Source-linked AI summary
O-HAZE: a dehazing benchmark with real hazy and haze-free outdoor images
Codruta O. Ancuti, Cosmin Ancuti, Radu Timofte, Christophe De Vleeschouwer
TL;DR
Dehazing lacks reliable real-world evaluation because available datasets predominantly use synthetic haze and corresponding real haze-free references are difficult to capture. The paper introduces O-HAZE, a 45-scene outdoor benchmark of real hazy and haze-free pairs recorded under matched illumination, and uses it to evaluate representative methods. The evaluation finds that current techniques do not accurately recover the ground-truth images and continue to exhibit structural and color distortions.
Problem
Dehazing evaluation lacks corresponding real hazy and haze-free reference images, while existing assessment datasets rely on synthesized haze.
Method
O-HAZE provides 45 outdoor scenes with real hazy and corresponding haze-free images, and evaluates representative state-of-the-art single-image dehazing methods using objective image-quality metrics.
Results
Current dehazing techniques do not accurately reconstruct the original image from its hazy version, according to evaluations using SSIM and CIEDE2000.
Takeaways & Limitations
O-HAZE provides reference data for more representative outdoor dehazing assessment and supports comparison of methods using ground-truth images.
Takeaways & Limitations
All tested methods introduce structural artifacts and color shifting, and none recovers the ground-truth image.
Abstract
from arXiv · showhide
Haze removal or dehazing is a challenging ill-posed problem that has drawn a significant attention in the last few years. Despite this growing interest, the scientific community is still lacking a reference dataset to evaluate objectively and quantitatively the performance of proposed dehazing methods. The few datasets that are currently considered, both for assessment and training of learning-based dehazing techniques, exclusively rely on synthetic hazy images. To address this limitation, we introduce the first outdoor scenes database (named O-HAZE) composed of pairs of real hazy and corresponding haze-free images. In practice, hazy images have been captured in presence of real haze, generated by professional haze machines, and OHAZE contains 45 different outdoor scenes depicting the same visual content recorded in haze-free and hazy conditions, under the same illumination parameters. To illustrate its usefulness, O-HAZE is used to compare a representative set of state-of-the-art dehazing techniques, using traditional image quality metrics such as PSNR, SSIM and CIEDE2000. This reveals the limitations of current techniques, and questions some of their underlying assumptions.
1. Introduction
Single-image dehazing is an ill-posed problem whose evaluation is hindered by the lack of real hazy/haze-free reference pairs. O-HAZE addresses this gap with 45 outdoor scene pairs and evaluates current methods using objective metrics.
- Haze scatters and attenuates light, reducing distant-object visibility, local contrast, and spectral fidelity.
- Single-image dehazing is ill-posed because haze degradation varies by pixel and depends on scene-point distance from the camera.
- The Koschmieder model represents hazy image intensity using haze-free reflectance, transmission, and atmospheric light.
- Real-world evaluation remains unsatisfactory because corresponding haze-free references are generally unavailable, forcing reliance on non-reference quality assessment.
- O-HAZE introduces 45 outdoor scenes with real hazy and corresponding haze-free images, then compares state-of-the-art methods using SSIM and CIEDE2000.
2. Related Work
Prior dehazing methods use multiple images, atmospheric or depth cues, hardware, explicit priors, fusion, or learning-based approaches. Their assessment is limited by missing real reference images, synthetic data, and narrow existing real-scene datasets.
- Early methods address ill-posed dehazing through multiple atmospheric conditions, atmospheric cues, geometric depth models, or polarization filters.
- Single-image methods avoid special setups and side information, motivating approaches based on transmission priors, fusion, and machine learning.
- Dehazing assessment: Reliable algorithm assessment is impeded by the absence of haze-free ground truth captured under identical scene illumination.
- Dehazing assessment: Existing evaluation uses non-reference metrics or synthesized haze generated from optical models and known depth.
- Dehazing assessment: O-HAZE extends real-scene benchmarking to 45 controlled-illumination outdoor scenes, complementing CHIC’s two indoor scenes.
3. O:HAZE: Recording the outdoor hazy scenes
O-HAZE records 45 outdoor scenes with and without machine-generated haze under matched camera settings and controlled environmental conditions. The protocol is designed to preserve correspondence while approximating outdoor hazy conditions.
- The dataset contains 45 outdoor scenes recorded both with and without haze, covering visibility from camera proximity to distances up to 30m.
- Recording spanned more than 8 autumn weeks during cloudy mornings or sunsets, with wind speeds below 3 km/h to limit haze spreading.
- Hazy and haze-free acquisitions used identical shutter speed, aperture, ISO, and white-balance settings.
- A gray card and Macbeth color checker supported manual white-balance setup and post-processing of the recorded images.
- Two professional 1500 W haze machines generated 1–10 micron vapor particles, with denser liquid chosen to simulate water haze over 20–30m.
4. Evaluated Dehazing Techniques
The paper evaluates seven single-image dehazing techniques spanning explicit priors, fusion, and CNN-based approaches, using O-HAZE comparisons against haze-free references.
- Explicit-prior methods: He et al. estimate transmission with the dark channel prior and refine it using guided filtering to reduce artifacts.The dark channel prior assumes most local regions contain a low-intensity pixel in at least one color channel, except sky or hazy regions.
- Explicit-prior methods: Meng et al. combine a DCP-based transmission estimate with a boundary constraint and weighted L1 regularization.The method targets limited transmission-map resolution and reduces halo artifacts around sharp edges relative to He et al.
- Explicit-prior methods: Fattal estimates transmission from RGB color-lines and refines it with a Markov random field to filter noise and scattering artifacts.
- Learning-based methods: Cai et al. use an end-to-end CNN mapping hazy patches to haze-free patches through feature extraction, multi-scale mapping, local extrema, and nonlinear regression.The network is trained on synthesized hazy images.
- Fusion and color-consistency methods: Ancuti et al. estimate local airlight and fuse multiple image versions, while Berman et al. infer transmission from haze-lines formed by RGB color clusters.Ancuti et al. developed their approach for difficult night-time conditions, whereas Berman et al. exploit color consistency across non-local pixels.
- Learning-based methods: Ren et al. estimate transmission with a coarse-to-fine multi-scale CNN trained on synthetically generated hazy images.Training uses haze-free images and associated depth maps with a simplified light propagation model.
5. Evaluation and Discussion
Using O-HAZE scenes and haze-free references, the evaluation compares visual and objective performance across dehazing methods. Results show method-specific artifacts, no clear class-wide winner, and substantial room for improvement.
- Experimental comparison: Seven methods are compared on 11 randomly selected O-HAZE scenes, with hazy inputs, haze-free ground truth, and dehazed outputs shown side by side.The comparison includes He, Meng, Fattal, Cai, Ancuti, Berman, and Ren.
- Qualitative results: He et al. recover image structure well but introduce color shifts in hazy regions, especially lighter regions where the dark channel prior fails.Meng et al. improve these results through more accurate transmission estimation, while Fattal introduces color artifacts.
- Qualitative results: Ancuti et al. produce higher contrast and more intense colors with a slight yellow/red shift, while Berman et al. yield sharper edges with fewer color artifacts.
- Qualitative results: Ren et al. generate visually more compelling results than Cai et al. among the learning-based approaches.
- Discussion: No clear winner emerges between learning-based and explicit-prior methods: all tested methods introduce structural or color distortions, though their artifacts and benefits differ.The authors identify complementarity between methods such as Ancuti et al. and Berman et al., leaving room for combinations.
- Quantitative evaluation: Meng, Berman, Ancuti, and Ren perform best on average across SSIM, PSNR, and CIEDE indices, while He, Fattal, and Cai are less competitive.The averages are computed over all 45 O-HAZE scenes; none of the techniques performs best for every image.
- Discussion: Relatively poor SSIM, PSNR, and CIEDE2000 values confirm the difficulty of single-image dehazing and indicate substantial room for improvement.