DSwinIR: Rethinking Window-Based Attention for Image Restoration.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning models, especially transformers with window-based self-attention, dominate image restoration.
• Existing methods suffer from rigid window partitioning, limiting feature interaction and receptive fields.

Purpose of the Study:

• To introduce a novel attention mechanism, Deformable Sliding Window (DSwin) Attention, for improved image restoration.
• To address limitations of fixed window schemes in transformer-based image restoration models.

Main Methods:

• Propose the Deformable Sliding Window Transformer for Image Restoration (DSwinIR).
• Implement a token-centric sliding window paradigm to mitigate boundary artifacts.
• Incorporate content-aware deformable sampling for adaptive receptive field shaping.

Main Results:

• DSwinIR achieves state-of-the-art performance on multiple image restoration benchmarks.
• Outperforms GridFormer by 0.53 dB (3-task) and 0.87 dB (5-task) in all-in-one image restoration.
• Demonstrates superior feature interaction and receptive field adaptability.

Conclusions:

• DSwinIR offers a more adaptive and flexible attention mechanism for image restoration.
• The proposed method effectively overcomes limitations of traditional window-based transformers.
• Code and models are publicly available for further research and application.