Abstract
Visibility Restoration for Hazy Images (VRHI) presents a novel approach to image dehazing that bypasses the complexities of neural network training. By integrating a specialized haze density model with a recursive quadtree search strategy, the method achieves accurate atmospheric light and transmission map estimation. This physical-model-based strategy effectively restores clarity in real-world scenes while avoiding common pitfalls such as color distortion, over-saturation, and halo artifacts.
Introduction
Haze is a frequent atmospheric nuisance that scatters light, leading to low-contrast, washed-out images that hinder both human viewing and computer vision tasks. Existing solutions generally fall into two categories: prior-based methods, which rely on physical heuristics like the Dark Channel Prior (DCP), and learning-based methods, which utilize deep neural networks.
However, both approaches have significant drawbacks. Learning-based models often suffer from "domain shift," where they fail to generalize from synthetic training data to real-world conditions. Conversely, simpler prior-based methods frequently produce unrealistic colors, "halo" artifacts around edges, or loss of detail in shadowed regions. VRHI bridges this gap by utilizing a sophisticated physical model that is robust enough for real-world application without requiring extensive training datasets.
Methodology
The VRHI framework is built upon the standard Atmospheric Scattering Model (ASM), which defines a hazy image as a combination of clear scene radiance and atmospheric "airlight." The primary challenge lies in accurately estimating the atmospheric light and the transmission map.
Haze Density Model and Transmission Estimation
The authors propose a new exponential model to estimate haze distribution. This model considers both the minimum color channel and the local contrast (the difference between maximum and minimum channels), providing a more accurate proxy for haze than traditional methods. From this density model, a global transmission map is derived, incorporating a vertical adjustment factor ($\lambda$) to account for the physical reality that haze is often denser toward the horizon in landscape imagery.
Recursive Quadtree for Atmospheric Light
To identify the global atmospheric light, VRHI employs a recursive quadtree search. A unique "centrifugation parameter" biases the search toward the upper portion of the image, where the sky and primary light sources are typically located. To prevent bright local objects (like white buildings or cars) from skewing the results, the algorithm averages the brightest pixels (ranging from 5% to 40%) within the final identified region.
Results and Discussion
VRHI was evaluated using both real-world hazy images and the Synthetic Objective Testing Set (SOTS) from the RESIDE dataset. It was compared against state-of-the-art (SOTA) methods, including deep learning models like EPDN and MSBDN.
Qualitative and Quantitative Performance
In visual comparisons, VRHI consistently produced images with more natural color reproduction. While competing methods often resulted in images that appeared too dark or artificially over-saturated, VRHI maintained the original scene's tonal balance. Quantitatively, the method demonstrated superior performance in restoring details across diverse luminance levels, specifically avoiding the "over-enhancement" that plagues many other dehazing algorithms.
Key Insights
- Spatial Priors: Incorporating the vertical gradient of haze density leads to significantly more natural results than assuming uniform distribution.
- Robustness: The weighted averaging in the quadtree search makes the light estimation highly resilient to local outliers and bright objects.
- Efficiency: As a "plug-and-play" physical algorithm, VRHI is more predictable and easier to deploy in resource-constrained environments compared to heavy deep learning models.
Conclusion
VRHI offers a robust, training-free solution for visibility restoration. While the model assumes a somewhat predictable haze distribution—which may pose challenges for extremely patchy fog—it represents a significant step forward in physical-model-based dehazing. Future work may focus on optimizing the recursive calculations for real-time video applications and adapting the centrifugation parameters for specialized imagery, such as underwater or aerial photography.
Summary updated