Direct Unsupervised Super-Resolution Using Generative Adversarial Network (DUS-GAN) for Real-World Data.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Supervised deep learning models for Single Image Super-Resolution (SISR) require paired low-resolution (LR) and high-resolution (HR) images.
• Training with known degradations causes a domain shift when applied to real-world images with unknown degradations.

Purpose of the Study:

• To propose an unsupervised approach for SISR that addresses the domain shift problem.
• To enhance the perceptual quality of super-resolved images using a novel loss function.

Main Methods:

• Developed a Generative Adversarial Network (GAN)-based unsupervised SISR method named DUS-GAN.
• Introduced a Mean Opinion Score (MOS) loss to improve human perception-based quality.
• Validated the approach on diverse reference-based and no-reference datasets.

Main Results:

• DUS-GAN demonstrates improved performance over existing unsupervised SR methods.
• Achieved superior subjective and quantitative evaluations using metrics like LPIPS, PI-RMSE, NIQE, BRISQUE, and PIQE.
• The method effectively handles real-world image degradations without prior estimation.

Conclusions:

• The proposed DUS-GAN offers a robust unsupervised solution for SISR.
• The integration of MOS loss significantly boosts the perceptual quality of super-resolved images.
• The open-source implementation facilitates reproducible research in unsupervised image super-resolution.