Multiple Cycle-in-Cycle Generative Adversarial Networks for Unsupervised Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Supervised learning for single image super-resolution (SR) typically requires paired low-resolution (LR) and high-resolution (HR) images.
• Traditional methods struggle when the down-sampling process is unknown or the LR input is degraded by noise and blurring.
• Unsupervised image-to-image translation techniques offer potential for learning from unpaired data.

Purpose of the Study:

• To develop an unsupervised method for single image super-resolution that addresses unknown down-sampling and image degradations.
• To propose a novel network architecture capable of handling real-world degraded LR images.

Main Methods:

• A multiple Cycle-in-Cycle network structure is proposed, utilizing generative adversarial networks (GANs).
• The network comprises sequential cycles: the first cleans noisy and blurry LR images, and subsequent cycles perform super-resolution (e.g., ×2, ×4, ×8).
• All network modules are trained end-to-end for seamless integration.

Main Results:

• The proposed method demonstrates effective handling of unknown down-sampling and image degradations.
• Quantitative and qualitative evaluations show performance comparable to state-of-the-art supervised SR models.
• The unsupervised approach successfully generates high-resolution images from degraded inputs.

Conclusions:

• The Cycle-in-Cycle network provides a robust solution for single image super-resolution in challenging, real-world scenarios.
• This unsupervised method overcomes the limitations of paired data requirements in traditional SR.
• The approach offers a viable alternative for image enhancement when dealing with unknown degradations.