DCAF-Net: Density-Conditioned Attention Fusion Network for Single-Image Dehazing

Area of Science:

• Computer Vision
• Image Processing

Background:

• Single-image dehazing is crucial for visual sensor preprocessing.
• Current deep learning methods struggle with non-uniform haze, illumination, and texture variations, limiting real-world generalization.
• Scarcity of real paired hazy/clear data hinders training.

Purpose of the Study:

• To propose a novel deep learning network, DCAF-Net, for robust single-image dehazing.
• To enhance the generalization ability of dehazing models in real-world scenarios.
• To provide high-quality image preprocessing for intelligent perception systems.

Main Methods:

• Developed a haze-density conditional attention fusion network (DCAF-Net).
• Introduced an adaptive haze density perception module to generate a spatial haze density guidance map.
• Integrated multi-scale feature modulation and attention fusion with conditional haze density information.
• Employed a residual dense cascaded feature enhancement module for improved feature representation.
• Utilized a joint optimization objective with Charbonnier, perceptual contrast, and structural similarity losses.

Main Results:

• DCAF-Net demonstrated competitive performance against existing methods on synthetic and real-world datasets.
• The network showed promising restoration capabilities on representative real-world hazy scenes.
• Achieved adaptive restoration for regions with varying haze degradation levels.

Conclusions:

• DCAF-Net offers an effective solution for single-image dehazing, overcoming limitations of previous methods.
• The proposed approach improves image quality for visual-sensor-based intelligent perception systems.
• The model exhibits strong generalization and restoration performance in complex real-world conditions.