Source-linked AI summary
I-HAZE: a dehazing benchmark with real hazy and haze-free indoor images
Codruta O. Ancuti, Cosmin Ancuti, Radu Timofte, Christophe De Vleeschouwer
TL;DR
Image dehazing lacks objective comparison because paired hazy and haze-free references are generally unavailable. The paper introduces I-HAZE, a 35-pair real-haze indoor dataset, and evaluates dehazing methods against its references using image-quality metrics, finding that leading methods still produce distortions and no single technique performs best on every image.
Problem
Dehazing methods are difficult to validate objectively because corresponding hazy and haze-free ground-truth images are generally unavailable.
Method
The paper constructs I-HAZE from 35 real hazy and haze-free indoor image pairs and compares state-of-the-art methods using reference-based image-quality metrics.
Results
No single technique performs best for all images, and the tested methods introduce structural and color distortions despite some methods performing best on average.
Takeaways & Limitations
I-HAZE enables objective comparison of dehazing results using known references and exposes substantial room for improving single-image dehazing.
Abstract
from arXiv · showhide
Image dehazing has become an important computational imaging topic in the recent years. However, due to the lack of ground truth images, the comparison of dehazing methods is not straightforward, nor objective. To overcome this issue we introduce a new dataset -named I-HAZE- that contains 35 image pairs of hazy and corresponding haze-free (ground-truth) indoor images. Different from most of the existing dehazing databases, hazy images have been generated using real haze produced by a professional haze machine. For easy color calibration and improved assessment of dehazing algorithms, each scene include a MacBeth color checker. Moreover, since the images are captured in a controlled environment, both haze-free and hazy images are captured under the same illumination conditions. This represents an important advantage of the I-HAZE dataset that allows us to objectively compare the existing image dehazing techniques using traditional image quality metrics such as PSNR and SSIM.
I. INTRODUCTION
Single-image dehazing is an ill-posed problem, yet existing validation remains unsatisfactory because corresponding hazy and haze-free references are generally unavailable. The paper introduces I-HAZE to provide real paired data for objective evaluation of dehazing methods.
- Motivation: Single-image dehazing estimates haze without scene geometry or atmospheric side information, making it mathematically ill-posed.Haze degradation varies by pixel and depends on scene-point distance through transmission coefficients.
- Prior approaches: Existing dehazing methods use priors such as surface-shading statistics, maximal contrast, dark channels, color attenuation, and learned mappings.These approaches estimate transmission using different assumptions or synthesized training data.
- Evaluation gap: Validation remains largely unsatisfactory because corresponding hazy and haze-free ground-truth images are absent, leading researchers to use non-reference assessment strategies.Examples include edge-gradient measures, subjective assessment, and visibility prediction.
- Evaluation gap: Existing datasets were built from synthesized haze because recording reference and hazy images under identical illumination is practically difficult.They rely on optical models and known depth, including synthetic road scenes and depth-derived real scenes.
- Contribution: I-HAZE contributes 35 pairs of real hazy and corresponding haze-free indoor images and evaluates state-of-the-art single-image dehazing methods using reference-based metrics.The dataset uses professional haze and supports SSIM and CIEDE2000 comparisons with known references.
II. I-HAZE DATASET
I-HAZE comprises 35 indoor domestic scenes containing varied objects and a Macbeth color checker. Each scene is recorded haze-free and then immediately under real haze using controlled capture procedures.
- Scene design: The dataset contains 35 indoor domestic scenes with objects differing in color and specularity.Every scene also includes a 24-square Macbeth color checker arranged in a 4×6 grid.
- Scene design: A Macbeth color checker is included in every scene for color-reference purposes.The checker measures 11 by 8.25 inches and contains 24 painted samples.
- Haze generation: For each scene, the haze-free ground truth is recorded first, followed immediately by haze introduction using two professional 1500 W fog machines.The machines generate water droplets approximately 1–10 microns in diameter, similar in size to atmospheric haze.
- Image capture: Haze-free and hazy images are captured with a stationary remotely controlled Sony A5000 on a tripod.The camera is calibrated in the haze-free scene and its aperture, shutter speed, ISO, and white balance are retained for the hazy scene.
III. QUANTITATIVE EVALUATION
The evaluation applies the Koschmieder image-formation model and compares multiple single-image dehazing methods against I-HAZE references. It uses visual comparisons and SSIM, PSNR, and CIEDE2000 measurements on selected and full-dataset images.
- Optical model: The Koschmieder model represents hazy pixel intensity using the haze-free image, transmission, and atmospheric light.Transmission describes depth-dependent attenuation, while atmospheric light is a color constant.
- Visual evaluation: Figure 2 compares hazy inputs, ground truth, and results from seven dehazing methods across five scenes.The methods shown are He, Meng, Fattal, Cai, Ancuti, Berman, and Ren.
- Compared methods: The evaluated methods include dark-channel-prior, Markov-random-field, CNN, local-airlight, and multiscale convolutional approaches.Several learned methods use synthetic hazy images generated from haze-free images and known depth maps.
- Quantitative evaluation: Table I computes SSIM and CIEDE2000 between ground truth and enhanced results for five randomly selected I-HAZE image sets.The corresponding hazy images, ground truth, and dehazed outputs are shown in Figure 2.
IV. RESULTS AND DISCUSSION
Evaluation against haze-free ground truth shows clear differences among dehazing methods, but persistent artifacts and low metric values indicate that no method consistently succeeds across all images.
- Qualitative comparison: He et al. recovers scene structure visually but introduces color shifts in hazy regions because of poor airlight estimation.The problem is especially visible with lighter close-up patches or small reflections.
- Qualitative comparison: Ancuti et al. produces more accurate results and fewer structural artifacts through local airlight estimation and multi-scale fusion.Berman et al. also produces visually compelling results with less color shifting than the color-lines method.
- Quantitative evaluation: The evaluation compares dehazed outputs directly with haze-free ground truth using PSNR, SSIM, and CIEDE2000.SSIM measures structural similarity, while CIEDE2000 evaluates color restoration.
- Quantitative evaluation: Berman et al., Ancuti et al., and Ren et al. perform best on average across SSIM, PSNR, and CIEDE indexes.Meng et al. and He et al. form a second group performing relatively well in structure and color restoration.
- Limitations of tested methods: All tested methods introduce structural distortions such as halo artifacts near edges, amplified in faraway regions, alongside possible color distortions.No single technique performs best for every image, and relatively low SSIM, PSNR, and CIEDE2000 values show that single-image dehazing remains difficult.