SRFormerV2: Taking a Closer Look at Permuted Self-Attention for Image Super-Resolution

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Transformer models significantly improve image super-resolution (SR).
• Increasing window size in SR models boosts performance but raises computational costs.
• Existing methods face a trade-off between performance gains and computational overhead.

Purpose of the Study:

• Introduce SRFormer, a novel method for efficient Transformer-based image super-resolution.
• Leverage large window self-attention benefits while minimizing computational burden.
• Explore the potential of scaled Transformer models for further SR performance improvements.

Main Methods:

• Developed SRFormer, featuring permuted self-attention (PSA) for balanced channel and spatial information processing.
• Implemented PSA to achieve efficient large window self-attention.
• Scaled the SRFormer model (SRFormerV2) by increasing window size and channel numbers.

Main Results:

• SRFormer achieved 33.86dB PSNR on Urban100, surpassing SwinIR by 0.46dB.
• SRFormer demonstrated superior performance with fewer parameters and reduced computations compared to SwinIR.
• The scaled SRFormerV2 achieved state-of-the-art results, indicating significant potential.

Conclusions:

• SRFormer offers an effective and computationally efficient approach to Transformer-based image super-resolution.
• Permuted self-attention (PSA) provides a balance between performance and computational load.
• The proposed method and its scaled version (SRFormerV2) represent advancements in super-resolution model design.