Single image denoising with a feature-enhanced network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Transformer networks excel in single image denoising by capturing long-range features.
• High computational complexity is a limitation of Transformer models.
• Integrating CNNs with Transformers offers a potential solution for efficient image denoising.

Purpose of the Study:

• To propose a Feature-Enhanced Denoising Network (FEDNet) that combines CNNs and Transformers.
• To improve the efficiency and performance of single image denoising models.
• To reduce the computational complexity while maintaining high denoising quality.

Main Methods:

• Developed FEDNet by integrating CNN architectures with Transformer blocks.
• Introduced a cross-channel attention mechanism to enhance channel feature interaction.
• Incorporated Transformer blocks into minimum-scale layers to capture long-distance dependencies and reduce MACs.
• Designed a structure-preserving block for improved structural feature extraction.

Main Results:

• FEDNet demonstrated state-of-the-art denoising performance on synthetic and real-world datasets.
• The proposed model achieved low computational costs compared to existing methods.
• Experimental results validated the effectiveness of the cross-channel attention and structure-preserving blocks.

Conclusions:

• FEDNet effectively combines CNNs and Transformers for superior single image denoising.
• The network achieves high performance with significantly reduced computational complexity.
• FEDNet represents an efficient and effective solution for modern image denoising challenges.