Abstract
Model-driven single image dehazing is widely studied for its extensive applications, yet it faces two inherent challenges: the ambiguity between object radiance and haze, and noise amplification in sky regions. This paper proposes a Dark Direct Attenuation Prior (DDAP) to address the former. A novel haze line averaging technique is introduced to reduce morphological artifacts caused by DDAP, allowing for a weighted guided image filter with a smaller radius to preserve fine structures. To tackle noise amplification, a multi-scale dehazing algorithm utilizes Laplacian and Gaussian pyramids to decompose the hazy image. Different restoration and noise reduction approaches are applied at various levels before collapsing the pyramid to restore the final haze-free image. Experimental results show that the proposed algorithm outperforms state-of-the-art methods, effectively preventing noise amplification in sky regions.
Introduction
Single image dehazing is a fundamentally ill-posed problem. Following Koschmieder’s law, a hazy image is modeled as a combination of true scene radiance and atmospheric light. Since multiple unknowns exist for each pixel in a single input, researchers rely on "priors"—statistical assumptions about natural images. However, existing methods often struggle with two issues:
- Ambiguity: Distinguishing between bright objects and haze is difficult. Priors like the Dark Channel Prior (DCP) often fail in sky or bright regions, causing halo artifacts.
- Noise Amplification: In heavy haze, the transmission signal is weak. Restoring these regions involves dividing by small values, which significantly amplifies sensor noise, particularly in the sky.
Methodology
The proposed approach utilizes three core components:
- Dark Direct Attenuation Prior (DDAP): Unlike DCP, which assumes dark pixels in the final image, DDAP assumes the direct attenuation (light from the object reaching the camera) is dark. This constraint is more robust in sky regions.
- Haze Line Averaging: To mitigate blocky artifacts, the authors group pixels with similar intrinsic colors but different depths. Non-local averaging across these "haze lines" refines the transmission map without requiring large-radius filters that blur fine details.
- Multi-Scale Pyramid Decomposition: The image is split into a Gaussian pyramid (low-frequency) and a Laplacian pyramid (high-frequency). Dehazing is applied to low-pass components where haze is dominant, while noise reduction is applied to high-pass components.
Experimental Results
The method was evaluated on real-world datasets, composite benchmarks (including RESIDE), and vehicle detection tasks. Quantitatively, the proposed method achieved a Dehazing Quality Index (DHQI) of 60.97, surpassing FFA-Net (55.33) and MSBDN (54.32).
Qualitatively, the refined transmission map significantly reduces artifacts compared to the initial estimation. The multi-scale approach ensures that the sky remains clean and grain-free while maintaining sharp edges on foreground objects. Furthermore, using this method as a pre-processor improved Faster R-CNN vehicle detection accuracy on hazy footage.
Conclusion
This paper presents a robust model-driven dehazing framework that addresses ambiguity and noise amplification. By combining the DDAP with pyramid-based multi-scale processing, the algorithm achieves superior photo-realism and detail preservation. Future work will focus on automating the selection of the dehazing strength parameter ($\eta$) to enable fully autonomous real-time applications.
Summary updated