Most modern image dehazing techniques rely on a simple physical model: the hazy image you see is a combination of the clear scene and a veil of scattered light. Mathematically, this is expressed as $I(x) = J(x)t(x) + A(1 - t(x))$, where $I$ is the hazy input, $J$ is the clear scene, $t$ is the transmission (how much light reaches the camera), and $A$ is the atmospheric light (or airlight).
To recover $J$, an algorithm must first estimate $A$. If $A$ is estimated incorrectly, the entire restoration collapses, leading to images that are unnaturally dark, over-saturated, or plagued by "halo" artifacts around bright objects. Recent research focuses on moving beyond simple heuristics to robustly estimate this critical parameter in complex, real-world environments.
The Failure of the Brightest Pixel Heuristic
Early dehazing methods, such as the classic Dark Channel Prior (DCP), often assumed that the brightest pixels in an image represent the global atmospheric light. The intuition was that in a hazy scene, the most distant points would be the most obscured, and thus their color would be dominated by the ambient light scattered by the atmosphere.
However, as Codruta Ancuti and Radu Timofte demonstrate in their O-HAZE and I-HAZE benchmarks, real-world physics is rarely that simple. In practice, the "brightest pixel" is often not the sky or distant haze, but a white car, a reflective sign, or a localized light source. When an algorithm mistakes a white object for the global airlight, it overestimates the haze intensity. This causes the model to "subtract" too much light, resulting in a restored image that is excessively dark and loses all detail in high-luminance regions.
Segmenting the Sky and Masking Interference
One primary strategy for robust estimation is to treat different regions of the image with specialized logic. The sky is a frequent culprit for estimation errors because its natural brightness mimics the appearance of dense haze.
Wei Jie and Somaiyeh Mahmoud Zadeh address this in Single Image Dehazing Algorithm Based on Sky Region Segmentation by explicitly segmenting the image into sky and non-sky regions. By applying different priors to each—using deep learning for the sky and optimized statistical priors for the landscape—they prevent the sky's luminance from biasing the transmission map of the foreground objects.
Similarly, point light sources like streetlamps or headlights create "glow" effects that traditional models cannot distinguish from uniform haze. Yanyang Yan and Wenqi Ren proposed a Point Light Source Interference Removal Method that identifies and masks these localized highlights. By excluding these "false positives" from the atmospheric light calculation, the algorithm can accurately estimate the ambient haze without being blinded by the glare of a lamp.
Adapting to the Night
Nighttime dehazing presents a unique challenge because the atmospheric light is no longer a single, global constant. Instead, multiple colored light sources create a non-uniform field of illumination.
To solve this, Jing Zhang and Yang Cao introduced the Maximum Reflectance Prior (MRP). While standard dehazing looks for the darkest pixels to estimate haze, MRP looks at the brightest pixels in local patches. This shift allows the algorithm to track how light from different sources (like a yellow streetlamp vs. a blue neon sign) scatters through the haze, enabling high-speed restoration that avoids the "halo" artifacts typical of nighttime processing. Shuai Zheng further refined this in New Nighttime Dehazing Algorithm Based On Dark Channel Prior Enhancement by using hybrid filtering to fuse light-source and non-light-source regions into a cohesive, natural-looking result.
Hierarchical and Geometric Constraints
When simple pixel statistics fail, researchers turn to more complex search patterns and global constraints to find the true airlight.
- Recursive Quadtree Search: Instead of looking at individual pixels, methods by Mingye Ju in VRHI and Yanhai Wu in A Quick Study of a Single Image Defogging Algorithm use a quadtree decomposition. They repeatedly divide the image into quadrants and search for the region with the highest haze density. This hierarchical approach is much more robust to high-frequency noise and small bright objects than a simple global maximum.
- Depth Ordering: Huaian Chen and Yi Jin suggest that the relative "near-to-far" perception of a scene is a powerful constraint. In Single Image Dehazing Using Scene Depth Ordering, they ensure that the restored image maintains the same depth hierarchy as the original. If a restoration step would make a "far" object appear "closer" than a "near" one, the atmospheric light estimate is adjusted to maintain geometric consistency.
By moving from a "brightest pixel" heuristic to a spatially-aware, source-sensitive estimation of atmospheric light, these methods allow computer vision systems to see through haze in environments—from nighttime city streets to rocket launch pads—where traditional algorithms would fail.