Spatially Varying Atmospheric Light for Image Restoration

29 concepts18 papers1 proposal

The classical Atmospheric Scattering Model (ASM) has long been the bedrock of image dehazing and visibility enhancement. In its traditional formulation, an observed hazy image $I(x)$ is modeled as a combination of the clear scene radiance $J(x)$ and the atmospheric light $A$, attenuated by a transmission map $t(x)$:

$$I(x) = J(x)t(x) + A(1 - t(x))$$

For decades, standard algorithms operated under a crucial simplification: they treated the atmospheric light $A$ as a global constant across the entire image. While mathematically convenient, this assumption assumes an idealized world illuminated uniformly by a single, distant light source.

Real-world environments routinely violate this assumption. Whether dealing with the multi-colored artificial lights of a nighttime cityscape, the uneven sunlight of a wide-angle aerial view, or non-homogeneous patches of fog, the ambient illumination changes from pixel to pixel. When a model forces a single global constant onto a heterogeneous scene, the restoration fails—resulting in severe color distortions, halo artifacts, and loss of fine details.

To solve this, a powerful research direction has emerged: replacing the global constant assumption with spatially varying atmospheric light modeling. By treating illumination as a localized, pixel-dependent field, these modern frameworks allow image restoration to adapt dynamically to the complex geometry and lighting of the real world.


Why the Global Assumption Fails

When a dehazing model assumes $A$ is uniform, it lacks the mathematical flexibility to distinguish between variations in haze density and variations in local illumination. As noted by Wenqi Ren et al. in their work on gated fusion networks, this violation causes models to produce highly inconsistent colors and contrast across different regions of an image.

This failure manifests in three distinct ways:

  1. Sky Distortions and Color Shifts: In expansive outdoor scenes, treating atmospheric light as a global constant fails to account for the natural gradient of sunlight across the horizon. As Shengdong Zhang et al. demonstrate, this causes traditional models to over-saturate colors and introduce grainy artifacts into sky regions.
  2. Aerial Overestimation: In wide-angle photography or drone imaging, lighting conditions vary significantly across vast landscapes. Tao Gao et al. show that global estimation leads to an overestimation of haze in brighter areas, causing unnatural dimming and loss of contrast in the restored output.
  3. Nighttime Glow and Halos: At night, the environment is illuminated by localized, heterogeneous sources like streetlights or neon signs. A global model cannot reconcile these distinct color temperatures and intensities, leading to severe halo artifacts around light sources.

Redefining Light as a Spatially Variant Field

The core mechanics of this research direction rely on transforming $A$ from a single vector into a spatially variant matrix or continuous field, $A(x)$, where $x$ represents the pixel coordinate.

Traditional ASM:  [Image] ---> [Single Global Light Vector A] -------> [Uniform Correction]
Spatially Varying: [Image] ---> [Pixel-wise / Local Light Map A(x)] ---> [Adaptive Correction]

To capture this spatial distribution, researchers have engineered several distinct strategies:

Pixel-Location-Dependent Coefficients

Instead of calculating a single global value based on the brightest pixels, Shengdong Zhang et al. introduce a pixel-location-dependent coefficient. This allows the dehazing equations to scale adaptively across the image plane, eliminating distortions in expansive regions like the sky. Similarly, Aiping Yang et al. estimate a spatially variant 3-channel matrix to represent non-uniform lighting conditions, effectively neutralizing sunlight glow and preventing color imbalances.

Multi-Scale Local Airlight Estimation

Rather than moving directly to pixel-level matrices, another approach divides the image into localized regions. Cosmin Ancuti et al. propose a multi-scale local airlight estimation framework. By calculating airlight within bounded local windows and fusing the results, their method adapts to diverse light sources within a single frame, making it equally viable for daytime and nighttime environments.

Illumination-Adaptive Scattering Models

To explicitly decouple light transport from environmental interference, Huaizhuo Liu et al. propose an Illumination-Adaptive Scattering Model (IASM). This mechanism replaces global uniform light with distinct, spatially varying parameters for both local incident illumination and environmental scattering, allowing neural networks to resolve complex shadows and non-homogeneous environmental conditions.


Taming Nighttime and Non-Uniform Environments

The power of spatially varying illumination modeling is most evident in extreme environments, such as low-light or nighttime scenes.

Nighttime images suffer from scattering-induced "glow"—the blurry halos that surround streetlights. Yu Li et al. address this by extending the atmospheric scattering model to include a dedicated term for multiple scattering (glow) while simultaneously treating the background atmospheric light as a spatially varying map. This allows their framework to decompose and strip away the localized glow before performing targeted dehazing.

Nighttime Pixel Intensity = Scene Radiance + Spatially Varying Airlight + Localized Glow Term

For low-light object detection, Xin Xu et al. model environmental light as a point-wise variable to accommodate the complex flickering and varying intensities typical of artificial night scenes. In a similar vein, the authors of the Maximum Reflectance Prior model ambient illumination as a spatially varying component, isolating local light colors and intensities to prevent glow artifacts without sacrificing computational speed.

This pixel-wise scaling is also crucial for handling non-uniform color casts caused by varied atmospheric particles like smoke, dust, or sand. Akshay Dudhane et al. employ channel-wise convolutions to generate pseudo-color maps for illumination estimation, allowing the model to neutralize localized, varicolored hazes.


Beyond Dehazing: 3D Reconstruction and Neural Fields

The intuition behind spatially varying illumination extends far beyond 2D image defogging; it is equally vital for recovering 3D structure and material properties.

When mapping a 3D environment, traditional inverse rendering methods often assume a single, distant environment map. However, as Changwoon Choi et al. point out, this assumption fails to capture how lighting changes based on an observer's position relative to local sources like lamps or windows. By modeling radiance as a spatially varying field within a Neural Radiance Field (NeRF) framework, their method achieves highly accurate material and irradiance decomposition.

Furthermore, ignoring local lighting variations can corrupt geometric optimization. If an algorithm assumes uniform lighting, it will frequently mistake a local shadow or a bright reflection for physical geometry. Robert Maier et al. solve this in 3D reconstructions by jointly optimizing geometry and a spatially varying lighting field, ensuring that non-uniform indoor lighting is correctly attributed to the illumination field rather than being baked into the physical surface texture.

Key Foundational Work

Nighttime Haze Removal with Glow and Multiple Light Colors (2015) — contributed 2 concept elements to this direction.
A Variational Framework for Single Image Dehazing (2014) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_spatially_varying_atmospheric_light_mode_22d4ca72,
  title = {spatially varying atmospheric light modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatially-varying-atmospheric-light-modeling/rdr_22d4ca72b73d421fb802ec5d6bebc2e2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_spatially_varying_atmospheric_light_mode_22d4ca72_v1,
  title = {spatially varying atmospheric light modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatially-varying-atmospheric-light-modeling/rdr_22d4ca72b73d421fb802ec5d6bebc2e2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake