Abstract
In recent years, single image dehazing models (SIDM) based on the atmospheric scattering model (ASM) have achieved remarkable results. However, ASM-based SIDM performance degrades when dehazing real-world hazy images due to the limited modeling ability of ASM, where the atmospheric light factor (ALF) and the angular scattering coefficient (ASC) are assumed to be constants for an entire image. Real-world hazy images rarely satisfy this assumption, creating a modeling mismatch that sets an upper bound on performance. In this study, a new Fully Non-Homogeneous Atmospheric Scattering Model (FNH-ASM) is proposed to better model hazy images under complex conditions where ALF and ASC are pixel-dependent. To address the difficulty of parameter estimation in FNH-ASM, two new cost-sensitive loss functions, beta-Loss and D-Loss, are developed to limit parameter bias in sensitive positions. Finally, an end-to-end CNN-based dehazing network, FNHD-Net, is developed. Experimental results demonstrate the effectiveness and superiority of FNHD-Net on both synthetic and real-world images, with performance gains being most significant in dense and heterogeneous haze scenes.
The Problem: The Limits of Uniform Haze Modeling
Most single-image dehazing methods rely on the classic Atmospheric Scattering Model (ASM). While effective for uniform fog, ASM makes a major simplifying assumption: that the Atmospheric Light Factor (ALF) and the Angular Scattering Coefficient (ASC) are constant across the entire image.
In the real world, haze is rarely uniform. Phenomena like patchy advection fog or varying shadows in a forest create non-homogeneous conditions. When a model assuming uniform haze encounters these real-world scenarios, it suffers from a "modeling mismatch," leading to two common failures:
- Haze Residue: Areas where the fog is not fully removed.
- Artifacts: Distortions or strange color shifts caused by incorrect parameter estimation.
Methods: The FNH-ASM Framework and FNHD-Net
To overcome these limitations, the authors propose a three-pronged solution:
1. FNH-ASM (The Model)
The researchers redefined the hazing equation so that every parameter—ALF, ASC, and depth—is pixel-dependent. This allows the model to mathematically represent complex, uneven haze distributions that traditional models ignore.
2. FNHD-Net (The Architecture)
They developed an end-to-end CNN based on the U-Net architecture. It features a single "Haze Encoder" feeding into three parallel decoders. Each decoder is responsible for estimating one specific map: the ALF map, the Depth map, and the ASC ($\beta$) map. These are then combined in an Estimate Haze-free Image Module (EHIM) to produce the final clean image.
3. Cost-Sensitive Losses ($\beta$-Loss and D-Loss)
A key insight of this research is the Asymmetry of Error: overestimating haze parameters (positive bias) creates much more severe visual artifacts than underestimating them. Furthermore, the relationship between ASC/depth bias and image error is non-linear—small mistakes in dense haze can ruin the output. To combat this, the authors developed exponential loss functions ($\beta$-Loss and D-Loss) that heavily penalize positive estimation biases, particularly in areas where the haze is already dense.
Experimental Evaluation
The model was evaluated using a mix of synthetic (NYU-Depth v2) and real-world datasets, including O-Haze (homogeneous) and NH-Haze (non-homogeneous).
Key Results
The proposed FNHD-Net outperformed 11 state-of-the-art methods, including DCP, DehazeNet, and FFA-Net. On the challenging NH-Haze dataset, FNHD-Net achieved a PSNR of 22.3 dB, significantly higher than the next best method, FFA-Net (19.9 dB).
As shown in the figure above, FNHD-Net demonstrates superior color preservation and significantly less haze residue in dense areas compared to traditional models like DCPDN, which often introduces unnatural color casts (such as the blue tint seen in the top-right quadrant).
Conclusion and Future Work
FNHD-Net proves that modeling the scattering coefficient as a pixel-dependent variable is the key to handling complex weather phenomena. While the model is computationally more complex due to its triple-decoder structure and relies on synthetic data for initial training, its robustness in extreme, non-homogeneous conditions represents a significant step forward for single-image dehazing. Future work may involve refining the output stage with more sophisticated attention mechanisms to further reduce noise in high-frequency image areas.
Summary updated