Source-linked AI summary
Domain Adaptation for Image Dehazing
Yuanjie Shao, Lerenhan Li, Wenqi Ren, Changxin Gao, Nong Sang
TL;DR
Synthetic-trained dehazing models often generalize poorly to real hazy images because of domain shift. The paper introduces an end-to-end framework combining bidirectional image translation with two dehazing networks and reports favorable results on synthetic and real-world images.
Problem
Dehazing models trained mainly on synthetic hazy images often generalize poorly to real hazy images because large paired real datasets are impractical to acquire and domain shift remains.
Method
The framework translates images bidirectionally between synthetic and real domains, trains two domain-related dehazing networks with consistency, and incorporates real hazy images using clear-image properties.
Results
The method obtains the highest PSNR and SSIM on both SOTS and HazeRD, exceeding EPDN by 3.94 dB PSNR and 0.04 SSIM on SOTS.
Takeaways & Limitations
Experiments on synthetic datasets and real-world images show favorable performance against state-of-the-art dehazing approaches.
Abstract
from arXiv · showhide
Image dehazing using learning-based methods has achieved state-of-the-art performance in recent years. However, most existing methods train a dehazing model on synthetic hazy images, which are less able to generalize well to real hazy images due to domain shift. To address this issue, we propose a domain adaptation paradigm, which consists of an image translation module and two image dehazing modules. Specifically, we first apply a bidirectional translation network to bridge the gap between the synthetic and real domains by translating images from one domain to another. And then, we use images before and after translation to train the proposed two image dehazing networks with a consistency constraint. In this phase, we incorporate the real hazy image into the dehazing training via exploiting the properties of the clear image (e.g., dark channel prior and image gradient smoothing) to further improve the domain adaptivity. By training image translation and dehazing network in an end-to-end manner, we can obtain better effects of both image translation and dehazing. Experimental results on both synthetic and real-world images demonstrate that our model performs favorably against the state-of-the-art dehazing algorithms.
1. Introduction
Single-image dehazing is important but difficult because transmission estimation is ill-posed and synthetic-trained models often generalize poorly to real hazy images. The paper addresses this domain shift with an end-to-end translation-and-dehazing framework.
- Single-image dehazing aims to recover clean images from hazy inputs for downstream tasks including object recognition and scene understanding.
- Transmission estimation is generally ill-posed, and inconsistent image priors can produce inaccurate transmission approximations and undesirable restored images.
- Deep learning dehazing methods usually train on synthetic hazy datasets because large paired real hazy and haze-free datasets are impractical to acquire.
- The proposed framework uses bidirectional image translation, two domain-related dehazing modules, and consistency training on original and translated images.Real hazy images are also incorporated through clear-image properties such as the dark channel prior and image-gradient smoothing.
- The work reports favorable performance against state-of-the-art dehazing approaches on synthetic datasets and real-world hazy images.Its stated contributions include bridging synthetic and real domains and improving performance by incorporating real hazy images during training.
2. Related Work
Prior dehazing methods include image-statistics-based approaches and CNN-based models, while domain adaptation methods align feature or pixel distributions. The paper positions its approach as combining image translation with dehazing to address synthetic-to-real domain shift.
- Prior-based methods: Prior-based dehazing estimates transmission maps and atmospheric light from clear-image statistics, including contrast, dark-channel, attenuation, and color-line assumptions.
- Learning-based Methods: Learning-based methods use deep CNNs to estimate transmission and atmospheric light or directly restore clean images, often leveraging large synthetic datasets.
- Learning-based Methods: CNN models trained on synthetic images can suffer significant performance drops in the real domain because of the domain gap.Using real haze images alone does not fully solve the domain-shift problem according to the related-work discussion.
- Domain adaptation: The paper distinguishes its approach by translating between domains before using domain-related dehazing networks to address the domain-shift problem.
- Domain adaptation: Domain adaptation methods reduce discrepancies through feature-level alignment, pixel-level image translation, or joint use of both strategies.
3. Proposed Method
The proposed framework adapts image dehazing across synthetic and real domains using bidirectional image translation, two domain-specific dehazing networks, and jointly trained losses. Depth-guided translation, supervised and unsupervised dehazing losses, and cross-network consistency support the framework.
- Method Overview: The method combines bidirectional image translation with separate dehazing networks for synthetic and real domains.The translators bridge domain differences, while each dehazing network processes domain-specific and translated images.
- Method Overview: The framework inputs synthetic images with depth maps and real hazy images, then produces translated images for the two dehazing networks.Synthetic-to-real and real-to-synthetic translation precede domain-specific dehazing.
- Image Translation Module: Depth-guided synthetic-to-real translation uses an SFT layer to modulate image features according to depth-derived conditioning maps.The SFT layer predicts modulation parameters γ and β from depth features and applies them to convolutional features.
- Image Translation Module: The translation module uses image-level and feature-level adversarial losses, cycle consistency, and identity mapping to reduce domain discrepancy while preserving content.The translators align image and feature distributions, while cycle consistency constrains inverse translations to reconstruct the inputs.
- Training Losses: The dehazing modules use supervised reconstruction losses on synthetic data and unsupervised total-variation and dark-channel losses on real-related data.The unsupervised losses encourage outputs with statistical characteristics associated with clear images, including smooth gradients and dark channels.
- Training Losses: A consistency loss encourages the two dehazing networks to produce similar outputs for real hazy images, and all losses are combined in an overall objective.The complementary networks are trained on source and translated images before their real-domain predictions are aligned.
4. Experimental Results
Experiments evaluate the framework on synthetic and real hazy images, with quantitative gains on synthetic datasets and visually improved restoration across domains.
- Evaluation Setup: The experiments evaluate the method on SOTS and HazeRD synthetic datasets and real hazy images, using visual, quantitative, and ablation comparisons.The study compares against DCP, MSCNN, DehazeNet, NLD, AOD-Net, GFN, DCPDN, and EPDN.
- Synthetic Datasets: The proposed method restores sharper structures and details closer to ground truth than competing methods on SOTS and produces better visual effects on HazeRD.NLD and GFN show color distortion, EPDN can appear darker, and several methods retain haze.
- Synthetic Datasets: The method obtains the highest PSNR and SSIM values on both synthetic datasets.Compared with EPDN, gains are 3.94 dB PSNR and 0.04 SSIM on STOS, and up to 0.7 dB PSNR and 0.07 SSIM on HazeRD.
- Real Images: On real hazy images from URHI, the proposed method restores more details and obtains visually pleasing images overall.Compared methods exhibit color distortion, darker results, remaining haze artifacts, or generally lower brightness.
- Ablation Study: Ablations show that image translation and unsupervised loss improve dehazing across synthetic and real domains, with both producing cleaner and more visually pleasing results.The combined design can make scenes such as skies brighter, and the quantitative results agree with the qualitative comparisons.
5. Conclusions
The paper proposes an end-to-end domain adaptation framework that translates between synthetic and real domains before using domain-specific dehazing modules. Experiments on synthetic and real-world images show favorable performance against state-of-the-art methods.
- 5. Conclusions: The framework contains an image translation module and two image dehazing modules for reducing discrepancy between synthetic and real domains.Translated and original images are used as inputs to the dehazing networks, while real hazy images contribute through clean-image properties.