EWT: Efficient Wavelet-Transformer for single image denoising.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Transformer-based methods excel at image denoising but are computationally expensive.
• High computational cost and memory footprint limit their practical application.

Purpose of the Study:

• To develop a resource-efficient Transformer-based image denoising method.
• To maintain high denoising performance while reducing computational demands.

Main Methods:

• Proposed the Efficient Wavelet Transformer (EWT).
• Incorporated a Frequency-domain Conversion Pipeline (FCP) for resolution reduction.
• Utilized a Multi-level Feature Aggregation Module (MFAM) with a Dual-stream Feature Extraction Block (DFEB) for hierarchical feature extraction.

Main Results:

• Achieved over 80% faster processing speed compared to original Transformers.
• Reduced GPU memory usage by more than 60%.
• Denoising performance is on par with state-of-the-art methods.

Conclusions:

• EWT significantly improves the efficiency of Transformer-based image denoising.
• Offers a balanced approach between performance and resource consumption.
• Provides a practical solution for efficient high-performance image denoising.