HAT: Hybrid Attention Transformer for Image Restoration

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Transformer networks show promise in image restoration but have limited spatial information utilization.
• Existing Transformer models may not fully exploit their potential for tasks like super-resolution and denoising.

Purpose of the Study:

• To propose a novel Hybrid Attention Transformer (HAT) to improve image restoration by maximizing input information usage.
• To enhance feature interaction across windows for better global context aggregation.

Main Methods:

• The Hybrid Attention Transformer (HAT) combines channel attention and window-based self-attention mechanisms.
• An overlapping cross-attention module is introduced to improve cross-window information aggregation.
• A same-task pre-training strategy is employed during the training phase.

Main Results:

• Extensive experiments validate the effectiveness of the proposed HAT modules.
• Scaling up the model significantly improves performance on super-resolution (SR) tasks.
• HAT demonstrates state-of-the-art quantitative and qualitative results across various image restoration applications.

Conclusions:

• The proposed Hybrid Attention Transformer (HAT) effectively addresses the limitations of existing Transformer-based image restoration methods.
• HAT achieves superior performance in super-resolution, denoising, and artifact reduction.
• The model's architecture and pre-training strategy unlock greater potential for Transformer networks in image restoration.