In traditional image restoration, the goal is often to "clean" a signal—removing a light dusting of noise or a thin veil of mist. However, recent research has pivoted toward a much more difficult regime: extreme occlusion. In scenarios involving dense fog, heavy rain, or opaque stains, the underlying scene information isn't just distorted; it is often physically obliterated.
When the signal-to-noise ratio (SNR) drops toward zero, restoration ceases to be a filtering problem and becomes a reconstruction problem. To solve this, researchers are moving beyond simple pixel-to-pixel mapping, instead using recursive refinement, contrastive disentanglement, and generative priors to "hallucinate" missing structures from the few remaining visual cues.
The Information Loss Threshold
The fundamental challenge in extreme restoration is the loss of structural cues. As Yanyun Qu and colleagues note in their work on the Enhanced Pix2pix Dehazing Network, high haze density can obscure nearly all original scene geometry. This creates a mathematical "white-out" where the mapping from a hazy image to a clear one becomes ill-posed.
This phenomenon isn't limited to haze. In document restoration, Xuhang Chen and the team behind StainDoc point out that opaque ink or tea stains physically block the underlying text. Similarly, Kaige Wang observes that large raindrops in autonomous driving scenarios create total occlusion that is significantly harder to reconstruct than light streaks. When the input signal is this weak, models must rely on internal "knowledge" of what the world should look like to fill in the blanks.
Iterative and Recursive Refinement
One way to handle extreme degradation is to break the restoration into smaller, manageable steps. Rather than attempting a single-shot "miracle" reconstruction, models like T-Net, developed by Lirong Zheng and Kaihao Zhang, use a recursive "stacking" strategy. By iteratively refining the image and sharing weights across stages, the network can progressively peel back layers of haze, using the partially restored edges from one stage to guide the next.
A similar intuition drives the Iterative Filter Adaptive Network (IFAN) by Junyong Lee. For extreme defocus blur—where light from a single point is smeared across a wide area—IFAN uses separable filters that adaptively "tighten" the blur over multiple iterations. This prevents the model from being overwhelmed by the massive search space of a single-step deblurring operation.
Disentanglement and Contrastive Learning
In real-world scenarios, we rarely have "paired" data (the exact same photo with and without fog). This has led to a surge in unsupervised methods that learn to separate the "clean" scene from the "degradation" layer.
Xiang Chen and Zhentao Fan proposed CDD-GAN, which uses contrastive disentanglement learning. The model is forced to pull apart "task-relevant" features (the scene) and "task-irrelevant" features (the haze). By ensuring that the restored image is "far" from the hazy input but "near" the distribution of clear images in a latent space, the model learns to ignore the occlusion.
Deniz Engin and Hazim Kemal Ekenel took this further with Cycle-Dehaze, adding a cyclic perceptual-consistency loss. This ensures that if you "re-haze" a restored image, you should get back exactly what you started with. This constraint is vital when the haze is so dense that the model might otherwise "drift" and start generating a completely different scene.
Generative Priors and the Hallucination Boundary
When occlusion is total, the model must transition from restoration to inpainting. This is where generative models like Diffusion and GANs excel. Yunwei Lan and colleagues introduced Diff-Dehazer, which leverages the massive prior knowledge of Stable Diffusion.
The intuition here is that if a model has seen millions of clear outdoor photos, it knows that a blurry, white-ish blob behind a wall of fog is likely a building or a tree. By combining these generative priors with physical constraints—like the atmospheric scattering model—these systems can reconstruct plausible textures even when the original pixels are gone.
However, this comes with a risk. As Junkai Fan notes in Non-aligned Supervision for Real Image Dehazing, when information is physically lost, models may generate "hallucinated artifacts"—details that look real but don't actually exist in the original scene. The current frontier of this research is finding the balance: using enough generative power to see through the "white-out" while maintaining enough fidelity to ensure the restored image remains a faithful representation of reality.