Abstract
In photographs captured in outdoor environments, particles in the air cause light attenuation and degrade image quality. This effect is especially obvious in hazy environments. In this study, a fuzzy theory is proposed to estimate the transmission map of a single image. To overcome the problem of oversaturation in dehazed images, a quadratic-function-based method is proposed to refine the transmission map. In addition, the color vector of the atmospheric light is estimated using the top 1% of the brightest light area. Finally, the dehazed image is reconstructed using the transmission map and the estimated atmospheric light. Experimental results demonstrate that the proposed hybrid method performs better than the other existing methods in terms of color oversaturation, visibility, and quantitative evaluation.
Introduction
Dehazing is a critical task in computer vision, aimed at recovering clear scene radiance from images obscured by atmospheric particles. Most modern dehazing frameworks rely on the standard haze model, which represents a hazy image as a combination of attenuated scene radiance and scattered atmospheric light.
The primary challenge in this model is accurately estimating the transmission map, which represents the portion of light that reaches the camera without being scattered. Existing methods, such as the Dark Channel Prior (DCP), often struggle with two specific artifacts:
- Halo Effects: Unnatural glowing edges around sharp depth transitions, such as buildings against a sky.
- Color Oversaturation: A tendency for dehazed regions to appear unnaturally vibrant or distorted, particularly in bright or heavily hazy areas.
This paper proposes a hybrid approach that leverages fuzzy logic for initial estimation and a quadratic function for refinement to mitigate these issues.
Methodology
The proposed pipeline consists of four distinct stages: atmospheric light estimation, fuzzy transmission estimation, quadratic refinement, and final image reconstruction.
Fuzzy Transmission Estimation
Rather than using a rigid mathematical prior, the authors employ a Fuzzy Inference System (FIS) to estimate the initial transmission map. This system maps RGB values into fuzzy sets (e.g., "Extremely Low" to "Extremely High") and applies expert-derived rules to determine transmission. This "soft" logic better handles the ambiguity of how color relates to depth in complex scenes.
Quadratic Refinement
To address the common problem of color shift and oversaturation, the initial transmission map is refined using a nonlinear quadratic function:
$$t_{ref}(x) = a(t(x)^2) + b(t(x)) + \frac{p}{1+p}$$
In this equation, $a$ and $b$ are tuning constants, and $p$ represents the average gray value of the image. This quadratic curve effectively "stretches" transmission values, preventing the dehazing process from pushing pixel values into unnatural, oversaturated ranges. The term $\frac{p}{1+p}$ specifically aids in maintaining natural gradients in bright sky regions.
Atmospheric Light Estimation
Instead of selecting the single brightest pixel—which may represent a localized light source rather than global atmospheric light—the authors identify the top 1% brightest pixels in the image's minimum-channel map. The average color vector of these pixels ($A_c$) provides a more robust reference for the reconstruction phase.
Results and Discussion
The method was evaluated using standard benchmarks, including urban cityscapes ("ny12", "ny17") and landscape images ("y01", "y16").
Qualitative Performance
Visually, the hybrid method produces images with significantly fewer halo artifacts around complex building geometries compared to the DCP method. The quadratic refinement ensures that sky regions retain natural color gradients rather than turning black or becoming grainy.
Quantitative Evaluation
Using the S-CIELAB color difference metric, where lower scores indicate better preservation of natural color, the proposed method consistently outperformed established algorithms. For the "ny17" dataset, the proposed method achieved a score of 345, vastly superior to Tan’s method (43,478) and He’s DCP (1,357).
Conclusion
By combining the flexible decision-making of fuzzy logic with the stabilizing properties of a quadratic refinement function, this research provides a robust solution to the persistent problems of halos and oversaturation in single-image dehazing. Future work will focus on automating the selection of quadratic parameters ($a$ and $b$) using machine learning to allow the algorithm to adapt dynamically to varying haze profiles.
Summary updated