Most computer vision models for clearing atmospheric interference—like fog, haze, or rain—suffer from a "synthetic trap." Because it is nearly impossible to capture perfectly aligned pairs of hazy and clear images in the wild, researchers traditionally rely on the Atmospheric Scattering Model (ASM) to synthesize training data.
The ASM assumes that a hazy image is a linear combination of the clear scene and the "atmospheric light," modulated by a transmission map. While mathematically elegant, this model is a drastic simplification. In the real world, haze is rarely homogeneous; it is a stochastic cocktail of varying particle sizes, non-uniform densities, and complex light scattering that the ASM fails to capture. This creates a significant domain gap, where models that perform perfectly on synthetic benchmarks fail when deployed on a real, smoggy street.
Embedding Physics into the Architecture
Rather than treating the neural network as a black box, a prominent direction involves explicitly embedding physical laws into the network's layers. This approach, often termed differential programming, ensures the model remains grounded in the optics of the real world.
In DehazeGAN: When Image Dehazing Meets Differential Programming, Hongyuan Zhu and colleagues moved away from pure black-box estimation. By embedding the physical scattering model directly into a Generative Adversarial Network (GAN), the architecture is forced to jointly learn physical parameters and image recovery. Similarly, Jin Xie and Xuelong Li proposed HRGAN, which unifies the estimation of atmospheric light and transmission maps within a single framework. The intuition here is that by constraining the network with physical priors, it becomes less likely to produce the "hallucinated" artifacts common in purely data-driven models.
Handling Non-Uniformity and Scale
Real-world weather is "patchy." A single transmission map often fails to account for localized fog or heavy smoke. To address this, researchers have turned to multi-scale and progressive architectures that refine the image in stages.
Hang Dong and Jinshan Pan introduced the Multi-Scale Boosted Dehazing Network (MSBDN), which uses a "Strengthen-Operate-Subtract" strategy to progressively recover spatial details. This mirrors how humans might clean a window—scrubbing away the thickest grime first before polishing fine details. For rain, Xin Yu proposed using the residual between the restored and rainy image as an attention map, allowing the model to focus its "energy" on rain streaks of varying densities without needing explicit classification.
Bridging the Gap with Generative Priors
When physical models are too simple and real data is too scarce, a third path emerges: using high-quality generative priors to "hallucinate" realistic textures that the ASM would otherwise wipe away.
Recent work has leveraged foundation models to provide these priors. Ruiyi Wang and Xiaohong Liu developed HazeCLIP, which uses the visual-language priors of a frozen CLIP model to guide dehazing. Because CLIP has "seen" millions of real-world images, it understands what a clear sky or a sharp building edge should look like, even if the training data is synthetic.
In a similar vein, RIDCP by Ruiqi Wu and Chongyi Li uses a pre-trained VQGAN codebook. Instead of trying to "math" their way out of the haze, they treat dehazing as a matching problem: finding the high-quality "clean" discrete priors that best correspond to the hazy features.
From Restoration to Understanding
A critical realization in this field, highlighted by Yanting Pei and Yaping Huang, is that visual quality does not always equal machine utility. An image that looks "clean" to a human might have lost the subtle gradients a classification model needs.
This has led to "unified" frameworks like those proposed by Boyi Li and Zhangyang Wang, which jointly optimize for dehazing and object detection. By training the restoration model to satisfy a downstream detector, the system learns to prioritize the features that matter for real-world applications like autonomous driving, where Christos Sakaridis and Dengxin Dai have shown that curriculum adaptation from clear to foggy scenes is essential for safety.