Physics-Aware Dehazing Algorithm

Updated 20 January 2026

Physics-aware dehazing is a method that integrates the atmospheric scattering model into the image restoration process to recover clear scene radiance from hazy inputs.
It employs branch-based fusion and physics-driven loss functions to enhance image quality metrics like PSNR and SSIM under varying haze conditions.
The approach improves interpretability and robustness, making it suitable for real-time dehazing in edge devices, remote sensing, and autonomous systems.

A physics-aware dehazing algorithm is a computational method for recovering underlying scene radiance from images degraded by atmospheric haze, rooted in the explicit modeling of the physical processes underlying haze formation and propagation. Such algorithms are characterized by direct incorporation of the atmospheric scattering model into the network structure, loss function, or data pipeline, often providing interpretability, robustness to real-world complexities, and improved generalization over purely data-driven or heuristic approaches.

1. Atmospheric Scattering Model Foundations

Physics-aware dehazing algorithms universally rely on the classical atmospheric scattering model (ASM), formalized as

$I(x) = J(x)\,t(x) + A\,(1 - t(x))$

where:

$I(x)$ is the observed hazy image at pixel $x$ ,
$J(x)$ is the latent, haze-free scene radiance,
$t(x)$ is the medium transmission map, typically modeled as $t(x) = e^{-\beta d(x)}$ with scene depth $d(x)$ and scattering coefficient $\beta$ ,
$A$ is the global atmospheric light, commonly estimated as a constant RGB vector.

Dehazing is thus equivalent to estimating both $t(x)$ and $A$ from a single observed image, then solving for $J(x)$ via the closed-form inversion: $\hat{J}(x) = \frac{I(x) - A\,(1 - \hat{t}(x))}{\hat{t}(x)}$ Certain recent methods relax the constant-airlight assumption and estimate airlight as a spatial field to better handle sky-dominated and non-uniform scenes (Zhang et al., 2018).

2. Model Architectures and Fusion of Physical Priors

Modern physics-aware dehazing architectures integrate the scattering model at multiple structural levels:

Branch-based Fusion Models: CL2S constructs multiple parallel operator branches for dehazing, each corresponding to an elementary function: multiplicative, additive, exponential, and (notably) sinusoidal. The atmospheric-model branch enforces strict scattering-based correction, and the outputs of all branches are fused by learned pixelwise attention weights. The sine operator is empirically found to enhance the model’s ability to capture complex haze distributions that deviate from log/exp monotonicity, yielding superior edge and color fidelity (Rohn, 2024).
Frequency-Aware Designs: FALCON introduces a continuous density mask (CDM) to provide a spatial haze-density prior, derived via a differentiable relaxation of the dark channel prior. The architecture employs a frequency adjoint link (FAL) bottleneck to capture global haze patterns with efficiency, merging local spatial convolutions and global Fourier transforms at feature level (Kim et al., 2024).
Feature Space Physics-Injection: Networks like C²PNet and FriendNet embed the physical scattering equation directly in feature space via specialized attention or dual-branch units, forcing intermediate representations to structurally obey physics-derived forms and facilitating interpretability within deep networks (Zheng et al., 2023, Fan et al., 2024).
Parameter-Efficient Alternatives: GUSL-Dehaze implements the ASM within a lightweight, non-deep-learning framework using a green U-shaped architecture. Initial transmission and airlight estimates are computed via an improved DCP and Random Forest regressor, followed by feature engineering and multiscale regression without backpropagation, trading computation for interpretability and low model complexity (Movaheddrad et al., 23 Oct 2025).
Transformers and High-Resolution Approaches: Ultra-high-definition dehazing is accomplished in anDehazeFormer by dynamically extracting ASM-relevant features (e.g., refined dark-channel estimates, learnable scattering coefficients) and guiding attention mechanisms and adaptive normalization based on the estimated haze profile for both computational and convergence gains over standard transformers (Wang et al., 20 May 2025).

3. Enhancements via Elementary Function Fusion

Physics-aware models often employ fusion of diverse nonlinear elementary functions to capture multi-modal haze distributions:

$\begin{align*} J_0(x) &= I(x) - \hat{A}_0(1 - \hat{T}_0(x))\ J_1(x) &= I(x) \times R_1(x)\quad\text{(multiplicative)}\ J_2(x) &= I(x) + R_2(x)\quad\text{(additive)}\ J_3(x) &= [I(x)]^{R_3(x)}\quad\text{(exponential)}\ J_4(x) &= \sin(I(x) + R_4(x))\quad\text{(sinusoidal)} \end{align*}$

The composite dehazed output is: $J_f(x) = \sum_{i=0}^{4}W_i(x)\,J_i(x)$ where $W_i(x)$ are softmax-normalized pixelwise attention weights.

Ablation confirms that the trigonometric component enables superior modeling of spatially inhomogeneous and periodic haze patterns, surpassing classical log-transform variants and yielding measurable PSNR/SSIM gains (Rohn, 2024).

4. Physics-Aware Loss Functions and Training Strategies

Robust training of physics-aware dehazing algorithms is achieved by integrating physically motivated losses:

Reconstruction Fidelity: Standard pixelwise $L_2$ loss is typically employed, targeting accurate recovery of haze-free ground truth.
Density Map Constraints: Auxiliary losses penalize discrepancies between predicted and ground-truth haze-density maps—often derived from DCP or novel differentiable pooling operations—to directly supervise the network’s respect for local haze structure (Kim et al., 2024).
Contrastive and Perceptual Regularization: Advanced regularization frameworks, including curricular contrastive losses, use self-consensual negatives and curriculum learning to constrain the solution space to physically plausible reconstructions with enhanced interpretability in learned feature space (Zheng et al., 2023).
Physical Consistency in Diffusion Models: For diffusion-based dehazing, RPD-Diff and DehazeDDPM integrate the ASM into both the forward and reverse Markov chains, e.g., by defining intermediate state targets via physical interpolation and enforcing consistency of the generative process with physical haze priors (Zhang et al., 23 Aug 2025, Yu et al., 2023).

5. Quantitative Outcomes and Empirical Validation

Physics-aware algorithms consistently exhibit superior performance on standard dehazing metrics:

Method	RESIDE PSNR (dB)	SSIM	O-HAZE PSNR	HazeRD CIEDE2000	Notable Features
DM2F	34.29–34.99	0.98	24.41–25.19	~12.9–11.6	Multi-branch, log operator
CL2S	35.36	0.98	24.58	11.42	Adds sine, ablation-strong
FALCON	NH-Haze2: 22.41	0.84	—	—	182 FPS, CDM + freq domain
GUSL-Dehaze	39.14–36.83	0.991	—	—	Non-DL, interpretability
RPD-Diff	Dense-HAZE: 20.06	0.63	—	FID 126.3	ASM in diffusion, adaptive

Ablation studies consistently corroborate the necessity of physics-driven modules. For instance, removing CL2S's sine branch dropped PSNR by 0.37 dB, and eliminating RPD-Diff's physics-guided targeting reduced Dense-HAZE PSNR by 23% relative to the full model. Such findings underscore that physical modeling is not merely auxiliary but fundamental in dense and non-uniform haze scenarios (Rohn, 2024, Zhang et al., 23 Aug 2025).

6. Generalization, Efficiency, and Limitations

Physics-aware dehazing is effective across synthetic and real-world datasets, demonstrating robustness to spatially non-uniform haze, strong color fidelity, and superior edge preservation. Efficient formulations (FALCON, GUSL-Dehaze) enable processing at real-time or edge-device scales, while more complex architectures (CL2S, RPD-Diff) trade some compute for SOTA quantitative benchmarks.

However, limitations persist. In scenarios with extremely dense haze or where model assumptions (e.g., homogeneous airlight) are violated, estimation of $t(x)$ , $A$ , or physically meaningful priors can be inaccurate, motivating ongoing improvements including:

Adaptive or spatially varying airlight field modeling (Zhang et al., 2018).
Enhanced unsupervised physics-based data augmentation and domain adaptation for transfer to previously unseen haze distributions (Tsai et al., 20 Jul 2025, Deng et al., 15 Jun 2025, Chang et al., 2024).

7. Extended Applications and Future Perspectives

Physics-aware dehazing is foundational in demanding domains such as UHD dehazing for remote sensing, HDR imaging under extreme illumination (rocket launches), and safety-critical tasks in autonomous vehicles. Extensions include multi-exposure fusion, entropy-constrained pyramid strategies for extreme dynamic range (Tao et al., 13 Jan 2026), and modular integration into edge-oriented frameworks. Advancements in physics-based feature learning and multi-domain generalization are expected to further enhance both interpretability and real-world robustness.

References