Source-linked AI summary
Haze Visibility Enhancement: A Survey and Quantitative Benchmarking
Yu Li, Shaodi You, Michael S. Brown, Robby T. Tan
TL;DR
Visibility degradation in hazy and foggy images impairs contrast and affects computer-vision algorithms that assume clear scenes. The paper surveys dehazing methods and benchmarks representative single-image techniques quantitatively and qualitatively, finding that recent prior-based methods generally perform best and are recommended for robustness in practice.
Problem
Hazy and foggy scenes degrade image contrast and visibility, limiting algorithms designed for clear-weather images and motivating dehazing methods.
Method
The paper surveys dehazing methods across imaging settings and quantitatively evaluates representative single-image methods using ground-truth datasets and SSIM, alongside qualitative real-image comparisons.
Results
Recent prior-based methods, especially Fattal 14 and Berman 16, generally perform best across the benchmark and consistently provide excellent dehazing results on real hazy images.
Takeaways & Limitations
The benchmark supports preferring prior-based methods over learning-based methods for practical robustness, while the dark channel prior is especially effective for denser haze.
Takeaways & Limitations
The visibility-enhancement limit in sufficiently thick fog remains an open theoretical question, and nonlinear camera tone curves complicate physically correct scene-reflection recovery.
Abstract
from arXiv · showhide
This paper provides a comprehensive survey of methods dealing with visibility enhancement of images taken in hazy or foggy scenes. The survey begins with discussing the optical models of atmospheric scattering media and image formation. This is followed by a survey of existing methods, which are grouped to multiple image methods, polarizing filters based methods, methods with known depth, and single-image methods. We also provide a benchmark of a number of well known single-image methods, based on a recent dataset provided by Fattal and our newly generated scattering media dataset that contains ground truth images for quantitative evaluation. To our knowledge, this is the first benchmark using numerical metrics to evaluate dehazing techniques. This benchmark allows us to objectively compare the results of existing methods and to better identify the strengths and limitations of each method.
I. INTRODUCTION
Haze and fog degrade image contrast and visibility through atmospheric scattering, motivating visibility-enhancement methods. The paper surveys these methods and benchmarks single-image techniques using datasets with ground truth.
- Visibility degradation affects computer-vision algorithms that assume clear scenes, making dehazing practically important.
- Atmospheric particles absorb and scatter light, weakening direct transmission and adding airlight that washes out scenes.
- Existing visibility-enhancement methods are grouped into multi-image, polarizing-filter, known-depth, and single-image approaches.
- The paper surveys scattering-media visibility enhancement, emphasizing single-image methods because they require no additional information.
- The paper quantitatively benchmarks single-image methods using Fattal’s dataset and a newly generated physically based scattering-media dataset.
- The paper concludes by discussing benchmark findings, current limitations, and future research directions.
A. Optical Modeling
The optical model explains visibility degradation as the combination of attenuated scene transmission and scattered airlight. It relates attenuation to particle scattering, distance, and atmospheric conditions.
- Atmospheric particles scatter incident light in directions determined by particle properties and illumination wavelength.
- Direct Transmission: The Beer-Lambert law models direct transmission as exponential attenuation with distance and atmospheric scattering coefficient.
- Airlight: Airlight is environmental illumination scattered toward the camera from sunlight, skylight, ground reflections, and related sources.
- Image Formation: The image-formation model combines direct transmission with airlight under an assumed linear relationship between incoming light and pixel values.
- Image Formation: Transmission is t(x) = e^−βd(x), where β is atmospheric attenuation and d(x) is object-to-camera distance.
- Image Formation: Dehazing estimates scene reflection, the component intended to represent the scene without medium-particle effects.
III. SURVEY
The survey organizes research on vision through atmospheric media into five categories according to input and output, presented chronologically.
- The survey covers depth estimation, multi-image dehazing, polarizing-filter dehazing, known-depth dehazing, and single-image dehazing.
A. Early Work in Depth Estimation
Early methods estimated scene depth from atmospheric effects using known reflection, multiple images, airlight, or color decomposition. Their depth estimates depend on assumptions about atmospheric parameters and available inputs.
- Cozman-Krotkov estimates atmospheric light from sky regions and transmission using an assumed scene reflection.
- Absolute depth remains unknown when the atmospheric attenuation coefficient β is unknown; a known reference depth can recover absolute depth.
- Nayar-Narasimhan proposed three depth-estimation algorithms using different scattering conditions, airlight, or color decomposition.
- The multi-image approach estimates relative depth from intensity ratios between images captured under different scattering coefficients.
- The relative-depth estimate from the light-source method applies only to light-source regions rather than the entire image.
- The airlight-based algorithm assumes atmospheric light and achromatic airlight are available, but obtaining airlight is its stated problem.
- The dichromatic scattering model decomposes images into direct-transmission and airlight chromaticities to support depth estimation.
B. Multiple Images
Multiple-image methods exploit changes across observations—haze density, weather, viewpoint, or video frames—to estimate atmospheric effects, depth, transmission, and restored visibility. Their assumptions range from known atmospheric conditions or segmentation to stereo geometry and iterative monocular estimation.
- Multiple haze densities: Different haze densities let Narasimhan–Nayar form intersecting RGB planes from two same-scene images to estimate atmospheric light.The images share atmospheric-light color but have different direct-transmission colors.
- Parameter requirements: The multi-image methods estimate scene structure from direct-transmission ratios, but some require reference transmission values or known atmospheric-light parameters.A known transmission pixel can provide a reference point, while other formulations require atmospheric-light values obtained from auxiliary equations.
- Different atmospheric conditions: Iso-depth contrast restoration assumes depth segmentation and known atmospheric light, then rescales same-depth regions after estimating atmospheric effects from two gray images.Pixels with equal depth form lines in a 2D image-intensity space whose intersection estimates atmospheric-light parameters.
- Stereo images: Stereo-based dehazing jointly estimates disparity and visibility using stereo and fog data terms, alternating their optimization until convergence.The method combines smoothness priors with iterative estimation of fog-related radiance and stereo disparity.
- Monocular video: Monocular-video dehazing jointly estimates depth and transmission with photoconsistency and Laplacian smoothness, producing depth maps and defogged images after iteration.Synthetic intensities are generated from estimated transmission and inverse depth for neighboring-frame comparisons.
C. Polarizing Filter
Polarizing-filter methods avoid waiting for haze density to change by acquiring differently polarized observations. They estimate airlight from polarization behavior, recover transmission, and address missing sky regions through ICA, while retaining a key separation challenge.
- Motivation: Polarization-based capture replaces changing fog density with multiple images taken through rotating polarizing filters.This addresses the requirement for same-scene images under substantially different particle densities.
- Polarization model: The method assumes direct transmission is weakly polarized, so polarized airlight dominates the observed intensity variation.Maximum and minimum intensities correspond to the filter passing or blocking airlight most effectively.
- Dehazing procedure: Summed polarization states provide the observed image, from which atmospheric light, per-pixel airlight, transmission, and the dehazed result are computed.The filter must be rotated automatically or manually to obtain the maximum and minimum intensity states.
- ICA-based estimation: A later method removes the sky-region requirement by estimating airlight and direct transmission with independent component analysis.Its central challenge is estimating the transformation that maps maximum and minimum polarized intensities to the two components accurately.
D. Known Depth
Known-depth methods use scene geometry, 3D models, or strong flat-world assumptions to estimate transmission and atmospheric light. These approaches can restore visibility, but their applicability is constrained by depth availability, reflectance assumptions, or scene structure.
- Single-image parameter estimation: Oakley–Satherley estimates physical-model parameters by minimizing intensity error, then applies contrast enhancement in a two-stage single-image procedure.Its parameter estimation approximates scene reflectance with a mean value.
- Limitations: Known-depth approaches remain limited by required depth information and reflectance simplifications, including whole-image mean reflectance in parameter estimation.The mean-reflectance assumption is considered acceptable for uniform-color scenes but not general scenes.
- Geometric assumptions: Flat-world methods estimate attenuation and depth from image geometry using known camera properties such as height and horizon position.The formulation replaces scattering-model depth with a vertical image-coordinate relationship under the flat-ground assumption.
- Geometric assumptions: The Oakley–Satherley approach works best for scenes dominated by flat planes because its depth heuristics do not generally fit trees, vehicles, houses, or other objects.The paper explicitly identifies rural road scenes as a setting where the method can operate at full potential.
- Exact 3D models: Kopf et al. combines an exact 3D scene model with Landsat textures to estimate transmission, atmospheric light, and a global correction vector.Average model-texture and image haze intensities are compared at corresponding depths, with distant pixels used for atmospheric-light estimation.
E. Single-Image Methods
Single-image dehazing methods estimate visibility from one hazy image using statistical, prior-based, optimization, fusion, or learned cues. The survey also records practical boundaries, including assumptions about scene statistics, artifacts, unexplained design choices, and computational constraints.
- Tan 2008: Tan’s method estimates atmospheric light from the brightest pixels and normalizes image color before dehazing.Its success depends on an optimum statistical decomposition of shading and transmission, and it was not tried on foggy scenes.
- Fattal 2008: Fattal derives shading and transmission by assuming they are locally and statistically uncorrelated.The method estimates transmission and atmospheric light to recover the scene reflection.
- He et al. 2009: The dark channel prior estimates transmission from the observation that clear outdoor scenes usually contain a locally dark color channel.Atmospheric light is selected from the top 0.1% brightest dark-channel pixels, followed by matting to smooth transmission and reduce halo effects.
- Tarel-Hautiere 2009: Tarel and Hautiere introduce a visibility-restoration method with complexity linear in the number of image pixels.It rapidly estimates normalized airlight using a heuristic bounded by W(x), then applies bilateral or median filtering for smoothness.
- Kratz-Nishino 2009: Kratz and Nishino formulate dehazing as a factorial MRF that estimates depth and albedo as statistically independent hidden fields.Their interpretation connects the dark channel prior to initialization from an upper-bound depth estimate rather than only to shadows or color variation.
- Ancuti-Ancuti 2010: Image-fusion methods combine a white-balanced image and a globally contrast-enhanced image using luminance, chromaticity, and saliency weights.The rationale for the two inputs and three weights is not fully explained and requires further investigation.
- Learning and optimization methods: CNN-based DehazeNet uses feature extraction, multi-scale mapping, local extremum, and nonlinear regression to estimate dehazing features.Other surveyed methods use local contrast or MRF optimization, but can produce halos at depth discontinuities or fail when texture is absent.
IV. QUANTITATIVE BENCHMARKING
The benchmark focuses on recent single-image dehazing methods because they are more practical and potentially applicable than approaches requiring additional information. Quantitative comparison requires datasets with ground-truth clear images, but collecting matched real atmospheric scenes across conditions is difficult.
- Benchmark scope: The benchmark focuses on recent single-image methods because they are more practical and have greater potential for applications.The authors argue that quantitative benchmarking makes method comparisons available numerically.
- Dataset requirements: Ground-truth evaluation requires datasets containing clear reference images paired with hazy observations.Ideally, such data would cover atmospheric conditions from light mist to dense fog across varied scenes.
- Evaluation setup: The study compares 11 dehazing methods using author code when available, reproduced implementations for several methods, and supplied results for others.Evaluation uses representative methods published in major venues.
A. Evaluation on Fattal’s Dataset [1]
The benchmark evaluates transmission estimation and final dehazing on Fattal’s synthetic dataset across haze levels and finds newer methods generally more accurate, with Fattal 14 strongest overall.
- Evaluation setup: The evaluation measures transmission-map MAD against ground truth, excluding sky regions, and final dehazing error against clean images.The benchmark directly uses ground-truth transmission and clean-image references for quantitative comparison.
- Transmission estimation: Fattal 14 and Berman 16 provide the most accurate transmission estimates in most cases, while no method is best for every image.Fattal 08 is unstable across scenes, and Tan 08 is generally less precise.
- Signed-error analysis: Mean signed differences reveal directional errors: Tan 08 tends to underestimate transmission, whereas Fattal 08 tends to overestimate it.Underestimation is associated with oversaturated outputs, while overestimation leaves residual haze.
- Final dehazing: Fattal 14 and Berman 16 produce dehazed images closest to ground truth, while newer model-based methods generally outperform early methods.The quantitative ranking is consistent with the visual quality shown for the church example.
- Haze-level evaluation: Fattal 14 consistently achieves the lowest errors in both transmission estimation and final dehazing across haze levels.Fattal 08 performs well only at low haze, whereas its performance declines at medium and high haze.
- Haze-level evaluation: He 09 and Fattal 14 retain relatively low final-dehazing error under heavy haze, when most methods incur large error.He 09 and Meng 13 slightly underestimate transmission because they predict a lower bound.
B. Evaluation on our Dataset
The authors construct a physically rendered scattering-media dataset with haze-free ground truth and evaluate nine methods using SSIM across increasing haze levels. Newer methods generally perform better, with He 09 particularly strong under heavier haze.
- Dataset construction: The dataset uses physically based rendering with Monte Carlo ray tracing in volumetric scattering media across five scene types and haze levels.Scenes use freely available 3D models and include substantial depth variation.
- Dataset construction: Haze-free renders serve as ground truth, while additional renders increase absorption and scattering coefficients to produce denser haze.Visibility, especially for distant objects, decreases as haze level increases.
- Evaluation setup: Nine methods are evaluated by comparing their outputs with corresponding ground-truth images using SSIM.SSIM is used because it is more consistent with human visual perception than pixelwise MAD in dehazing cases.
- Results: Tang 14, Cai 16, and Berman 16 generally outperform the other methods on SSIM, while He 09 performs especially well at heavier haze levels.This pattern is consistent with the Fattal-dataset experiment.
C. Qualitative results on real images
Qualitative results on real hazy images reinforce the benchmark findings: Fattal 14 and Berman 16 are the most consistently effective, while other methods exhibit distinct visual artifacts or residual haze.
- Qualitative comparison: Fattal 14 and Berman 16 consistently provide excellent dehazing results on real hazy images.The qualitative comparison agrees with the earlier quantitative benchmarking.
- Qualitative comparison: Kratz 09, Tarel 09, and Ancuti 13 show limitations including oversaturation, boundary artifacts, and color shifts.Tang 14 and Cai 16 generally produce appealing results but tend to leave haze in the outputs.
V. SUMMARY AND DISCUSSION
The paper surveys visibility-enhancement methods across multiple imaging modalities and emphasizes single-image dehazing. Its benchmark finds stronger performance from recent prior-based methods, while also identifying limits in dense fog, scattering blur, and camera tone-curve effects.
- The survey covers multiple images, known approximated depth, stereo, polarizing filters, and single-image methods for visibility enhancement.
- Recent methods, generally perform better in dehazing tasks, while learning-based methods, can produce decent results but may be affected by white balancing.
- The authors recommend prior-based methods, over learning-based methods, in practical use for robustness, and find the dark channel prior effective especially for denser haze.
- The benchmark evaluates representative single-image dehazing methods using Fattal’s dataset and a newly introduced synthetic dataset with ground-truth clean and hazy images.
- Thick fog makes visibility enhancement harder because low transmission and dominant airlight hide scene-reflection information.
- Absorption and scattering away from the line of sight produce image blur that the two-flux scattering model does not model.
- Camera gamma correction and tone curves break the model’s assumed linear relation between incoming light flux and pixel intensities, affecting physically correct scene-reflection recovery.