GlobalSR: Global context network for single image super-resolution via deformable convolution attention and fast...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Vision Transformers (ViTs) excel in image super-resolution (SR) but are slow due to Multi-Head Self-Attention (MHSA).
• Convolutional Neural Network (CNN) methods often miss global context, leading to blurry details.

Purpose of the Study:

• To develop a novel architecture that combines the strengths of ViTs and CNNs for efficient and high-quality image super-resolution.
• To investigate the impact of architectural design versus specific components like MHSA on Transformer-based SR performance.

Main Methods:

• A general architecture, GlobalSR, was proposed, focusing on architectural design rather than specific MHSA components.
• Ablation studies replaced MHSA with large kernel convolutions, yielding competitive results.
• GlobalSR was instantiated as GlobalSR-DF, using Deformable Convolution Attention Block (DCAB) and Fast Fourier Convolution Domain Embedding (FCDE) for efficient global context extraction.

Main Results:

• GlobalSR architecture demonstrated a superior trade-off between image super-resolution quality and inference speed.
• GlobalSR-DF outperformed state-of-the-art CNN-based and ViT-based SISR models in accuracy-speed.
• The method produced sharp and natural-looking super-resolved images.

Conclusions:

• The general architectural design is crucial for efficient and high-performance Transformer-based SR.
• GlobalSR provides practical guidelines for lightweight SR networks leveraging global contextual information.
• GlobalSR-DF offers a promising solution for balancing SR quality and computational efficiency.