SRGAT: Single Image Super-Resolution With Graph Attention Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Single-image super-resolution (SISR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs.
• Current Convolutional Neural Network (CNN)-based SISR methods often overlook internal patch recurrence, limiting representational power.
• Exploiting patch recurrence can potentially improve texture detail recovery in SISR.

Purpose of the Study:

• To propose a novel SISR network, SRGAT, that effectively utilizes internal patch recurrence.
• To enhance the representational capacity of deep neural networks for SISR tasks.
• To improve the recovery of fine texture details in super-resolved images.

Main Methods:

• Developed a novel Super-Resolution network based on Graph Attention (SRGAT).
• Incorporated a feature mapping block with a recurrent structure to refine representations.
• Employed a parallel graph similarity branch and a content branch to exploit patch symmetry and enhance details.
• Utilized Graph Attention Networks (GAT) to model interactions between image feature patches.

Main Results:

• SRGAT effectively leverages internal patch recurrence for improved SISR.
• The graph similarity branch provides priors that enhance texture details.
• Quantitative and qualitative evaluations show favorable performance against state-of-the-art SISR methods.
• The proposed method demonstrates superior performance on five benchmark datasets.

Conclusions:

• SRGAT offers a powerful approach to single-image super-resolution by incorporating internal patch recurrence.
• Exploiting patch similarity and symmetry via graph attention networks significantly improves texture reconstruction.
• The proposed method represents a promising advancement in deep learning for image super-resolution.