What Is Image Dehazing?

Image dehazing is an image processing task that recovers a clear scene from a photograph degraded by atmospheric particles such as fog, mist, smoke, or airborne dust. Haze reduces contrast, shifts colors toward the ambient light, and suppresses fine detail in proportion to distance, so a hazy photograph loses information unevenly across the frame. The problem is treated as a restoration task rather than an enhancement task, because the goal is to estimate the radiance the camera would have recorded in clear air rather than simply to make the picture look better. Dehazing matters in engineering contexts because most outdoor vision systems, including surveillance cameras, driver assistance sensors, and satellite instruments, degrade sharply once scene contrast collapses.

The Atmospheric Scattering Model

Nearly all dehazing work builds on the same physical model, in which the observed intensity at a pixel is the sum of two terms: the scene radiance attenuated by transmission through the medium, and airlight, the atmospheric illumination scattered into the line of sight. Transmission falls exponentially with the product of scene depth and a scattering coefficient, which is why haze thickens with distance. The model has three unknowns per pixel, the true radiance, the transmission, and the global atmospheric light, against a single observed color, so the inversion is underdetermined. Every method therefore introduces additional assumptions, whether a statistical prior on natural images, multiple observations of the same scene, or a learned mapping fitted on paired data.

Prior-Based Methods

The most influential single-image approach constrains the problem with the dark channel prior, introduced in Single Image Haze Removal Using Dark Channel Prior. The observation is that in most local patches of a haze-free outdoor photograph, at least one color channel contains pixels with very low intensity, so a bright dark channel indicates haze and its magnitude estimates transmission directly. Refining the resulting coarse transmission map with soft matting or a guided filter produces a per-pixel estimate that yields both a restored image and, as a by-product, a usable depth map. The prior fails predictably where it is violated: over sky regions, snow, and large white or bright surfaces, transmission is underestimated and the restored image acquires artifacts. Later variants pair it with a bright channel prior, adapt the patch size to local structure, or add regularization drawn from natural image statistics.

Learning-Based Methods

Data-driven methods replace the hand-designed prior with a network trained to estimate transmission, atmospheric light, or the clear image end to end. Because matched hazy and haze-free pairs of real scenes are difficult to capture, training relies heavily on synthetic haze rendered onto depth-annotated photographs, which raises the question of how well a model transfers to genuine atmospheric conditions. The RESIDE benchmark described in Benchmarking Single Image Dehazing and Beyond addressed this by combining synthetic and real-world hazy images and evaluating with full-reference metrics, no-reference metrics, subjective ratings, and downstream task performance. Recent work has pushed toward lighter architectures suitable for real-time use, including the efficient designs examined in work on single image dehazing and desnowing, which treat haze and snow as related degradations under a shared framework.

Applications

Image dehazing has applications across imaging and vision systems, including:

• Video surveillance and traffic monitoring in fog-prone locations
• Advanced driver assistance and autonomous vehicle perception
• Remote sensing, where atmospheric correction restores usable surface reflectance
• Aerial and drone photography over long viewing distances
• Maritime and aviation navigation aids under reduced visibility
• Consumer photography, as a contrast and color restoration step