Abstract
Accurate estimation of haze density is vital for meteorological monitoring and the safety of autonomous systems. This paper proposes a novel framework based on the Normalized Scattering Coefficient (NSC), which provides a relative measure of atmospheric visibility. By utilizing superpixel segmentation to identify regions of uniform depth and a reference-based calibration technique, the method effectively resolves the ambiguities inherent in the standard Atmospheric Scattering Model. Evaluations on benchmark datasets show that the approach achieves high precision in outdoor scenes, especially when filtering for dark regions to reduce estimation noise.
Introduction
Measuring haze density typically requires specialized hardware like haze meters, which are often too expensive for dense, real-time deployment. Vision-based alternatives using standard cameras are scalable but mathematically challenging due to the Atmospheric Scattering Model. In this model, the light captured by a camera is a product of scene albedo, object depth, and the scattering coefficient ($\beta$). Solving for $\beta$ is an "ill-posed" problem because these variables are coupled. Existing methods often fail under varying lighting conditions or camera movement.
Proposed Methodology
The core of the proposed framework is the transition from absolute physical measurements to a relative scale. The authors define the Normalized Scattering Coefficient (NSC), represented by $\gamma$, which ranges from 0 (a clear reference image) to 1 (a heavy haze reference image).
The process involves four main stages:
- Reference Calibration: Two images of the same scene under different known haze conditions serve as the baseline.
- Iso-depth Segmentation: The target image is partitioned into superpixels (SPs). This assumes that pixels within a single SP share approximately the same depth, simplifying the scattering equations.
- Mathematical Elimination: By comparing the target SPs to the corresponding regions in the reference images, the unknown albedo and depth variables are canceled out, leaving the ratio of scattering coefficients.
- Dark SP Selection: To improve robustness, the algorithm filters for "dark" superpixels. Based on the Dark Channel Prior, these regions are less likely to be affected by the noise typical of bright areas like the sky, leading to more stable depth and haze estimates.
Experimental Results
The model was evaluated using the Fattal et al. dataset, comprising 11 diverse scenes. To ensure a rigorous test, 55 hazy variations were synthesized for each scene, covering various lighting (airlight) colors and scattering intensities.
The results indicate that the NSC estimation is highly accurate for outdoor environments such as roads, churches, and lawns. In these cases, the estimated values closely track the ground truth, as shown by the tight diagonal alignment in the precision plots. A key finding was the necessity of the "Dark SP" selection; without it, the model tended to overestimate haze in clear conditions.
Discussion and Limitations
While the framework excels in outdoor settings, its performance degrades in indoor environments (e.g., the couch or flowers scenes). This is primarily due to the absence of a visible sky, which complicates the estimation of "airlight"—the ambient light scattered by the haze. Future research will focus on improving airlight estimation in enclosed spaces, potentially extending the utility of this method to tunnels and large industrial facilities.
Conclusion
This research demonstrates that a relative, reference-based approach to haze estimation can provide high-precision results without the need for specialized sensors. By combining superpixel segmentation with robust region selection, the proposed NSC framework offers a scalable solution for real-time atmospheric monitoring.
Summary updated