When capturing images in adverse conditions—such as pitch-black nights, murky underwater environments, or heavy fog—the resulting images suffer from severe visibility degradation. Historically, computer vision treated low-light enhancement and atmospheric dehazing as entirely separate problems governed by different physical principles.
However, a profound mathematical symmetry connects these domains: intensity inversion duality. By inverting the pixel values of an underexposed or dark image, the mathematical behavior of the darkness transforms into a perfect analogue for atmospheric haze or a virtual absorption medium. This duality allows researchers to repurpose powerful, physics-based scattering models to clear away darkness as if it were fog.
The Core Intuition: Turning Darkness into Haze
To understand why this duality works, consider what happens physically during atmospheric scattering versus underexposure. In a hazy scene, light from objects is scattered and attenuated by particles in the air, while ambient light is scattered into the camera sensor, creating a bright, obscuring "atmospheric veil." In a low-light scene, the problem appears opposite: there is a severe deficit of light due to absorption or lack of illumination.
As discovered by Galdrán et al. and Guo et al., if you mathematically invert a low-light image (subtracting each pixel value from a maximum constant), the dark regions become bright veils, and the light attenuation behaves identically to light scattering. Specifically, when the atmospheric light constant is set to unity, the mathematical equations for low-light inversion align precisely with standard atmospheric scattering models (Guo et al.).
Instead of designing a completely new framework to estimate non-uniform illumination, researchers can treat darkness as a virtual scattering or absorption medium (Kar et al.; Liu et al.). Dehazing algorithms can then be applied directly to this inverted input to "remove" the virtual veil.
The Mathematical Bridge: Retinex and Dehazing
This inversion mechanism bridges two prominent image processing frameworks: Retinex theory (which models human color constancy and local contrast) and Koschmieder's atmospheric scattering model (used for dehazing).
The authors of Galdrán et al. demonstrated that traditional dehazing heuristics, such as the Dark Channel Prior (DCP), are actually simplified instances of Retinex theory operating under local contrast estimation. Because of this underlying relationship:
- Retinex algorithms possess an inherent brightness-increasing bias. When applied to an inverted hazy image, this bias effectively subtracts the additive light component caused by fog (Galdrán et al.).
- Dehazing algorithms can be run in reverse. When applied to an inverted low-light image, they correct lighting irregularities as if they were atmospheric veils, restoring visibility by modeling light absorption as the inverse of scattering (Shi et al.; Galdrán et al.).
The Inversion Pipeline
In practice, this duality yields a highly efficient, three-step pipeline for restoring single images without requiring complex paired training data:
[Low-Light Input] ──> (Pixel Inversion) ──> [Virtual Haze Image]
│
▼
[Enhanced Output] <── (Pixel Inversion) <── [Dehazed Image]
- Inversion: The input pixel intensities of an underexposed or low-light image are inverted (Marques et al.; Li et al.).
- Dehazing: A standard scattering-based dehazing algorithm or transmission estimation model is applied to this inverted image, stripping away the "haze" (the inverted darkness) (Marques et al.; Liu et al.).
- Re-inversion: The restored image is inverted back to the original space, producing a naturally brightened scene with recovered contrast and details (Liu et al.; Li et al.).
This pipeline has proven highly versatile. For instance, Marques et al. utilized this exact inversion logic to restore visibility in non-uniform, low-light underwater environments by treating darkness as a scattering phenomenon. Similarly, Liu et al. leveraged radiative transfer equations—typically reserved for hazy or underwater media—to clear away virtual absorption media in neural radiance fields ($I^2$-NeRF), achieving superior visual fidelity in underexposed 3D scenes.
Symmetry Beyond Low Light: Correcting Overexposure
The elegance of intensity inversion duality is that it is truly bidirectional. Just as underexposure can be treated as a dehazing problem via inversion, overexposure can be treated as an underexposure problem.
Zhang et al. established that overexposed regions are the mathematical inverse of underexposed regions. This symmetry implies that specialized low-light enhancement techniques can be generalized to highlight recovery through simple image inversion.
By processing the inverted version of an image, blown-out highlights are transformed into manageable shadows (Wang et al.). This allows a unified architecture or a single illumination estimation framework to correct both lighting extremes simultaneously, fixing pitch-black shadows and overexposed highlights using the exact same underlying computational logic (Zhang et al.; Wang et al.).
Key Foundational Work
On the Duality Between Retinex and Image Dehazing (2018) — contributed 8 concept elements to this direction.
Nighttime low illumination image enhancement with single image using bright/dark channel prior (2018) — contributed 2 concept elements to this direction.
LIME: A Method for Low-light IMage Enhancement (2016) — contributed 1 concept elements to this direction.