Multi-attention fusion transformer for single-image super-resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Transformer-based methods are increasingly used for image super-resolution (SR).
• Existing Transformer SR models have limitations in fully utilizing spatial information due to local attention mechanisms.
• Analysis using local attribution maps highlights the need for improved spatial information integration.

Purpose of the Study:

• To propose a novel Transformer model, the Multi-Attention Fusion Transformer (MAFT), for enhanced image super-resolution.
• To address the limitations of current methods in leveraging the spatial extent of input information.
• To expand the activation range of pixels during image reconstruction for better SR performance.

Main Methods:

• Introduced the Multi-Attention Fusion Transformer (MAFT) model.
• Developed Multi-attention Adaptive Integration Groups to transition between local and global attention.
• Incorporated Local Attention Aggregation and Global Attention Aggregation blocks with alternating connections.
• Expanded the network's receptive field to better utilize input information.

Main Results:

• MAFT demonstrated superior performance on benchmark datasets for image super-resolution.
• Achieved 0.09 dB gain on the Urban100 dataset for 4x SR compared to state-of-the-art methods like HAT.
• Reduced parameter count by 32.55% and FLOPs by 38.01% compared to existing methods.

Conclusions:

• The proposed MAFT effectively integrates multiple attention mechanisms for improved image super-resolution.
• MAFT enhances the utilization of input information space, leading to better reconstruction quality.
• MAFT offers a more efficient and effective solution for image super-resolution tasks.