Progressive Upsampling Generative Adversarial Network with Collaborative Attention for Single-Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Single-image super-resolution (SISR) is crucial for enhancing image quality.
• Existing SISR methods often struggle with real-world noise and degradation.
• There is a need for robust SISR models that handle complex image degradations.

Purpose of the Study:

• To develop an advanced SISR model addressing limitations of current methods.
• To improve the reconstruction of high-resolution images from low-resolution inputs.
• To introduce a novel network architecture for superior image detail enhancement.

Main Methods:

• Proposed PUGAN (Progressive Upsampling Generative Adversarial Network) with a collaborative attention mechanism.
• Utilized Residual Multiscale Blocks (RMBs) with stacked Mixed-Pooling Multiscale Structures (MPMSs) for feature extraction.
• Implemented a progressive upsampling strategy and a discriminator for balanced reconstruction and detail enhancement.

Main Results:

• PUGAN achieved competitive PSNR/SSIM/LPIPS scores on NTIRE 2020, Urban 100, and B100 datasets for ×2 and ×4 scaling factors.
• Demonstrated superior qualitative and quantitative performance compared to state-of-the-art SISR methods.
• Showcased potential for pathological image super-resolution applications.

Conclusions:

• PUGAN offers a significant advancement in single-image super-resolution.
• The proposed architecture effectively handles image degradations and noise.
• PUGAN shows promise for real-world applications, including medical imaging.