Abstarct: Face super-resolution, also known as face hallucination, refers to generating high-resolution face images from low-resolution face images. Face hallucination is a specific domain of the zooming problem. Most advanced methods do not pay much attention to restoring the original structure of low-resolution images. Therefore, in this research, an artificial intelligence model baxsed on Generative Adversarial Networks is proposed to focus on high-frequency details to retrieve better global and local information in super-resolving face images. One of the most critical topics in preserving a low-resolution image structure is the proper use of high-frequency details and edges of the input image. For this reason, some preprocessing, such as a Local Binary Pattern and Unsharp Masking, have been used to enhance the edges of the low-resolution images and global and local features. In addition, in the proposed generator network, the Edge block, the Self-attention block, and the Residual block have been used in a residual way to increase network performance by preserving the structure and using global and local features. The evaluations show that the proposed method has been able to preserve the structure of the face images better than the existing face super-resolution methods. Another issue in any face super-resolution method is the generating high-quality images. To a large extent, this depends on appropriate loss functions. In this thesis, five different loss functions have been used to evaluate the generated images and train the generator network. A new loss function called Frechet VGG Distance (FVD) has been proposed to generate more realistic and higher quality face images baxsed on extracted features from the VGG-19 and Frechet distance, which minimizes the distance between the features of ground truth and generated face images. The edge loss function is also used to reduce the distance between the high-frequency details of the generated image and the ground truth image. Illumination loss, mean squared error, and discriminator loss also compares the generated image and the ground truth in terms of brightness distance, pixel difference, and the difference in feature level. The ultimate goal of face super-resolution is to increase face recognition accuracy. Many methods have been proposed to super-resolving the face images, but they have improper performance in face recognition accuracy. The main reason is the lack of preserving the structure of the face image and generating fake details that are only visually compatible with other components of the generated face. The results show the performance of the proposed method in increasing the accuracy of various face recognition methods. The results also show the proper performance of the proposed method in low-resolution images in the real world.