Semantic-Driven Global-Local Fusion Transformer for Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Transformer architectures have advanced image super-resolution (SR).
• High computational costs limit current transformer SR methods to local attention.
• This restricts modeling of long-range dependencies and global structures in SR.

Purpose of the Study:

• Propose a novel SR framework, Semantic-Driven Global-Local Fusion Transformer (SGLFT).
• Enhance receptive field to capture both local textures and global context for improved SR.
• Strengthen semantic consistency in reconstructed images using high-level semantic priors.

Main Methods:

• Combine Hybrid Window Transformer (HWT) and Scalable Transformer Module (STM) for global-local feature fusion.
• Introduce a Semantic Extraction Module (SEM) to distill semantic priors from input images.
• Utilize an Adaptive Feature Fusion Semantic Integration Module (AFFSIM) for integrating semantic cues with visual features.

Main Results:

• SGLFT demonstrates effectiveness on standard SR benchmarks.
• Achieved visually faithful and structurally consistent super-resolved images.
• The proposed method overcomes limitations of local-window attention in transformer-based SR.

Conclusions:

• SGLFT offers a powerful approach for image super-resolution by integrating semantic information.
• The model effectively balances local texture detail and global context understanding.
• Future work can explore further semantic-driven enhancements for SR tasks.