Face Super-Resolution in the field of image processing has become an attractive and practical research field, which plays an important role in other fields such as identity recognition. In recent years, various deep learning methods have been proposed to increase the resolution of images, and Many of these methods are developed with the assumption of constant degradation. Therefore, these approaches have problems in real-world applications of face images. One additional factor contributing to the limitations of these methods is their reliance on prior geometric and reference facial information for image restoration. However, such comprehensive information is often lacking in low-quality input images. Furthermore, the availability of high-quality images is not always guaranteed. In this research, we try to overcome these shortcomings to some extent through the structure of pre-trained generative network (StyleGAN2). The information obtained from this network is combined with the features extracted from the input image in order to create a suitable balance between the degree of fidelity to the preservation of face identity and the degree of realness of the face. Also, for the dynamic performance of the network to select from both sides of the obtained information, a section for determining the amount of degradation has been considered, which uses the learning strategy of representation of degradation in an unsupervised manner. The eva‎luation of the solution used in this model on the face images of the real world as well as other databases has shown an increase in efficiency and quality, especially in images with a high level of degradation.

محمدعلي خسروي فرد , محمدرضا حيدرپور

شادرخ سماوي , نادر كريمي