WSformer: Wavelet-Based Sparse Transformer for Blind Image Restoration.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Blind Image Restoration (BIR) is crucial but challenging due to unknown degradations.
• Transformer models struggle with BIR, often incorporating irrelevant information via attention mechanisms.
• Existing sparse transformers lack flexibility due to rigid sparsification strategies.

Purpose of the Study:

• To propose WSformer, a Wavelet-based Sparse transformer for improved BIR.
• To address the limitations of existing sparse transformers in handling complex degradations.
• To achieve a balance between restoration performance and computational efficiency.

Main Methods:

• Introduced a Sparse Reciprocal Multi-head Self-Attention (SR-MSA) mechanism for adaptive information selection and reduced complexity.
• Developed a Recalibrated Feed-Forward Network (RFFN) to effectively fuse local and global information for robust feature learning.
• Integrated wavelet transform and a U-shaped architecture to optimize performance and inference time.

Main Results:

• WSformer demonstrated superior effectiveness across multiple BIR tasks.
• Achieved significant improvements in both quantitative metrics and visual quality of restored images.
• The proposed model successfully balances high performance with efficient inference.

Conclusions:

• WSformer offers a flexible and effective solution for blind image restoration.
• The novel attention and feed-forward mechanisms enhance feature learning and information selection.
• Wavelet integration and U-shaped architecture contribute to an optimal performance-efficiency trade-off.