A PDE-Based Image Dehazing Method via Atmospheric Scattering Theory

Basic Information

• Paper ID: 2506.08793
• Title: A PDE-Based Image Dehazing Method via Atmospheric Scattering Theory
• Authors: Liubing Hu, Pu Wang, Guangwei Gao, Chunyan Wang, Zhuoran Zheng
• Categories: cs.CV (Computer Vision), eess.IV (Image and Video Processing)
• Publication Date: 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2506.08793

Abstract

This paper proposes a novel single-image dehazing framework based on partial differential equations (PDEs). The method embeds the atmospheric scattering model into a PDE incorporating edge-preserving diffusion and nonlocal operators to maintain local details and global structure. A key innovation is an adaptive regularization mechanism based on dark channel priors, which adjusts smoothing strength according to haze density. The mathematical well-posedness of the framework is rigorously established by proving the existence and uniqueness of weak solutions in the H₀¹(Ω) space. An efficient GPU-accelerated fixed-point solver is employed for implementation. Experiments confirm that the method effectively removes haze while maintaining high image fidelity, providing a principled alternative to purely data-driven techniques.

Research Background and Motivation

Problem Definition

Image dehazing is a critical challenge in computer vision, addressing image degradation caused by atmospheric scattering. Hazy images follow the atmospheric scattering model: I(x) = J(x)t(x) + A(1-t(x)), where I(x) is the observed hazy image, J(x) is the clear image, t(x) is the transmission map, and A is the atmospheric light value.

Problem Significance

1. Practical Application Demands: Hazy weather severely impacts outdoor vision system performance, such as autonomous driving and surveillance systems
2. Image Quality Enhancement: Dehazing significantly improves image contrast and visibility
3. Foundation for Subsequent Processing: Provides high-quality inputs for advanced vision tasks such as object detection and scene understanding

Limitations of Existing Methods

1. Traditional Physics-Based Methods:
   • Dark channel prior (DCP) fails in sky regions and uniform scenes
   • Results in inaccurate transmission map estimation
2. Data-Driven Methods:
   • Require large amounts of annotated data
   • Lack interpretability
   • Perform poorly under out-of-distribution haze conditions

Research Motivation

Combine physical modeling with mathematical rigor to provide a dehazing framework that offers both theoretical guarantees and practical effectiveness, bridging the gap between traditional methods and deep learning approaches.

Core Contributions

1. Innovative PDE Framework: First to embed the atmospheric scattering model into a PDE with edge-preserving diffusion and nonlocal operators
2. Adaptive Regularization Mechanism: Adaptive regularization parameters based on dark channel priors that dynamically adjust smoothing strength according to haze density
3. Mathematical Theoretical Guarantees: Rigorously proves the existence and uniqueness of weak solutions to the PDE in the H₀¹(Ω) space
4. Efficient Implementation: GPU-accelerated fixed-point iterative solver enabling real-time processing
5. Superior Performance: Achieves state-of-the-art performance on multiple no-reference image quality assessment metrics

Method Details

Task Definition

Input: Hazy image I(x) ∈ ℝᴴˣᵂˣ³ Output: Dehazed image u(x) ∈ ℝᴴˣᵂˣ³ Constraints: Comply with atmospheric scattering physics model, preserve image structure and details

Model Architecture

1. Atmospheric Scattering Model Embedding

The reconstruction operator is defined as:

Φ(I, t, A) = (I - A(1-t)) / max(t, t₀)

where t₀ > 0 is a small threshold to avoid numerical instability.

2. Core PDE Model

-div(D(∇u)∇u) + λ(t)G(u) = Φ(I, t, A)

Components:

• Edge-Preserving Diffusion Term: -div(D(∇u)∇u)
• Nonlocal Regularization Term: λ(t)G(u)
• Data Fidelity Term: Φ(I, t, A)

3. Edge-Preserving Diffusion Mechanism

The diffusion coefficient is designed as:

D(∇u) = (|∇u| + ε)⁻¹

where ε = 10⁻³. This design ensures:

• Suppression of diffusion at strong edges (large gradients)
• Promotion of diffusion in smooth regions (small gradients)

4. Nonlocal Regularization

Gaussian convolution operator:

G(u)(x) = ∫_Ω K(x,y)u(y)dy

Gaussian kernel:

K(x,y) = (1/2πσ²)exp(-‖x-y‖²/2σ²)

5. Adaptive Regularization Parameter

Adaptive parameter based on dark channel priors:

λ(t) = λ₀exp(-β(1-t))

where λ₀ = 0.5, β = 3.0. This design ensures:

• Dense haze regions (small t values): Strong regularization
• Clear regions (large t values): Weak regularization, preserving details

Technical Innovations

1. Physical Consistency: Directly embeds the atmospheric scattering model into PDE constraints, ensuring physical validity
2. Adaptivity: Dynamically adjusts regularization strength according to local haze concentration
3. Multi-Scale Processing: Combines local diffusion and nonlocal operators to handle both local details and global structure
4. Mathematical Rigor: Provides complete theoretical analysis proving existence and uniqueness of solutions

Experimental Setup

Datasets

• Real-World Datasets: Containing diverse hazy images
• Scene Types: Urban streets, natural landscapes, buildings
• Haze Concentration: Ranging from light to severe
• Evaluation Approach: No-reference image quality assessment due to lack of real corresponding clear images

Evaluation Metrics

Six widely recognized no-reference image quality assessment (NR-IQA) metrics are employed:

• NIQE, BRISQUE, PIQE, FADE: Lower is better
• DHQI, NRBP: Higher is better

Comparison Methods

• Traditional Methods: DCP, CAP, Color-Lines, ICAP, BCCR
• Fusion Methods: VarFusion
• Deep Learning Methods: CORUN, DiffDehaze

Implementation Details

• Platform: Python + PyTorch, NVIDIA RTX 4090 GPU
• Parameter Settings:
  • Dark channel block size: 15×15
  • Diffusion stability term: ε = 10⁻³
  • Gaussian kernel: 5×5, σ = 2.0
  • Relaxation parameter: τ = 0.2
  • Adaptive parameters: λ₀ = 0.5, β = 3.0

Experimental Results

Main Results

The proposed method achieves best performance on 5 out of 6 NR-IQA metrics:

Ablation Study

Systematic removal of components validates their contributions:

Key Findings:

• Removing PDE optimization causes NRBP to drop from 76.82 to 41.49
• Nonlocal regularization and edge-preserving terms are critical for image naturalness

User Study

• Participants: 5 experts + 16 general observers
• Evaluation Dimensions: Haze removal degree, detail clarity, color naturalness
• Results: The proposed method received 62% of "best visual quality" votes, significantly outperforming other methods

Case Analysis

Qualitative results demonstrate:

• Traditional methods (DCP, CAP) tend to produce color shifts and artifacts
• Deep learning methods occasionally fail to preserve details
• The proposed method consistently produces visually superior results with effective haze removal and natural color preservation

Related Work

Main Research Directions

1. Physics-Model-Based Methods: Utilizing atmospheric scattering models and various priors (dark channel, color attenuation, etc.)
2. Data-Driven Methods: CNNs, attention mechanisms, generative adversarial networks
3. Hybrid Methods: Combining physical constraints and learned representations

Advantages of This Work

Compared to existing work:

• Provides a PDE framework with mathematical theoretical guarantees
• Adaptive regularization mechanism outperforms fixed-parameter methods
• Stronger physical consistency than purely data-driven methods
• Higher computational efficiency than complex deep networks

Conclusions and Discussion

Main Conclusions

1. Successfully embeds atmospheric scattering theory into a mathematically rigorous PDE framework
2. Adaptive regularization mechanism significantly improves dehazing performance
3. Achieves state-of-the-art performance in multiple objective and subjective evaluations
4. Provides a principled alternative to data-driven methods

Limitations

1. Computational Complexity: PDE solving requires iterative optimization, potentially slower than direct CNN inference
2. Parameter Sensitivity: Multiple hyperparameters require tuning, potentially affecting generalization
3. Extreme Conditions: May fail under extremely heavy haze or special scenes
4. Dark Channel Assumption: Still relies on the validity of dark channel priors

Future Directions

1. Hybrid Architecture: Combining PDE framework with deep learning for end-to-end training
2. Real-Time Optimization: Developing more efficient solving algorithms
3. Multi-Modal Fusion: Incorporating information from other sensors (depth, polarization)
4. Theoretical Extension: Generalizing the framework to other image restoration tasks

In-Depth Evaluation

Strengths

1. Theoretical Innovation: First to provide mathematically rigorous analysis of dehazing PDE, including existence and uniqueness proofs
2. Method Design: Skillfully combines physical models, mathematical theory, and computational implementation
3. Comprehensive Experiments: Includes quantitative comparisons, ablation studies, user evaluations, and multi-dimensional verification
4. Practical Value: GPU-accelerated implementation has potential for real-world applications

Weaknesses

1. Computational Efficiency: Iterative solving may be slower compared to end-to-end deep networks
2. Generalization Ability: Robustness in diverse real-world scenarios requires further verification
3. Parameter Tuning: Multiple hyperparameter settings may require optimization for different scenarios
4. Baseline Comparison: Lacks sufficient comparison with latest SOTA deep learning methods

Impact

1. Academic Contribution: Provides a new mathematical framework for image dehazing, potentially inspiring related research
2. Methodological Value: Demonstrates how to rigorously embed physical models into PDE frameworks
3. Practical Prospects: Advantageous in application scenarios requiring high interpretability
4. Reproducibility: Provides detailed implementation details and parameter settings

Applicable Scenarios

1. Research Applications: Research scenarios requiring theoretical guarantees and interpretability
2. Real-Time Systems: Systems with certain computational efficiency requirements but needing stable performance
3. Special Environments: Scenarios with scarce training data but applicable physical models
4. Hybrid Systems: As a supplement or initialization for deep learning methods

References

The paper cites 25 relevant references covering multiple aspects including image dehazing, PDE methods, and deep learning, providing a solid theoretical foundation and comparison benchmarks for this research.

Overall Assessment: This is a high-quality paper with significant contributions in both theory and practice. The authors successfully combine physical models, mathematical theory, and computational implementation to provide a novel and rigorous solution to the image dehazing problem. While there is room for improvement in computational efficiency and generalization ability, its theoretical innovations and experimental validation meet very high standards.