Source-linked AI summary
Dense Haze: A benchmark for image dehazing with dense-haze and haze-free images
Codruta O. Ancuti, Cosmin Ancuti, Mateu Sbert, Radu Timofte
TL;DR
Single-image dehazing lacks satisfactory validation because real hazy images rarely have corresponding haze-free references. The paper introduces Dense-Haze, a 33-pair real dataset of dense homogeneous outdoor haze, evaluates existing techniques, and finds that they perform poorly, leaving substantial room for improvement.
Problem
Validation of dehazing methods remains unsatisfactory because corresponding real hazy and haze-free ground-truth image pairs are scarce.
Method
The paper constructs Dense-Haze with 33 real hazy and haze-free outdoor image pairs, generating dense haze with professional machines while matching scene and illumination conditions.
Results
Existing single-image dehazing techniques perform poorly on Dense-Haze’s dense homogeneous hazy scenes in qualitative and quantitative evaluation.
Takeaways & Limitations
Dense-Haze provides a benchmark for assessing and advancing robust dehazing methods for real dense haze scenes.
Abstract
from arXiv · showhide
Single image dehazing is an ill-posed problem that has recently drawn important attention. Despite the significant increase in interest shown for dehazing over the past few years, the validation of the dehazing methods remains largely unsatisfactory, due to the lack of pairs of real hazy and corresponding haze-free reference images. To address this limitation, we introduce Dense-Haze - a novel dehazing dataset. Characterized by dense and homogeneous hazy scenes, Dense-Haze contains 33 pairs of real hazy and corresponding haze-free images of various outdoor scenes. The hazy scenes have been recorded by introducing real haze, generated by professional haze machines. The hazy and haze-free corresponding scenes contain the same visual content captured under the same illumination parameters. Dense-Haze dataset aims to push significantly the state-of-the-art in single-image dehazing by promoting robust methods for real and various hazy scenes. We also provide a comprehensive qualitative and quantitative evaluation of state-of-the-art single image dehazing techniques based on the Dense-Haze dataset. Not surprisingly, our study reveals that the existing dehazing techniques perform poorly for dense homogeneous hazy scenes and that there is still much room for improvement.
I. INTRODUCTION
Dense-Haze addresses weak validation in single-image dehazing by providing real hazy/haze-free references for dense, homogeneous outdoor scenes. It also evaluates existing methods and finds substantial difficulty under these conditions.
- Haze degrades visibility through contrast loss, selective spectral attenuation, and added noise, affecting applications such as traffic safety and visual surveillance.
- Single-image dehazing is mathematically ill-posed, while earlier approaches use polarization filters or depth knowledge priors.
- Existing validation is unsatisfactory because corresponding real hazy and haze-free ground-truth pairs are scarce; most datasets therefore synthesize haze using simplified optical models and known depth.
- Dense-haze conditions have received less attention than relatively light haze, limiting the utility of methods validated primarily on lighter scenes.
- Dense-Haze contains 33 pairs of real hazy and haze-free outdoor images featuring dense, homogeneous haze generated with professional haze machines.Scenes were recorded under controlled illumination conditions and extend the lighter-haze O-HAZE dataset.
- Evaluation with Dense-Haze shows existing techniques perform poorly on dense hazy scenes, leaving substantial room for improvement.The study combines qualitative and quantitative assessment using the dataset.
II. RECORDING DENSE-HAZE DATASET
The Dense-Haze recording protocol was designed to produce paired hazy and haze-free outdoor images under matched scene, camera, and illumination conditions. Operational constraints, especially low wind, made acquisition difficult and prolonged the recording process.
- Recording required pixel-level hazy and haze-free pairs from static scenes under identical conditions, including the same camera settings and viewpoint.
- Scenes were recorded during cloudy days, mornings, or sunsets to help preserve comparable illumination between hazy and haze-free captures.
- Wind speeds below 2-3 km/h were required to limit rapid haze spreading, contributing to the eight-week recording period for 33 outdoor scenes.
- Two professional 1500 W haze machines generated vapor particles typically 1-10 microns in diameter, using denser liquid to simulate haze over 20-30 meters.
- A tripod-mounted, remotely controlled Sony A5000 captured JPG and RAW 5456 × 3632 images, using the same manually adjusted camera settings for each paired scene.
- Exposure and white balance were calibrated with light meters and an 18% gray card, while a Macbeth color checker supported post-processing.
III. EVALUATED DEHAZING TECHNIQUES
The study evaluates representative prior- and learning-based single-image dehazing techniques, including methods based on transmission priors, color structure, fusion, and synthetic-data-trained CNNs. The accompanying tables define quantitative evaluation on sampled and full-dataset image sets.
- The evaluation covers representative single-image dehazing techniques using qualitative comparisons and quantitative measures on Dense-Haze.
- He et al. use the Dark Channel Prior to estimate transmission from low-intensity color-channel observations, then refine it with guided filtering in this evaluation.
- The comparison includes seven methods across seven Dense-Haze scenes, with hazy inputs, dehazed outputs, and ground truth shown in the corresponding figure.
- Meng et al. refine DCP transmission with a boundary constraint and weighted L1-norm regularization to improve resolution and reduce edge artifacts.
- Fattal and Berman et al. estimate transmission from one-dimensional color-line structures and clustered RGB color distributions, respectively.
- DehazeNet maps hazy to haze-free patches through sequential feature extraction, multi-scale mapping, local extrema, and nonlinear regression trained on synthesized data.
- Table I reports PSNR and CIEDE2000 for 10 randomly selected image sets, while Table II reports average SSIM, PSNR, and CIEDE2000 across all 33 sets.
IV. EVALUATION AND DISCUSSION
Dense-Haze evaluates representative single-image dehazing methods qualitatively and quantitatively on dense-haze outdoor scenes. The methods generally perform poorly, with structural and color distortions and relatively low metric performance leaving substantial room for improvement.
- The study evaluates seven representative single-image dehazing methods on 33 hazy and haze-free outdoor image pairs.The evaluated methods include He et al., Meng et al., Fattal, Cai et al., Ancuti et al., Berman et al., and Ren et al.
- DCP-based methods recover global structure but introduce color shifts in hazy regions, especially where airlight estimation is poor.The distortions are more significant in lighter or whiter regions, where the dark channel prior usually fails.
- Fattal and Berman et al. produce color artifacts, while learning-based methods trained on synthetic haze fail to remove the hazy white appearance.The learning-based approaches do not introduce additional distortions but remain unable to remove the haze appearance.
- All analyzed methods perform quite poorly on Dense-Haze scenes, with structural distortions near sharp transitions and artifacts more visible in distant regions.Some dehazed results also contain unnatural color distortions.
- PSNR, SSIM, and CIEDE2000 enable objective evaluation of pixel fidelity, structural similarity, and color preservation against haze-free ground truth.PSNR measures absolute errors, SSIM assesses normalized local intensity patterns, and lower CIEDE2000 values indicate better color preservation.
- He et al. and Meng et al. perform slightly better for structure and color restoration, but all methods achieve relatively low SSIM, PSNR, and CIEDE2000 performance.The results underscore the complexity of dehazing dense-haze scenes and leave significant qualitative and quantitative room for improvement.