Source-linked AI summary
DW-GAN: A Discrete Wavelet Transform GAN for NonHomogeneous Dehazing
Minghan Fu, Huan Liu, Yankun Yu, Jun Chen, Keyan Wang
TL;DR
Non-homogeneous dehazing remains challenging because complex haze can remove texture details and limited paired data can cause over-fitting. DW-GAN uses a two-branch network combining discrete wavelet features with ImageNet-pretrained knowledge adaptation, and reported experiments show strong performance across synthetic and real-world haze settings.
Problem
Non-homogeneous haze complicates texture-detail restoration, while scarce training pairs make robust model training difficult.
Method
DW-GAN combines a DWT branch, an ImageNet-pretrained Res2Net knowledge adaptation branch, and adversarial training in a two-branch dehazing GAN.
Results
DW-GAN is reported to outperform state-of-the-art methods quantitatively and qualitatively, with strong performance on synthetic and real-world dense and non-homogeneous haze.
Takeaways & Limitations
Wavelet frequency information and pretrained classification knowledge are used to improve robustness for limited real-world dehazing data.
Takeaways & Limitations
The DWT branch alone has limited performance on non-homogeneous dehazing because the available data are limited.
Abstract
from arXiv · showhide
Hazy images are often subject to color distortion, blurring, and other visible quality degradation. Some existing CNN-based methods have great performance on removing homogeneous haze, but they are not robust in non-homogeneous case. The reasons are mainly in two folds. Firstly, due to the complicated haze distribution, texture details are easy to be lost during the dehazing process. Secondly, since the training pairs are hard to be collected, training on limited data can easily lead to over-fitting problem. To tackle these two issues, we introduce a novel dehazing network using 2D discrete wavelet transform, namely DW-GAN. Specifically, we propose a two-branch network to deal with the aforementioned problems. By utilizing wavelet transform in DWT branch, our proposed method can retain more high-frequency knowledge in feature maps. In order to prevent over-fitting, ImageNet pre-trained Res2Net is adopted in the knowledge adaptation branch. Owing to the robust feature representations of ImageNet pre-training, the generalization ability of our network is improved dramatically. Finally, a patch-based discriminator is used to reduce artifacts of the restored images. Extensive experimental results demonstrate that the proposed method outperforms the state-of-the-arts quantitatively and qualitatively.
1. Introduction
Non-homogeneous dehazing is difficult because complex haze can erase texture and color details, while limited training pairs make robust learning challenging. DW-GAN addresses these issues with a two-branch GAN using wavelet features, pretrained knowledge, and adversarial learning.
- Complex haze distributions make texture and color details difficult to preserve, while scarce training pairs complicate robust non-homogeneous dehazing.
- The DWT branch retains frequency-domain information and reduces parameters, supporting texture restoration and reducing over-fitting on small datasets.
- The knowledge adaptation branch uses ImageNet-pretrained Res2Net features to introduce prior knowledge for small-scale training datasets.
- Adversarial loss encourages the network to produce natural and photo-realistic restored images.
- The proposed two-branch end-to-end GAN combines a DWT branch with a knowledge adaptation branch for non-homogeneous dehazing.
- Extensive experiments and ablation analyses are presented to evaluate the proposed method.
2. Related Works
Prior work on dehazing includes hand-crafted priors, learned mappings, frequency-domain networks, and adversarial training. These approaches provide complementary strategies for estimating haze, restoring details, and generating realistic images.
- Dehazing methods are broadly categorized into prior-based and learning-based approaches.
- Prior-based methods use statistical knowledge or hand-crafted features, including dark channel and color attenuation priors, to estimate haze-related structure.
- Learning-based methods explore mappings from hazy inputs to clear images using preprocessing, feature fusion, smoothed dilation, or fusion discriminators.
- Frequency-domain methods embed wavelet information into network architectures to learn subbands, recover details, enlarge receptive fields, or improve image restoration.
- GAN-based approaches use generator–discriminator competition and adversarial loss for image translation, super-resolution, dehazing, and deraining.
3. Proposed Method
DW-GAN combines a DWT-based branch for preserving texture information with a knowledge adaptation branch for small-data dehazing. Its training objective combines pixel, perceptual, structural, and adversarial losses.
- 3.1. Network Architecture: The network uses two branches: a DWT branch for dehazing and a knowledge adaptation branch that supplies additional information when training data are limited.The branches are fused by a 7 × 7 convolution to produce clear images.
- 3.1. Network Architecture: The knowledge adaptation branch uses ImageNet-pre-trained Res2Net features, pixel-shuffle up-sampling, attention blocks, and skip connections to process small datasets.The branch is designed to transfer classification knowledge to the dehazing task and identify dynamic hazy patterns.
- 3.2. Discrete Wavelet Transform: Haar DWT decomposes feature maps into low- and high-frequency subbands, helping retain hazy-image details while convolution provides complementary spatial information.The method uses fixed stride-2 filters and emphasizes the xLH, xHL, and xHH subbands for detail preservation.
- 3.2. Discrete Wavelet Transform: Experiments report a great improvement from using DWT, supporting the inclusion of frequency-domain operations in the dehazing pipeline.The reported comparison is referenced as appearing in Section 4.3.
- 3.3. Loss Functions: The training objective combines smooth L1, perceptual, MS-SSIM, and adversarial losses to supervise pixel accuracy, feature similarity, structure, and haze-free realism.The adversarial loss uses the discriminator's probability that the reconstructed image is haze-free.
4. Experiments
DW-GAN is evaluated through ablations, comparisons with state-of-the-art methods, qualitative results, inference time, and challenge performance across synthetic and real-world dehazing datasets. The experiments show that its two-branch architecture, DWT components, combined losses, and pretrained knowledge support strong non-homogeneous dehazing performance.
- 4.3. Ablation Study: The two-branch architecture significantly improves PSNR and SSIM, while adding DWT further improves both metrics and supports texture restoration through frequency-domain information.The branches combine direct hazy-to-clear mapping with adaptation of pretrained knowledge.
- 4.3. Ablation Study: Integrating Smooth L1, perceptual, MS-SSIM, and GAN losses produces the best ablation performance in PSNR and SSIM.The losses provide pixel-wise, deep-feature, structural, and adversarial supervision, respectively.
- 4.4. Comparisons with State-of-the-art Methods: Our method achieves the best PSNR and SSIM on three real-world datasets, exceeding the second-ranked model by 1.07dB on NH-HAZE and 1.54dB on NH-HAZE2.On the synthetic dataset, it is slightly below FFA but remains second best and approaches FFA.
- 4.4. Comparisons with State-of-the-art Methods: Qualitatively, DW-GAN produces visually pleasing outputs closest to ground truth across datasets, while competing methods exhibit haze, color deviation, blur, or artifacts.The comparisons cover SOTS, DENSE-HAZE, NH-HAZE, and NH-HAZE2.
- 4.4. Comparisons with State-of-the-art Methods: DW-GAN runs faster than GCANet, FFA, and TDN while maintaining better qualitative and quantitative performance on 1600 × 1200 images.AOD-Net and DCP are faster but are reported to remove haze less effectively.
- 4.5. NTIRE2021 Dehazing Challenge: On the NTIRE2021 Dehazing Challenge, DW-GAN reports 21.08dB PSNR and 0.8393 SSIM and is among the top-performing methods.The reported visual results remove most haze and appear visually pleasing.