ICAFormer: An Image Dehazing Transformer Based on Interactive Channel Attention

Area of Science:

• Computer Vision
• Image Processing

Background:

• Traditional dehazing algorithms struggle with global feature association and local detail preservation.
• Single image dehazing is crucial for recovering clear scenes from hazy inputs.

Purpose of the Study:

• To propose a novel Transformer-based dehazing model.
• To enhance global feature association and local detail preservation in image dehazing.

Main Methods:

• A U-shaped encoder-decoder architecture with interactive channel attention.
• Multi-scale feature pyramid for multi-dimensional information extraction.
• Improved cross-channel attention for varying haze densities.

Main Results:

• Achieved PSNR gains of 0.53 dB (indoor) and 1.64 dB (outdoor).
• Demonstrated SSIM improvements of 1.4% (indoor) and 1.7% (outdoor).
• Successfully restored fine structural details and maintained color fidelity in dense haze.

Conclusions:

• The proposed Transformer-based model significantly outperforms existing methods.
• The interactive channel attention mechanism effectively fuses global and local features.
• The approach enhances both perceptual quality and quantitative accuracy in image dehazing.