Unsupervised Dark-Channel Attention-Guided CycleGAN for Single-Image Dehazing

Jiahao Chen, Chong Wu, Hu Chen, Peng ChengSensors 2020 · DOI

TL;DR

This paper introduces an unsupervised deep learning approach for single-image dehazing that leverages an attention mechanism derived from the dark-channel prior. By adaptively focusing on hazy regions and their intensity, the method effectively removes haze while preserving image details, outperforming existing GAN-based and traditional dehazing techniques.

What Problem Does This Paper Address?

Outdoor images captured in hazy or foggy conditions suffer from reduced visibility, contrast, and color distortion. This degradation significantly impacts the performance of computer vision tasks like object recognition and image segmentation. While traditional dehazing methods rely on physical models and prior knowledge, they often struggle with the complex and uneven distribution of haze in real-world scenarios. Deep learning methods, particularly Generative Adversarial Networks (GANs), have shown promise, but existing GAN-based approaches can lead to undesirable style migration, affecting haze-free regions or failing to adapt to varying haze concentrations. This paper aims to address these limitations by developing an unsupervised method that can accurately identify and process hazy areas of varying intensity while preserving image details and natural appearance.

Methods

The core of this work is an enhanced CycleGAN architecture incorporating a novel attention mechanism. Instead of relying on learned attention maps that might miss subtle haze variations, the authors propose using the dark-channel prior as a basis for their attention maps. The dark-channel prior, a well-established concept in traditional dehazing, naturally highlights hazy regions.

Here's a breakdown of their approach:

  • Dark-Channel Prior for Attention: The dark-channel prior is calculated from the input hazy image. This map is then enhanced and refined through training. The intensity of the dark-channel map is used to represent haze concentration, allowing the model to focus more on denser haze.
  • Attention Mechanism: This enhanced dark-channel map serves as an attention mask. The model uses this mask to selectively apply dehazing operations, ensuring that haze-free areas are not altered and that different levels of haze are treated appropriately.
  • CycleGAN Framework: The method builds upon the unsupervised CycleGAN framework, which allows for image-to-image translation without requiring paired hazy and haze-free images. This involves two generators (GA and GB) and two discriminators (DA and DB) to facilitate cyclic conversion between hazy and haze-free domains.
  • Perceptual Loss: To further preserve image details and textures, a perceptual loss using a pre-trained VGG16 network is incorporated into the loss function.
  • Combined Loss: The overall loss function comprises adversarial loss, cyclic consistency loss, and perceptual loss, optimized to achieve high-quality dehazed images.

Datasets

The authors evaluated their method on two datasets:

  • O-HAZE [10]: This dataset consists of 45 real-world hazy images with corresponding haze-free counterparts. It's noted for its high spatial resolution and diverse haze concentrations, reflecting actual dehazing scenarios.
  • I-HAZE [31]: This dataset includes 35 indoor hazy images with corresponding haze-free images.

The paper also mentions using synthetic hazy images [30] as a common practice in the field but opts for real-world datasets to better reflect natural haze characteristics.

Key Insights and Ideas

  • Dark-Channel Prior as a "Free" Attention Map: The insight that the dark-channel prior, a concept from traditional computer vision, can be effectively repurposed as an attention mechanism for deep learning models is a key contribution. This avoids the computational cost and potential inaccuracies of training complex attention networks from scratch.
  • Adaptive Haze Removal: By using the dark-channel map's intensity to guide the attention, the model can dynamically adjust the dehazing process based on local haze concentrations, leading to more natural results than uniform processing.
  • Preserving Haze-Free Regions: The attention mechanism is crucial for preventing the "style migration" issue common in GANs, ensuring that areas already clear of haze are not adversely affected.
  • Unsupervised Learning for Real-World Haze: The use of CycleGAN's unsupervised nature is vital for handling real-world hazy images where perfectly paired datasets are scarce and difficult to obtain.

Key Results

  • Qualitative Improvements: Visual comparisons show that the proposed method produces significantly clearer and more detailed dehazed images compared to traditional methods (He et al., Berman et al., Fattal) and other GAN-based approaches (pix2pix, CycleDehaze, Mejjati et al., Engin et al.). The results exhibit better color restoration and less artifacting.
  • Quantitative Performance: The method achieved superior performance across key metrics like PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity), and CIEDE2000 (color difference evaluation) on both O-HAZE and I-HAZE datasets. For instance, on the O-HAZE dataset, the proposed method achieved a PSNR of 18.732, SSIM of 0.674, and CIEDE2000 of 10.955, outperforming competitors.
  • Robustness: The model demonstrated robustness in handling misty and even haze-free images, with minimal color distortion, unlike some other methods.
  • Cross-Dataset Generalization: The model showed good generalizability, performing well even when trained on one dataset and tested on another.

Limitations and Future Work

  • Dark-Channel Limitations: The authors acknowledge that the dark-channel prior is not perfect. Specifically, white objects in the dark channel can be misidentified as hazy areas, and the attention map might not be entirely black for truly haze-free regions.
  • Future Directions: The paper suggests future research will focus on developing methods to more accurately identify hazy areas and their concentrations, further improving the dehazing effect and the robustness of the model. This could involve exploring more sophisticated attention mechanisms or integrating additional priors.

Summary updated

Abstract

In this paper, we propose a new unsupervised attention-based cycle generative adversarial network to solve the problem of single-image dehazing. The proposed method adds an attention mechanism that can dehaze different areas on the basis of the previous generative adversarial network (GAN) dehazing method. This mechanism not only avoids the need to change the haze-free area due to the overall style migration of traditional GANs, but also pays attention to the different degrees of haze concentrations that need to be changed, while retaining the details of the original image. To more accurately and quickly label the concentrations and areas of haze, we innovatively use training-enhanced dark channels as attention maps, combining the advantages of prior algorithms and deep learning. The proposed method does not require paired datasets, and it can adequately generate high-resolution images. Experiments demonstrate that our algorithm is superior to previous algorithms in various scenarios. The proposed algorithm can effectively process very hazy images, misty images, and haze-free images, which is of great significance for dehazing in complex scenes.
BibTeX
@inproceedings{chen2020unsupervised,
  author = {Jiahao Chen and Chong Wu and Hu Chen and Peng Cheng},
  title = {{Unsupervised Dark-Channel Attention-Guided CycleGAN for Single-Image Dehazing}},
  booktitle = {Sensors},
  year = {2020},
  doi = {10.3390/s20216000},
}
Figure