Parallel attention recursive generalization transformer for image super-resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Transformer architectures show promise in image super-resolution (SR).
• Existing Transformer models struggle with local feature modeling and fine detail restoration in high-resolution (HR) images.

Purpose of the Study:

• To propose a novel SR model, Parallel Attention Recursive Generalization Transformer (PARGT), to address limitations in current Transformer-based SR methods.
• To improve the reconstruction of fine-grained details and enhance feature representation capabilities in image SR.

Main Methods:

• Introduced the Parallel Local Self-attention (PL-SA) module, combining Shift Window Pixel Attention Module (SWPAM) and Channel-Spatial Shuffle Attention Module (CSSAM).
• Developed a Spatial Fusion Convolution Feed-forward Network (SFCFFN) for multi-scale information fusion.
• Incorporated Stationary Wavelet Transform (SWT) to optimize high-frequency detail reconstruction.

Main Results:

• PARGT effectively captures fine-grained interactions between local image features.
• The model achieves clearer and more coherent generated details compared to existing methods.
• Experimental results on benchmark datasets demonstrate PARGT's superiority over state-of-the-art SR models.

Conclusions:

• The proposed PARGT model significantly enhances image super-resolution performance.
• Combining parallel attention mechanisms with multi-scale feed-forward networks is effective for SR tasks.
• PARGT offers improved restoration of fine details and feature representation for high-resolution image generation.