Abstract
In foggy conditions, acquired images have low contrast ratio and low scene visibility problems. The traditional defogging algorithm has high time complexity and low speed characteristics and can not be applied in real-time image processing. Thus, a single image restoration method based on mean filter is proposed in this paper combined with the optical properties of atmospheric. In this method, on the basis of atmospheric scattering model, accurate atmospheric dissipation function can be achieved firstly by using mean filter. Atmosphere light is obtained by quadtree algorithm which is more efficient. Finally, restored image is calculated by the atmospheric scattering model and the brightness of the image is adjusted, and a clear image without fog is gained. Simulation results show that this algorithm has a strong adaptability to various scenes and restored images have plentiful colors. Compared with the classic defogging algorithm, this algorithm ensures the defogging effect while overcomes the high time complexity and low speed defects of guide filter algorithm.
Introduction
Images captured in foggy or hazy conditions suffer from significantly reduced contrast and poor visibility due to the scattering of light by atmospheric particles. This degradation poses a major challenge for computer vision applications such as autonomous driving, traffic monitoring, and outdoor surveillance.
Existing dehazing algorithms generally fall into two categories. High-quality methods, such as He et al.’s "Dark Channel Prior," produce visually impressive results but rely on computationally intensive refinement steps like soft matting or guided filtering. These steps are often too slow for real-time processing. Conversely, faster methods using simple filters (e.g., median filters) often suffer from color distortion, excessive darkening, or "halo" artifacts around object edges. This paper proposes a high-speed single-image defogging algorithm that bridges this gap by replacing expensive refinement steps with a compensated mean filter and an efficient quadtree-based search for atmospheric light.
Methodology
The proposed algorithm is grounded in the standard Atmospheric Scattering Model, which represents a hazy image as a combination of the attenuated clear scene and the ambient "airlight." The authors introduce three key technical improvements to optimize this model for speed and quality:
1. Compensated Mean Filtering
To estimate the atmospheric dissipation function without the overhead of the Dark Channel Prior, the authors utilize a simple mean filter. To mitigate the "flat" appearance typically associated with mean filtering, a compensation parameter ($\sigma$) is introduced. This parameter adaptively adjusts based on the image's global gray level, preserving local contrast while maintaining computational efficiency.
2. Quadtree Subdivision for Atmospheric Light ($A$)
Accurately estimating atmospheric light is critical, as bright objects (e.g., white buildings) can be erroneously identified as dense fog. The authors employ a quadtree subdivision method: the image is iteratively divided into four quadrants, and the quadrant with the highest mean brightness is selected. This hierarchical search quickly converges on the most representative "foggy" area, enhancing robustness against high-frequency bright details.
3. Post-Restoration Tone Adjustment
Physical restoration often results in images that appear unnaturally dark. To improve visual appeal, the algorithm includes a tone mapping step performed in the log-domain. This step boosts brightness and contrast, ensuring the final output is both physically accurate and aesthetically pleasing.
Experimental Results
The algorithm was evaluated using four groups of simulated and real-world foggy images representing diverse scenarios, including urban environments and indoor scenes.
Performance Comparison
The proposed method demonstrates a significant speed advantage over traditional approaches. In comparative tests, the proposed algorithm processed a test image in 1.4095 seconds, whereas He et al.’s method required 11.802 seconds.
Visual and Quantitative Quality
Qualitatively, the restored images exhibit fewer halo artifacts than median-filter-based methods and more vivid color reproduction than the Dark Channel Prior. Quantitatively, the method was assessed using the "visible edge gradient" metrics ($e$ and $\bar{r}$). In the "Fourth group" test, the proposed method achieved a gradient ratio ($\bar{r}$) of 2.5245, outperforming both He (1.9911) and Tarel (2.2961).
Conclusion
This study presents an efficient single-image defogging algorithm that achieves near real-time performance without sacrificing visual quality. By optimizing the estimation of the dissipation function and atmospheric light, the method overcomes the high time complexity of guided filter-based approaches. Future work will focus on improving parameter robustness for extreme weather conditions and ensuring temporal consistency for video-based applications like real-time road monitoring.
Summary updated