head in front of the camera has caused the performance of existing facial recognition systems to decrease significantly. The main focus of this thesis is on the challenge of positioning a person's head angle in front of the camera in video images. Just as a human being can identify that person by receiving images of a person at different angles, we also want to act in a similar way and by receiving several angled images of a person and extracting features from them, we estimate the features corresponding to the image in front of that person. so that facial recognition methods can perform identification with good accuracy. Therefore, in the presented method, the basic features of a person's face structure are extracted by the non-negative matrix decomposition (NMF) method. Then the feature vectors extracted from the angular images are combined with each other using the weighted averaging method. Therefore, a combined feature vector is obtained for each person. The optimal weight vector that is used to combine the feature vectors of the person's angular images is calculated by the genetic algorithm. This vector is the same size as the feature vector. Then, with the help of GAN neural networks, the feature vector of the front state of the person's face is estimated. To search for a face in the databaxse, the search is performed on the feature vectors of the face images in the databaxse. The proposed method has been evaluated using FERET databaxse images. The accuracy of identifying angular images of this databaxse with the latest face recognition method for images up to an angle of 15 degrees is equal to 23% and images up to an angle of 40 degrees is equal to 9%; that the accuracy of identification with the proposed method has increased to 80% for images up to an angle of 15 degrees and to 63% for images up to an angle of 40 degrees.