GSDAT-Net: Enhancing Image Super-Resolution with Grid-Spatial Dual Attention Hybrid Transformer

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image super-resolution (SR) is crucial for recovering high-fidelity details from low-resolution images.
• Existing Transformer-based SR methods have limitations in feature representation.
• There is a need for advanced models to improve reconstruction accuracy and visual quality in SR.

Purpose of the Study:

• To introduce GSDAT-Net, a novel grid-spatial dual-attention-driven residual hybrid Transformer for image super-resolution.
• To enhance the extraction of multi-scale features and improve local and global feature representation in SR tasks.
• To demonstrate the effectiveness of GSDAT-Net compared to existing state-of-the-art methods.

Main Methods:

• Integration of Grid Attention Block (GAB) for grid-level feature refinement.
• Incorporation of an Enhanced Spatial Attention (ESA) module for spatial feature enhancement.
• Utilization of SwinV2 Transformer layers (S2TL) for efficient multi-scale feature extraction and representation.

Main Results:

• GSDAT-Net demonstrates competitive reconstruction accuracy on benchmark datasets.
• The method achieves superior visual quality in recovered high-resolution images.
• A maximum Peak Signal-to-Noise Ratio (PSNR) improvement of up to 0.15 dB was observed.

Conclusions:

• GSDAT-Net effectively addresses limitations of existing Transformer-based SR methods.
• The proposed architecture significantly improves feature representation for image super-resolution.
• GSDAT-Net offers a promising approach for achieving high-fidelity image reconstruction.