Source-linked AI summary
FD-GAN: Generative Adversarial Networks with Fusion-discriminator for Single Image Dehazing
Yu Dong, Yihao Liu, He Zhang, Shifeng Chen, Yu Qiao
TL;DR
Existing learning-based dehazing methods commonly estimate intermediate transmission and atmospheric-light parameters, but inaccurate estimates can produce artifacts, color distortion, and insufficient haze removal. FD-GAN directly generates haze-free images with a frequency-aware Fusion-discriminator and a diverse indoor/outdoor training dataset. The method reports state-of-the-art performance on synthetic and real-world images with more visually pleasing results.
Problem
Existing learning-based methods often estimate intermediate transmission and atmospheric-light parameters, whose inaccurate estimation can cause artifacts, color distortion, and insufficient haze removal.
Method
FD-GAN is a fully end-to-end GAN that directly generates haze-free images and uses a Fusion-discriminator incorporating high- and low-frequency information as priors and constraints.
Results
FD-GAN achieves state-of-the-art performance on synthetic and real-world scenes, producing more natural and visually pleasing dehazed images with less color distortion.
Takeaways & Limitations
The Fusion-discriminator and diverse indoor/outdoor training dataset support clearer, more realistic dehazed results across synthetic and real-world images.
Abstract
from arXiv · showhide
Recently, convolutional neural networks (CNNs) have achieved great improvements in single image dehazing and attained much attention in research. Most existing learning-based dehazing methods are not fully end-to-end, which still follow the traditional dehazing procedure: first estimate the medium transmission and the atmospheric light, then recover the haze-free image based on the atmospheric scattering model. However, in practice, due to lack of priors and constraints, it is hard to precisely estimate these intermediate parameters. Inaccurate estimation further degrades the performance of dehazing, resulting in artifacts, color distortion and insufficient haze removal. To address this, we propose a fully end-to-end Generative Adversarial Networks with Fusion-discriminator (FD-GAN) for image dehazing. With the proposed Fusion-discriminator which takes frequency information as additional priors, our model can generator more natural and realistic dehazed images with less color distortion and fewer artifacts. Moreover, we synthesize a large-scale training dataset including various indoor and outdoor hazy images to boost the performance and we reveal that for learning-based dehazing methods, the performance is strictly influenced by the training data. Experiments have shown that our method reaches state-of-the-art performance on both public synthetic datasets and real-world images with more visually pleasing dehazed results.
1 Introduction
Single-image dehazing is ill-posed because haze degrades visibility and existing learning-based methods often estimate intermediate parameters before reconstruction. FD-GAN instead directly generates haze-free images using frequency-aware discrimination and a more diverse training dataset.
- Haze absorbs and scatters light, reducing contrast, color fidelity, and edge information and harming visibility in computer vision applications.
- Recovering the haze-free scene is ill-posed because only the observed hazy image is given.
- Most learning-based methods still estimate transmission and atmospheric light before recovering the haze-free image.
- FD-GAN directly generates haze-free images without estimating intermediate parameters.
- The Fusion-discriminator concatenates generated or ground-truth images with high- and low-frequency components as additional priors and constraints.
- The proposed COCO-based training dataset includes indoor and outdoor images, and models trained on it perform better than those trained on commonly used datasets.
2 Related Work
Prior-based dehazing uses hand-designed assumptions, while learning-based methods use CNNs to learn mappings from hazy to haze-free images. Prior work includes methods that estimate transmission and atmospheric light from learned or handcrafted cues.
- Prior-based methods use hand-designed priors or assumptions, including contrast maximization, albedo-transmission independence, and the dark channel prior.
- Learning-based methods use CNNs to extract dehazing features from large collections of hazy and haze-free image pairs.
- Several learning-based methods estimate the transmission map and atmospheric light, either separately or simultaneously.
- GANs consist of a generator and discriminator that contest within a zero-sum game framework.
3 Method
FD-GAN uses a fully end-to-end generator and a Fusion-discriminator that incorporates low- and high-frequency image information as additional priors and constraints. Its training combines pixel-wise, structural, perceptual, and adversarial losses, while variants compare the contributions of low- and high-frequency inputs.
- Generator: The densely connected encoder-decoder generator directly converts hazy inputs into clear images without estimating intermediate parameters.Its design aims to preserve content, recover details, and remove haze.
- Fusion-discriminator: Low-frequency components emphasize brightness, color, contrast, and structure, whereas high-frequency components represent edges, textures, and fine details.These components provide complementary image information for distinguishing hazy and clear results.
- Fusion-discriminator: The Fusion-discriminator concatenates generated or ground-truth images with their corresponding low-frequency and high-frequency components before classification as fake or real.This differs from traditional discriminators that use only the image itself.
- Results: Fusion-discriminator results show more natural and realistic dehazed images with less color distortion and fewer artifacts.The reported visual comparisons cover synthetic and real-world hazy images.
- Variant studies: The full Fusion-discriminator and single-component variants are evaluated quantitatively and visually, with Fusion-full achieving the highest PSNR and SSIM on SOTS and NTIRE’18.Fusion-LF follows Fusion-full quantitatively, while Fusion-HF removes more haze but can introduce slight color distortion.
4 Experiment
Experiments evaluate FD-GAN against established methods on synthetic and real-world images, while also introducing a diverse indoor–outdoor training dataset.
- Dataset for training: The new dataset is synthesized from Microsoft COCO and includes various indoor and outdoor hazy images.It contains 24,200 paired synthetic images, with depths generated using MegaDepth and haze synthesized using related methods.
- Experimental setting: FD-GAN is compared with DCP, AOD, DCPDN, GFN, and EPDN on synthetic datasets and real-world images.The evaluation uses the NTIRE’18 toolkit and reports PSNR and SSIM on synthetic datasets.
- Synthetic datasets: FD-GAN achieves the highest PSNR and SSIM values by a large margin on the synthetic evaluation.Qualitatively, it produces sharper textures and better color fidelity than the compared methods.
- Real-world datasets: On real-world hazy images, the model trained on synthetic data generalizes well and produces more natural dehazed results.Compared methods exhibit haze residuals, darkening, or color distortion in the reported examples.
5 Ablation Study
The ablation study examines training datasets and discriminator designs, finding benefits from both the proposed dataset and frequency-informed fusion.
- Effect of training datasets: Models trained on the proposed synthetic dataset outperform models trained on the indoor and outdoor RESIDE training sets.The compared models use the same network architecture and training settings except for the modifications specified in Table 3.
- Ablation study: Models with GAN losses surpass the L1-and-SSIM baseline in PSNR and SSIM on SOTS and NTIRE’18 Test sets.The study evaluates the models on both datasets under matched architectural and training conditions.
- Fusion-discriminator: Introducing the Fusion-discriminator removes haze more thoroughly and produces cleaner images with better color fidelity than the standard GAN model.The standard GAN reduces more haze than the baseline but tends to introduce artifacts and color distortion.
- Fusion-discriminator: The ablation study reports significant improvements from developing the Fusion-discriminator.Table 3 summarizes the quantitative ablation results.
6 Conclusion
FD-GAN is a fully end-to-end dehazing algorithm with a frequency-informed Fusion-discriminator. Experiments show favorable performance and visually pleasing results on synthetic and real-world images.
- Conclusion: FD-GAN directly generates dehazed images while integrating frequency information as additional priors and constraints.The method is presented as a fully end-to-end image dehazing algorithm.
- Conclusion: FD-GAN performs favorably against several state-of-the-art methods on synthetic datasets and real-world hazy images.The reported results include less color distortion and visually pleasing dehazed outputs.