Real-world face super-resolution based on generative adversarial and face alignment networks

Facial recognition performance is significantly limited when dealing with low-resolution face images, especially in real-world scenarios, due to the lack of precise knowledge about the degradation kernel. This research aims to enhance the resolution of real-world low-resolution face images by integrating a face alignment network into a semi-cycle generative adversarial network (GAN), which is conventionally known as face super-resolution. The proposed approach leverages the powerful capabilities of GANs to alleviate the domain discrepancy between real and synthetic images by introducing dual degradation pathways (forward and backward) that work collaboratively within a cycle-consistency learning framework. Additionally, a face alignment network is embedded within the GAN framework to refine the generated images by leveraging heatmap regression, which predicts the precise locations of facial landmarks. This allows our method to enforce structural consistency and preserve fine-grained facial details, such as the eyes, nose, and mouth, in the super-resolved images. As a result, the proposed method achieves significant improvements in generating high-resolution realistic face images. The experiments were conducted on both real-world and synthetic datasets; the results demonstrated the superiority of our method over existing approaches in generating high-resolution face images with exceptional degradation kernel and naturalness. Additionally, our method achieved the highest accuracy in face recognition and detection tasks, reflecting its capability to preserve essential identity features effectively, making it particularly well-suited for applications involving downstream facial analysis.