Source-linked AI summary
From Synthetic to Real: Image Dehazing Collaborating with Unlabeled Real Data
Ye Liu, Lei Zhu, Shunda Pei, Huazhu Fu, Jing Qin, Qing Zhang, Liang Wan, Wei Feng
TL;DR
Single-image dehazing suffers when models trained on synthetic haze encounter real images, while conventional CNN features may conflate multiple physical haze factors. The paper introduces DID-Net and a mean-teacher DMT-Net that disentangle these factors and use unlabeled real data, reporting improvements over existing methods across benchmark and real-world evaluations.
Problem
Synthetic-to-real domain shift degrades dehazing, and predicting only one haze factor overlooks the combined physical components in hazy images.
Method
DID-Net disentangles features into haze-free image, transmission, and atmospheric-light components, while DMT-Net trains with supervised synthetic data and consistency on unlabeled real data.
Results
The method clearly outperforms state-of-the-art dehazing methods on synthesized datasets and real-world photos.
Takeaways & Limitations
Disentangled physical-component prediction combined with unlabeled real images is effective for single-image dehazing across synthetic and real-world evaluations.
Takeaways & Limitations
The model has 51.79M parameters and 0.127s inference time, motivating future work to reduce complexity and inference time.
Abstract
from arXiv · showhide
Single image dehazing is a challenging task, for which the domain shift between synthetic training data and real-world testing images usually leads to degradation of existing methods. To address this issue, we propose a novel image dehazing framework collaborating with unlabeled real data. First, we develop a disentangled image dehazing network (DID-Net), which disentangles the feature representations into three component maps, i.e. the latent haze-free image, the transmission map, and the global atmospheric light estimate, respecting the physical model of a haze process. Our DID-Net predicts the three component maps by progressively integrating features across scales, and refines each map by passing an independent refinement network. Then a disentangled-consistency mean-teacher network (DMT-Net) is employed to collaborate unlabeled real data for boosting single image dehazing. Specifically, we encourage the coarse predictions and refinements of each disentangled component to be consistent between the student and teacher networks by using a consistency loss on unlabeled real data. We make comparison with 13 state-of-the-art dehazing methods on a new collected dataset (Haze4K) and two widely-used dehazing datasets (i.e., SOTS and HazeRD), as well as on real-world hazy images. Experimental results demonstrate that our method has obvious quantitative and qualitative improvements over the existing methods.
1 INTRODUCTION
Single-image dehazing methods face synthetic-to-real domain shift and often predict only one physical haze factor. This work addresses both issues by disentangling predictions and using unlabeled real images through consistency training.
- Haze distorts image content and degrades accuracy in subsequent visual analysis, motivating recovery of the latent haze-free image.
- Existing CNN methods commonly train on synthesized hazy images, causing degraded performance on real-world hazy photos because of domain shift.
- Because haze combines transmission, atmospheric light, and haze-free content, predicting only one factor can hinder dehazing performance.
- DID-Net disentangles features into transmission, latent-image, and light components, progressively integrating them to predict three corresponding maps.
- DMT-Net combines supervised loss on labeled synthetic data with disentangled consistency loss on unlabeled real-world data.
- The method is evaluated against 13 state-of-the-art methods across Haze4K, SOTS, HazeRD, and real-world hazy images, with reported performance improvements.
2 RELATED WORK
Traditional methods rely on hand-crafted priors, while later CNN and domain-adaptation methods estimate haze factors or clean images with distinct limitations. The proposed approach separates physical components and uses unlabeled data to improve dehazing.
- Traditional dehazing methods use hand-crafted priors to estimate transmission, but these priors do not always hold in diverse real-world photos.
- Some early CNN methods estimate transmission maps, yet inaccurate transmission estimation can limit haze-removal quality.
- Later CNN methods directly learn latent clean images from single hazy inputs using end-to-end architectures.
- DCPDN separately estimates transmission and atmospheric light before computing the haze-free image through the physical haze model.
- A domain-adaptation method translates between synthetic and real haze domains, but its dehazing modules predict only the latent haze-free image.
- This work instead learns disentangled feature representations and leverages unlabeled data for improving dehazing performance.
3 OUR APPROACH
The framework combines DID-Net’s physically motivated disentanglement and coarse-to-fine component prediction with DMT-Net’s consistency training on labeled synthetic and unlabeled real-world images.
- Disentangled Image Dehazing Network: DID-Net disentangles multiscale features into haze-free image, transmission, and atmospheric-light components for joint dehazing.The three components correspond to the physical haze model and are estimated from separate distilled feature representations.
- Disentangled Image Dehazing Network: Independent U-Net-based residual blocks refine the three coarse predictions, producing refined haze-free, transmission, and atmospheric-light estimates.The refinement networks share encoder-decoder structure but do not share parameters.
- Disentangled Image Dehazing Network: The network reconstructs coarse and refined hazy images from the three predicted components using the physical haze model, then applies an L1 reconstruction loss against the input hazy image.This loss compares the observed hazy image with both reconstructed versions.
- Disentangled Image Dehazing Network: Three independent branches progressively merge disentangled features across scales to generate coarse predictions for the haze-free image, transmission map, and atmospheric-light map.The merging process proceeds from deeper, lower-resolution features toward shallower, higher-resolution features before a 1×1 convolution produces each prediction.
- Our Network: DMT-Net trains with supervised losses on labeled synthetic images and consistency losses on unlabeled real-world images, using student predictions and an EMA-updated teacher.For unlabeled data, consistency is computed across clean-image, transmission, atmospheric-image, and reconstructed-hazy-image outputs.
4 EXPERIMENTAL RESULTS
Experiments compare the proposed network with 13 dehazing methods on synthetic benchmarks and visually on synthetic and real-world hazy images. Ablations and hyper-parameter studies examine disentangled features, refinement, unlabeled data, and loss weighting.
- Experimental setup: The evaluation compares the proposed network with 13 state-of-the-art methods using PSNR and SSIM on Haze4K, SOTS, and HazeRD.Haze4K contains 4,000 hazy images, with 3,000 for training and 1,000 for testing.
- Quantitative comparison: Our method consistently achieves the largest PSNR and SSIM scores on Haze4K, SOTS, and HazeRD.DID-Net using only labeled data already outperforms most existing CNN-based methods.
- Visual comparison: On synthetic hazy photos, competing methods show color distortions or darkened regions, whereas the proposed results are closest to the latent ground truth.The proposed DMT-Net results are reported to have higher visual quality and less color distortion.
- Visual comparison: On real-world hazy photos, several baselines retain substantial fog, darken areas, or produce color distortions, while our method removes haze more effectively with realistic colors.The comparison covers real-world photos from the RESIDE dataset.
- Ablation study: Disentangling features improves over the basic baseline, refinement further improves performance, and DMT-Net outperforms the labeled-data DID-Net baseline.The ablation attributes successive gains to feature disentangling, three refinement blocks, and unlabeled data.
- Further analysis: Different loss-weight hyper-parameter settings affect dehazed results, but the tested settings all achieve good performance; the teacher network reaches 28.34 PSNR and 0.96 SSIM.Inference uses the student network, whose performance is only slightly better than the teacher network.
5 CONCLUSION
The work presents DMT-Net, combining feature disentanglement, DID-Net’s coarse-to-fine predictions, and unlabeled real-world images for single-image dehazing. Experiments on synthesized datasets and real-world photos show clear improvements over state-of-the-art methods.
- DMT-Net leverages feature disentanglement and unlabeled real-world images to improve single-image dehazing.
- DID-Net disentangles hazy-image features to simultaneously predict clean images, transmission maps, and atmospheric images.
- DID-Net follows a coarse-to-fine strategy for predicting the three dehazing components.
- Experimental results on synthesized datasets and real-world photos demonstrate effectiveness and clear outperformance over state-of-the-art image dehazing methods.