Towards Domain Invariant Single Image Dehazing

Pranjay Shyam, Kuk-Jin Yoon, Kyung-Soo Kim

Abstract

Presence of haze in images obscures underlying information, which is undesirable in applications requiring accurate environment information. To recover such an image, a dehazing algorithm should localize and recover affected regions while ensuring consistency between recovered and its neighboring regions. However owing to fixed receptive field of convolutional kernels and non uniform haze distribution, assuring consistency between regions is difficult. In this paper, we utilize an encoder-decoder based network architecture to perform the task of dehazing and integrate an spatially aware channel attention mechanism to enhance features of interest beyond the receptive field of traditional conventional kernels. To ensure performance consistency across diverse range of haze densities, we utilize greedy localized data augmentation mechanism. Synthetic datasets are typically used to ensure a large amount of paired training samples, however the methodology to generate such samples introduces a gap between them and real images while accounting for only uniform haze distribution and overlooking more realistic scenario of non-uniform haze distribution resulting in inferior dehazing performance when evaluated on real datasets. Despite this, the abundance of paired samples within synthetic datasets cannot be ignored. Thus to ensure performance consistency across diverse datasets, we train the proposed network within an adversarial prior-guided framework that relies on a generated image along with its low and high frequency components to determine if properties of dehazed images matches those of ground truth. We preform extensive experiments to validate the dehazing and domain invariance performance of proposed framework across diverse domains and report state-of-the-art (SoTA) results.

Introduction

Single image dehazing is a critical preprocessing step for computer vision tasks in outdoor environments. Most modern dehazing algorithms rely on deep learning models trained on synthetic datasets created using the Atmospheric Scattering Model. This model typically assumes haze is uniform across the entire image. However, real-world haze is inherently non-uniform, influenced by depth, wind, and varying particle distributions. This discrepancy leads to a significant "domain gap," where models trained on synthetic data fail to generalize to real-world scenarios.

Comparison of synthetic uniform haze (b) and proposed localized data augmentation (c) against real-world hazy scenes (d-f).

Furthermore, standard convolutional neural networks (CNNs) are limited by fixed receptive fields, making it difficult to maintain structural and color consistency across large, non-uniformly hazy regions. This paper addresses these challenges through a novel framework designed for domain invariant dehazing.

Proposed Methodology

The proposed framework, termed DIDH (Domain Invariant Dehazing), introduces three key innovations to bridge the domain gap and improve feature recovery.

Spatially Aware Channel Attention (SACA)

To overcome the limitations of local receptive fields, the authors integrate SACA modules into the skip connections of a UNet-style encoder-decoder. SACA allows the network to capture long-range dependencies, ensuring that the dehazing process in one region is informed by the global context of the image. This ensures color and structural consistency across the entire recovered scene.

Greedy Localized Data Augmentation (GLDA)

To simulate the non-uniformity of real haze, GLDA is employed during training. Instead of applying global haze, GLDA "pastes" hazy patches of random shapes and sizes onto clean images. This forces the model to learn how to localize and specifically target hazy regions rather than applying a uniform correction, which significantly improves performance on real-world non-uniform haze.

Frequency Prior-based Dual Discriminators

The model is trained within an adversarial framework utilizing two distinct discriminators to guide the generator:

  1. Low Frequency (LF) Discriminator: Focuses on preserving global structure and color consistency.
  2. High Frequency (HF) Discriminator: Focuses on recovering sharp edges and fine details.
The proposed DIDH framework featuring an encoder-decoder architecture with SACA modules and frequency-based dual discriminators.

Experiments and Results

The DIDH framework was evaluated on both synthetic (SOTS-Indoor/Outdoor) and real-world datasets (NTIRE-18, 19, and 20). The model demonstrated exceptional domain invariance, particularly on real-world benchmarks where previous SoTA models often struggled.

On the NTIRE-19 (Real) dataset, DIDH achieved a PSNR of 22.80 dB, a substantial improvement over the previous best of 15.96 dB. Similarly, on NTIRE-20 (Real), the model reached 23.66 dB, outperforming competitors by nearly 4 dB. Qualitative analysis confirms that DIDH produces cleaner, more natural-looking images with fewer artifacts compared to models like FFA-Net or MSNet.

Conclusion

By combining spatially aware attention, localized data augmentation, and frequency-disentangled adversarial training, the DIDH framework effectively addresses the domain gap in single image dehazing. The results demonstrate that focusing on non-uniform haze distribution and frequency priors is essential for creating models that generalize from synthetic training environments to the complexities of the real world.

Summary updated

Abstract

Presence of haze in images obscures underlying information, which is undesirable in applications requiring accurate environment information. To recover such an image, a dehazing algorithm should localize and recover affected regions while ensuring consistency between recovered and its neighboring regions. However owing to fixed receptive field of convolutional kernels and non uniform haze distribution, assuring consistency between regions is difficult. In this paper, we utilize an encoder-decoder based network architecture to perform the task of dehazing and integrate an spatially aware channel attention mechanism to enhance features of interest beyond the receptive field of traditional conventional kernels. To ensure performance consistency across diverse range of haze densities, we utilize greedy localized data augmentation mechanism. Synthetic datasets are typically used to ensure a large amount of paired training samples, however the methodology to generate such samples introduces a gap between them and real images while accounting for only uniform haze distribution and overlooking more realistic scenario of non-uniform haze distribution resulting in inferior dehazing performance when evaluated on real datasets. Despite this, the abundance of paired samples within synthetic datasets cannot be ignored. Thus to ensure performance consistency across diverse datasets, we train the proposed network within an adversarial prior-guided framework that relies on a generated image along with its low and high frequency components to determine if properties of dehazed images matches those of ground truth. We preform extensive experiments to validate the dehazing and domain invariance performance of proposed framework across diverse domains and report state-of-the-art (SoTA) results.
BibTeX
@inproceedings{shyam2021towards,
  author = {Pranjay Shyam and Kuk-Jin Yoon and Kyung-Soo Kim},
  title = {{Towards Domain Invariant Single Image Dehazing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=NTdeB0UdXjW},
}
Figure