Source-linked AI summary
NH-HAZE: An Image Dehazing Benchmark with Non-Homogeneous Hazy and Haze-Free Images
Codruta O. Ancuti, Cosmin Ancuti, Radu Timofte
TL;DR
Image dehazing lacks objective evaluation because paired haze-free references are difficult to obtain, while existing realistic datasets generally assume homogeneous haze. This work introduces NH-HAZE, a realistic non-homogeneous paired dataset, and evaluates single-image methods on it. The assessment finds that CNN-based methods generally perform better, although no technique is best on every image.
Problem
Image dehazing lacks standardized objective evaluation because collecting paired hazy and haze-free images under identical conditions is burdensome, and existing realistic datasets assume homogeneous haze.
Method
The work introduces NH-HAZE with realistic non-homogeneous haze and haze-free outdoor pairs, then evaluates dehazing methods quantitatively using PSNR and SSIM.
Results
CNN-based methods generally perform better than the other considered techniques, while none performs better than all others on every image.
Takeaways & Limitations
NH-HAZE facilitates objective quantitative evaluation of dehazing methods on realistic scenes with non-homogeneous haze.
Takeaways & Limitations
The non-homogeneous haze makes dehazing challenging, with analyzed techniques producing low SSIM and PSNR values.
Abstract
from arXiv · showhide
Image dehazing is an ill-posed problem that has been extensively studied in the recent years. The objective performance evaluation of the dehazing methods is one of the major obstacles due to the lacking of a reference dataset. While the synthetic datasets have shown important limitations, the few realistic datasets introduced recently assume homogeneous haze over the entire scene. Since in many real cases haze is not uniformly distributed we introduce NH-HAZE, a non-homogeneous realistic dataset with pairs of real hazy and corresponding haze-free images. This is the first non-homogeneous image dehazing dataset and contains 55 outdoor scenes. The non-homogeneous haze has been introduced in the scene using a professional haze generator that imitates the real conditions of hazy scenes. Additionally, this work presents an objective assessment of several state-of-the-art single image dehazing methods that were evaluated using NH-HAZE dataset.
1. Introduction
Image dehazing lacks standardized, objective evaluation because paired hazy and haze-free reference images are difficult to collect. NH-HAZE addresses this gap with realistic non-homogeneous haze, paired outdoor images, and quantitative assessment of state-of-the-art methods.
- Image dehazing is ill-posed and aims to restore image content affected by haze.
- The field lacks standardized benchmarks and reference haze-free images for objective, quantitative performance evaluation.Collecting paired images under identical conditions is burdensome.
- Existing realistic datasets introduced through NTIRE challenges assume homogeneous haze across each scene.
- NH-HAZE is the first realistic dataset with non-homogeneous hazy and haze-free paired images, containing 55 outdoor pairs.A professional haze generator imitates real haze conditions.
- The work objectively evaluates several state-of-the-art single-image dehazing methods using ground-truth images and PSNR and SSIM.
2. Related Work
Prior image dehazing methods use physical cues, statistical priors, visibility restoration, or learned models. Existing datasets and evaluation practices remain constrained by synthetic assumptions, limited realism, and scarce paired haze-free references; NH-HAZE extends realistic benchmarking to non-homogeneous haze.
- Image dehazing methods: Traditional methods use atmospheric cues, polarization-filtered images, known depth, or priors based on hazy-image statistics.
- Image dehazing methods: Other methods restore visibility without explicitly assuming the optical model, including local-contrast maximization and image fusion.
- Image dehazing methods: Deep-learning methods learn dehazing mappings or transmission estimation, commonly using synthetically generated hazy images for training.
- Dehazing assessment: Evaluation was limited by missing haze-free references, motivating non-reference metrics, subjective assessment, and blind visibility measures.
- Dehazing assessment: Synthetic datasets generate haze from simplified optical assumptions, while RGB-NIR is small and indoor.FRIDA contains 66 pairs; RGB-NIR contains four indoor sets.
- Dehazing assessment: O-HAZE, I-HAZE, and DENSE-HAZE provide realistic paired images but use relatively light, indoor, or dense homogeneous haze; NH-HAZE adds non-homogeneous haze.
3. Recording the NH-HAZE dataset
NH-HAZE comprises 55 outdoor scenes captured with and without haze, under controlled conditions designed to reproduce realistic non-homogeneous haze. The acquisition used matched camera settings, professional haze machines, controlled wind, and color-checker calibration.
- NH-HAZE contains 55 outdoor scenes captured both with and without haze.
- Recordings spanned more than two months during autumn, on cloudy mornings or sunsets, with wind kept below 2–3 km/h.Low wind limited rapid haze spreading and contributed to the long recording duration.
- Hazy and haze-free images used identical manually set shutter speed, aperture, ISO, and white balance.Images were recorded as JPG and ARW files at 5456×3632 resolution with 24-bit depth.
- Two professional haze machines generated 1–10-micron vapor particles to simulate atmospheric haze over investigated distances of 20–30 m.A denser haze liquid was selected to imitate water haze over larger distances.
- A fan spread the generated haze for approximately 2–3 minutes, producing a nonuniform distribution 20–30 m in front of the camera.
- A Macbeth color checker was placed in every scene to support post-processing.The checker measured 11 by 8.25 inches and contained a 4×6 grid of painted square samples.
4. Evaluated Dehazing Techniques
The evaluation compares several classical, model-based, enhancement, and CNN-based dehazing techniques on NH-HAZE. Visual comparisons use hazy inputs, ground truth, and method outputs, while quantitative tables report PSNR and SSIM on selected and complete datasets.
- The study validates several competitive single-image dehazing techniques using the NH-HAZE dataset.
- Dark Channel Prior estimates the transmission map from low-intensity pixels in at least one color channel, with guiding-filter refinement used here.
- DehazeNet is an end-to-end CNN that estimates transmission maps by mapping synthesized hazy patches to haze-free patches.Its sequential stages are feature extraction, multi-scale mapping, local extrema, and nonlinear regression.
- Berman et al. estimate transmission and haze from haze-lines formed when haze-free RGB color clusters become lines in hazy images.
- Table 1 computes PSNR and SSIM for 12 randomly selected image sets, while Table 2 reports average PSNR and SSIM across all 55 scenes.
- The evaluated set also includes Retinex-based dehazing, a perceptual pyramid CNN, GridDehazeNet, 3C color-channel compensation, and Ancuti et al.’s local-airlight method.The perceptual pyramid network uses an encoder-decoder structure trained with mean squared error and perceptual losses.
5. Results and Discussion
NH-HAZE enables objective evaluation of dehazing methods on realistic non-homogeneous haze, revealing method-specific strengths and limitations. The results show that haze variation challenges existing techniques, while CNN-based methods generally perform better than the considered non-CNN approaches.
- Method-specific observations: DCP recovers image structure but amplifies color-shifting artifacts while removing varying haze layers.
- Method-specific observations: 3C preprocessing significantly reduces DCP-induced color shifting and produces visually pleasing results for non-homogeneous hazy scenes.
- Method-specific observations: Berman et al. produces increased contrast, sharper edges, and fewer color artifacts, whereas Ancuti et al. yields vivid colors but slight yellowish shifts.
- Method-specific observations: Cai et al. and Liu et al. struggle to restore contrast in hazy regions because their strategies assume homogeneous haze, while Zhang et al. handles haze variation better.
- Overall comparison: CNN-based methods generally perform better than the considered non-CNN techniques, which introduce more color and structural distortions and unnatural appearances.
- Dataset-based evaluation: NH-HAZE facilitates objective quantitative evaluation by providing ground-truth haze-free images for comparison.Table 1 evaluates dehazing outputs against haze-free images using PSNR and SSIM.