Haze Visibility Enhancement: A Survey and Quantitative Benchmarking

Yu Li, Shaodi You, Michael S. Brown, Robby T. Tan

Introduction

Images captured in outdoor environments often suffer from degraded visibility due to atmospheric phenomena such as haze, fog, and mist. These conditions are caused by particles in the air that scatter and absorb light, leading to a loss of contrast and color fidelity. The research paper "Haze Visibility Enhancement: A Survey and Quantitative Benchmarking" provides a comprehensive evaluation of the "dehazing" problem—the computational task of restoring clarity to these obscured images. While previous research often focused on individual algorithmic breakthroughs, this work categorizes decades of progress and establishes a rigorous numerical benchmark to objectively compare performance against ground truth.

Examples of atmospheric obscuration: Haze, Fog, and Mist.

The Physics of Haze

To understand dehazing, one must first understand the physical degradation process. Images taken in hazy conditions are primarily affected by two phenomena:

  1. Attenuation: The light reflected from the scene is absorbed or scattered by atmospheric particles before it reaches the camera sensor. This reduces the direct signal from the objects in the scene.
  2. Airlight: Ambient light (e.g., sunlight) is scattered into the camera's line of sight by the same atmospheric particles. This adds a "washout" effect or a veil of white/gray light over the image.
Illustration of light attenuation over distance x through a medium.

Mathematically, dehazing is an "ill-posed" problem. For every pixel in a hazy image, there is one known value but multiple unknowns: the true scene radiance, the depth of the object, and the global atmospheric light. Solving this requires "priors"—statistical assumptions about the properties of clear images. The challenge is that these assumptions often fail in specific scenarios, leading to artifacts like halos, color shifts, or over-saturation.

Categorizing Dehazing Methods

The authors categorize dehazing research into four primary domains:

  • Multi-image methods: These utilize multiple photographs of the same scene taken under different weather conditions to resolve ambiguities.
  • Polarizing filters: These exploit the physical property that airlight is often polarized, whereas scene reflection is typically not, allowing for hardware-based separation.
  • Known depth: These methods leverage external data, such as 3D models or GPS-based depth maps, to provide the missing distance information required by the physical model.
  • Single-image methods: Currently the most flexible and popular approach, these rely on sophisticated priors such as the Dark Channel Prior (DCP), which assumes that clear outdoor images contain pixels with very low intensity in at least one color channel. Other modern techniques include Color Lines (modeling pixel distributions in RGB space) and Machine Learning (using Convolutional Neural Networks like DehazeNet to learn haze features).

Quantitative Benchmarking and Insights

Historically, dehazing algorithms were evaluated qualitatively—essentially, "does it look better?" To move toward scientific objectivity, the authors introduced a new benchmark using Physically Based Rendering (PBRT). Unlike simple synthetic models, PBRT uses Monte Carlo ray tracing to simulate realistic light transport through volumes of particles, providing a high-fidelity "ground truth" for comparison.

The benchmark revealed several key insights:

  • Top Performers: The methods by Fattal (2014) and Berman et al. (2016) emerged as leaders in terms of Mean Absolute Difference (MAD) and Structural Similarity Index (SSIM).
  • Sensitivity of ML: Learning-based models like DehazeNet are highly sensitive to white balance. If the training data does not account for non-white atmospheric light, performance degrades significantly.
  • Non-Local Advantage: "Non-local" methods, which consider global image statistics rather than just local patches, are superior at handling depth discontinuities and avoiding "halo" artifacts common in older algorithms.

Future Frontiers

Despite significant progress, the survey identifies three critical areas for future research:

  1. Thick Fog: In extremely dense conditions, the signal-to-noise ratio is too low for current algorithms to recover meaningful scene information.
  2. Scattering-Induced Blur: Standard models assume light travels in straight lines. In reality, multiple scattering events cause a subtle blur that is not yet addressed by most dehazing equations.
  3. Non-Linear Camera Response: Most algorithms assume a linear relationship between light and pixel values, ignoring the non-linear "gamma correction" applied by real-world cameras. Integrating camera response functions into the dehazing process is a necessary step for real-world robustness.

This survey serves as a foundational roadmap for the field, shifting the focus from subjective enhancement to mathematically grounded restoration.

Summary updated

BibTeX
@article{li2017haze,
  author = {Yu Li and Shaodi You and Michael S. Brown and Robby T. Tan},
  title = {{Haze Visibility Enhancement: A Survey and Quantitative Benchmarking}},
  journal = {Comput. Vis. Image Underst.},
  year = {2017},
  doi = {10.1016/j.cviu.2017.09.003},
}
Figure