Semi-Cycled Generative Adversarial Networks for Real-World Face Super-Resolution

Area of Science:

• Computer Vision
• Image Restoration
• Generative Adversarial Networks

Background:

• Real-world face super-resolution (SR) is challenging due to ill-posed image restoration.
• Fully-cycled Cycle-GANs produce artifacts in real-world SR due to domain gaps.
• Degradation branches in Cycle-GANs impact performance with synthetic low-resolution (LR) images.

Purpose of the Study:

• To enhance real-world face SR performance using Generative Adversarial Networks (GANs).
• To mitigate artifacts caused by domain gaps between real and synthetic LR face images.
• To achieve accurate and robust face SR.

Main Methods:

• Introduced Semi-Cycled Generative Adversarial Networks (SCGAN).
• Established two independent degradation branches for forward and backward cycle-consistent reconstruction.
• Utilized a shared restoration branch regularized by both processes.

Main Results:

• SCGAN effectively alleviates adverse effects from domain gaps.
• Achieved accurate and robust performance in real-world face SR.
• Outperformed state-of-the-art methods in recovering face structures and details.

Conclusions:

• SCGAN offers superior performance for real-world face SR tasks.
• The proposed architecture improves artifact reduction and detail recovery.
• The method demonstrates significant improvements on both synthetic and real-world datasets.