Abstract
We introduce an effective fusion-based technique to enhance both day-time and night-time hazy scenes. When inverting the Koschmieder light transmission model, and by contrast with the common implementation of the popular dark-channel [1], we estimate the airlight on image patches and not on the entire image. Local airlight estimation is adopted because, under night-time conditions, the lighting generally arises from multiple localized artificial sources, and is thus intrinsically non-uniform. Selecting the sizes of the patches is, however, non-trivial. Small patches are desirable to achieve fine spatial adaptation to the atmospheric light, but large patches help improve the airlight estimation accuracy by increasing the possibility of capturing pixels with airlight appearance (due to severe haze). For this reason, multiple patch sizes are considered to generate several images, that are then merged together. The discrete Laplacian of the original image is provided as an additional input to the fusion process to reduce the glowing effect and to emphasize the finest image details. Similarly, for day-time scenes we apply the same principle but use a larger patch size. For each input, a set of weight maps are derived so as to assign higher weights to regions of high contrast, high saliency and small saturation. Finally the derived inputs and the normalized weight maps are blended in a multi-scale fashion using a Laplacian pyramid decomposition. Extensive experimental results demonstrate the effectiveness of our approach as compared with recent techniques, both in terms of computational efficiency and the quality of the outputs.
Introduction: The Challenge of Night-time Dehazing
Most image dehazing algorithms rely on the Koschmieder light transmission model, which describes a hazy image as a combination of the clear scene radiance and "airlight" (light scattered by the atmosphere). A common assumption in popular methods like the Dark Channel Prior (DCP) is that airlight is uniform across the entire image.
While this holds true for daytime scenes where the sun is the primary light source, it fails for night-time environments. Night-time scenes are illuminated by multiple, localized, and often colored artificial sources such as street lamps, neon signs, and headlights. Applying a global airlight estimate to these scenes leads to severe artifacts, such as unnatural color shifts and "glowing" halos around light sources. The challenge lies in accurately estimating this non-uniform airlight without requiring extra information like depth maps or multiple images.
Proposed Methodology: Multi-Scale Local Fusion
The proposed framework centers on Local Airlight Estimation to handle non-uniform lighting. Instead of finding the brightest pixels in the whole image to estimate airlight, the algorithm calculates it over local patches. To resolve the "patch size dilemma"—where small patches allow for fine spatial adaptation but large patches provide more accurate haze estimation—the authors employ a multi-scale fusion approach.
Fusion Inputs and Weighting
The system generates three distinct inputs for the fusion process:
- Small-Patch Dehazing: Dehazed using a small patch size (e.g., 20x20) to adapt to local light sources.
- Large-Patch Dehazing: Dehazed using a larger patch size (e.g., 80x80) to restore global contrast and chroma.
- Discrete Laplacian: The Laplacian of the original image is used to re-inject high-frequency details and mitigate "glowing" artifacts.
To ensure the best parts of each input are preserved, three weight maps are calculated: local contrast, saturation, and saliency. These inputs and normalized weight maps are then blended using a Laplacian pyramid decomposition, ensuring seamless transitions and avoiding blocky artifacts.
Key Insights
- Generalization of DCP: The method is a "night-friendly" generalization of the classic Dark Channel Prior; when the local neighborhood expands to the entire image, it reduces to a global estimator.
- Independent Color Channels: Airlight is estimated for each color channel independently, which is crucial for scenes with multi-colored artificial lights.
- The Laplacian "Fix": Using the Laplacian as a third input effectively targets the "glow" around artificial lights, restoring fine details lost during dehazing.
Experimental Results
The method was evaluated on both daytime (O-HAZE) and night-time benchmarks. On the O-HAZE dataset, it achieved competitive results with state-of-the-art deep learning models in terms of PSNR and SSIM, while often superior in color preservation (CIEDE2000).
Qualitatively, the method significantly outperforms daytime-specific and other specialized night-time dehazing algorithms. As shown in the comparison below, the proposed approach restores better contrast and exhibits fewer color distortions compared to recent techniques.
Furthermore, the approach is highly efficient. An unoptimized Matlab implementation processed an 800x600 image in less than 4 seconds, whereas other specialized night-time methods required 20 to 30 seconds.
Conclusion and Future Directions
By replacing global atmospheric light estimation with a multi-scale local approach, this technique provides a robust solution for both day and night dehazing. While the fusion approach mitigates many artifacts, selecting optimal patch sizes remains somewhat empirical. Future work may involve automating patch size selection or integrating noise-reduction techniques to handle extremely dark, high-noise regions.
Summary updated