Single image dehazing based on scene depth and physical model

Guo Jia, Xiaotong Wang, Chengpeng Hu, Xiaogang XuJournal of Image and Graphics 2012 · DOI

Introduction

Haze significantly degrades image quality by reducing contrast and washing out colors, posing a major challenge for computer vision systems in navigation, surveillance, and remote sensing. While the "Dark Channel Prior" (DCP) method revolutionized dehazing by observing that haze-free outdoor images typically contain pixels with very low intensity in at least one color channel, it has notable limitations. Specifically, DCP often fails in regions lacking dark objects, such as the sky or bright sea surfaces, leading to overestimation of haze, unnatural color shifts, and "halo" artifacts. This paper proposes a novel single-image dehazing method that addresses these issues by combining a geometric scene depth model with the classic atmospheric scattering physical model, particularly for maritime and flat-land environments.

Methodology

The proposed approach shifts from purely statistical estimation toward a physics-based geometric model to achieve more robust dehazing.

Geometric Scene Depth Model

Instead of inferring depth solely from pixel statistics, the authors derive a mathematical relationship between pixel coordinates and real-world distance. By modeling the sea surface as a plane and utilizing camera parameters—height ($H$), tilt angle ($\alpha$), and field of view ($\beta$)—a 3D-to-2D projection model is established. This allows for the calculation of a relative depth map for the entire scene, providing a more reliable prior than statistical methods.

Reference Depth and Sky Intensity Estimation

To integrate the geometric model with the physical scattering model, a "reference depth" ($d_0$) is selected. This is achieved by identifying a small, dark object in the image to serve as a baseline for the clearest point.

Furthermore, the estimation of atmospheric light ($E_\infty$) is improved. By analyzing the histogram of the dark channel and identifying regions with high brightness but low variance, the algorithm more accurately isolates the sky's true intensity. This prevents the "over-dehazing" of bright regions that typically plagues standard DCP implementations.

Image Restoration

The final step involves plugging the calculated depth and sky intensity into the standard atmospheric scattering equation. This process recovers the scene's original radiance, resulting in a haze-free image.

Evaluation and Results

The algorithm was evaluated on maritime hazy images and land/urban photos of flat terrain. Results were qualitatively compared against the seminal Dark Channel Prior method by He et al.

The proposed method demonstrates superior performance in several key areas:

  • Color Fidelity: Restored images maintain natural colors, particularly in the transition zones between the sea and sky.
  • Artifact Reduction: The "halo" effects and noise common in DCP results for bright areas are significantly minimized.
  • Versatility: While optimized for maritime scenes, the "flat-plane" assumption proved effective for general outdoor scenes with relatively flat terrain.

Discussion and Limitations

A key insight of this work is that utilizing known camera geometry (e.g., from ship-borne cameras) provides a far more robust depth prior than pixel-based inference. By treating the sky's depth as effectively infinite within the geometric model, the algorithm avoids introducing artifacts in bright sky regions. Additionally, the method allows for parameter tuning of sky intensity and reference depth, offering greater control over the final image's brightness and contrast.

The primary limitation is the reliance on accurate camera metadata (height and tilt angle). Inaccurate geometric models, caused by camera instability or extremely rugged terrain, can degrade performance. Future research may involve integrating dynamic sensors (IMUs) to update the geometric model in real-time or developing hybrid models for complex 3D structures.

Summary updated

BibTeX
@article{jia2012single,
  author = {Guo Jia and Xiaotong Wang and Chengpeng Hu and Xiaogang Xu},
  title = {{Single image dehazing based on scene depth and physical model}},
  journal = {Journal of Image and Graphics},
  year = {2012},
  doi = {10.11834/jig.20120104},
}
Figure