AMSA-Net: attention-based multi-scale feature aggregation network for single image dehazing.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Deep learning enhances single-image dehazing.
• Existing methods struggle with variable haze density and spatial distribution, limiting performance.

Purpose of the Study:

• To propose an attention-based multi-scale feature aggregation network (AMSA-Net) for improved single-image dehazing.
• To address limitations in current methods regarding haze density and spatial distribution.

Main Methods:

• Developed AMSA-Net, an encoder-decoder architecture utilizing multi-scale hybrid attention feature aggregation modules (MSHA-FAM).
• MSHA-FAM incorporates scale-aware coordinate residual modules (SCRM) for haze density/spatial capture and multi-scale feature refinement residual modules (MSFRRM) for feature enhancement.
• SCRM uses improved coordinate attention, while MSFRRM employs an improved pixel attention mechanism.

Main Results:

• AMSA-Net demonstrated superior dehazing quality compared to existing methods in experimental evaluations.
• Ablation studies confirmed the effectiveness of the proposed SCRM and MSFRRM modules within AMSA-Net.

Conclusions:

• AMSA-Net achieves high-quality single-image dehazing by effectively considering haze characteristics.
• The proposed network provides high-quality outputs suitable for subsequent computer vision applications.