The classic physics-based framework for image dehazing relies on the atmospheric scattering model. In this idealized formulation, an observed hazy image is a linear combination of the clean scene radiance and the atmospheric light, modulated by a transmission map. For decades, algorithms operated under a foundational assumption: the atmospheric medium is homogeneous. Under this assumption, the scattering coefficient is constant across the scene, meaning that haze density scales uniformly and exponentially with depth alone.
In the real world, this assumption breaks down. Real-world atmospheric interference—such as swirling fog, localized smoke, or patchy mist—is inherently non-homogeneous. Its density fluctuates unpredictably across a single frame due to micro-climates, wind, and local pollution sources.
When a model built on the homogeneity assumption encounters real-world patchy fog, it fails catastrophically. Because it assumes a uniform distribution, it applies a globally consistent or smoothly varying correction based strictly on depth. As a result, the model under-corrects areas with dense, localized pockets of fog (leaving behind residual hazy patches) while over-correcting clearer regions (introducing severe color distortions, artifacts, and over-saturation).
To resolve this, a modern research direction has emerged that abandons global uniformity in favor of spatially varying, density-aware degradation modeling.
The Core Intuition: Spatially Varying Density Estimation
If haze does not vary uniformly with depth, a model cannot rely solely on a global transmission map. Instead, it must explicitly perceive and map the localized density of the degradation at every pixel.
[Hazy Input] ──> [Density Estimation Module] ──> [Continuous Density Mask]
│
[Restored Image] <── [Spatially Adaptive Processing] <─────────┘
By treating the degradation as a continuous, space-variant field, networks can dynamically adjust their restoration intensity. For instance, Ye et al. introduced PMDNet, which explicitly encodes the intensity of degradation at each pixel using a Density Encoding Matrix. This allows the network to "know" how much haze is present in a specific region before attempting to remove it. Similarly, Kim et al. proposed FALCON, which leverages a physics-inspired "Continuous Density Mask" to capture the highly fluctuating, non-linear nature of patchy fog rather than relying on binary or smoothly varying assumptions.
By grounding the network's behavior in local density variations, these architectures ensure that computational resources and restoration strengths are concentrated precisely where the interference is thickest, preserving clean regions from over-processing.
Treating Pixels Unequally via Attention
Traditional convolutional layers apply the same learned weights across an entire image, which inherently struggles with non-uniform degradation. To handle space-variant haze, networks must treat different spatial regions and color channels unequally.
This is achieved primarily through advanced attention mechanisms:
- Pixel and Channel Attention: Qin et al. developed FFA-Net (Feature Fusion Attention Network), which pioneered the combination of channel and pixel attention for dehazing. Pixel attention allows the network to provide more weight to heavily obscured pixels, while channel attention accounts for the fact that haze affects different color channels differently under varying lighting conditions.
- Adaptive Feature Distillation: Chen et al. designed the Multi-scale Adaptive Dehazing Network (MADN), which employs an adaptive distillation module to prioritize critical features dynamically, bypassing the rigid constraints of traditional physical scattering models.
- Haze-Aware Representation: Deng et al. introduced HardGAN, utilizing a specialized Haze-Aware Representation Distillation (HARD) module. This architecture adaptively fuses global atmospheric brightness with local spatial structures, enabling the network to adjust normalization layers based on local haze intensity.
Multi-Scale and Patch-Based Hierarchies
Because non-homogeneous haze exhibits localized patterns that do not correlate with global image features, processing the image solely at a single global scale is insufficient. Modern networks break the image down hierarchically to analyze degradation at multiple granularities.
Das et al. addressed this with the Deep Multi-patch Hierarchical Network (DMPHN). Instead of relying on global scales, DMPHN processes images through a hierarchy of localized patches. This architecture allows the network to isolate dense pockets of haze within small local windows, removing them without affecting the broader global context.
From a non-deep-learning perspective, Shih-Chia Huang demonstrated a similar intuition by introducing a hybrid transmission map approach. By merging different patch-sized estimations using Laplacian pyramids and saliency masks, this method restores heavily obscured backgrounds without introducing artifacts into clear foreground objects.
Shifting the Benchmark: From Synthetic to Real Non-Homogeneity
For years, progress in image dehazing was bottlenecked by training data. Synthetic datasets typically generated haze by applying a uniform scattering coefficient to known depth maps. Models trained on these datasets learned to invert that specific, uniform mathematical relationship, rendering them ineffective when deployed in real-world, patchy environments.
A major catalyst for the shift toward non-homogeneous modeling was the creation of new benchmarking datasets that captured real-world irregularity. Ancuti et al. introduced NH-Haze, the first real-world dataset containing non-homogeneous hazy images paired with identical haze-free ground truths. To simulate authentic atmospheric conditions without relying on synthetic rendering, the authors used professional haze machines to generate vapor particles and fans to manually distribute them into non-uniform, patchy configurations.
Synthetic Uniform Haze: I(x) = J(x)e^{-β d(x)} + A(1 - e^{-β d(x)}) <-- β is constant
Real Non-Homogeneous Haze: I(x) = J(x)e^{-β(x) d(x)} + A(1 - e^{-β(x) d(x)}) <-- β(x) varies spatially
This benchmark exposed the limitations of standard global transmission models and forced the development of architectures capable of tracking highly fluctuating, non-linear atmospheric density.
Broad Implications for Vision Tasks
The principles developed for non-homogeneous dehazing—explicit density encoding, patch-based hierarchies, and spatially adaptive feature fusion—are fundamentally transferable to any vision task where environmental degradation varies across space.
For example, Du et al. applied spatial density estimation to the task of image deraining (Variational Image Deraining), demonstrating that treating rain streaks as a spatially uneven probabilistic inference problem prevents the over-smoothing of clean regions. Similarly, in probabilistic modeling, Im et al. introduced a Variational Bayesian framework that replaces constant scattering coefficients with spatially varying scattering parameters, providing a plug-and-play performance boost for existing networks.
By moving away from the mathematical convenience of the homogeneous medium assumption, these modern architectures achieve robust, real-world reliability by modeling the environment as it actually is: dynamic, localized, and non-uniform.