Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Depth-agnostic Single Image Dehazing

Honglei Xu, Yan Shu, Shaohui Liu

TL;DR

The paper tackles single-image dehazing by addressing the domain gap caused by depth priors in synthetic data. It introduces DA-HAZE, a depth-agnostic dataset generated via the Atmospheric Scattering Model with randomized depth mappings, and enhances generalization with a Global Shuffle Strategy that pairs haze-free images with multiple depth estimates. A Convolutional Skip Connection is proposed to improve feature fusion in U-Net-based dehazing while keeping computational costs low. Empirical results show that models trained on DA-HAZE with GSS achieve stronger performance on real-world benchmarks and exhibit reduced cross-distribution discrepancies, while CSC consistently boosts dehazing quality across methods and datasets.

Abstract

Single image dehazing is a challenging ill-posed problem. Existing datasets for training deep learning-based methods can be generated by hand-crafted or synthetic schemes. However, the former often suffers from small scales, while the latter forces models to learn scene depth instead of haze distribution, decreasing their dehazing ability. To overcome the problem, we propose a simple yet novel synthetic method to decouple the relationship between haze density and scene depth, by which a depth-agnostic dataset (DA-HAZE) is generated. Meanwhile, a Global Shuffle Strategy (GSS) is proposed for generating differently scaled datasets, thereby enhancing the generalization ability of the model. Extensive experiments indicate that models trained on DA-HAZE achieve significant improvements on real-world benchmarks, with less discrepancy between SOTS and DA-SOTS (the test set of DA-HAZE). Additionally, Depth-agnostic dehazing is a more complicated task because of the lack of depth prior. Therefore, an efficient architecture with stronger feature modeling ability and fewer computational costs is necessary. We revisit the U-Net-based architectures for dehazing, in which dedicatedly designed blocks are incorporated. However, the performances of blocks are constrained by limited feature fusion methods. To this end, we propose a Convolutional Skip Connection (CSC) module, allowing vanilla feature fusion methods to achieve promising results with minimal costs. Extensive experimental results demonstrate that current state-of-the-art methods. equipped with CSC can achieve better performance and reasonable computational expense, whether the haze distribution is relevant to the scene depth.

Depth-agnostic Single Image Dehazing

TL;DR

The paper tackles single-image dehazing by addressing the domain gap caused by depth priors in synthetic data. It introduces DA-HAZE, a depth-agnostic dataset generated via the Atmospheric Scattering Model with randomized depth mappings, and enhances generalization with a Global Shuffle Strategy that pairs haze-free images with multiple depth estimates. A Convolutional Skip Connection is proposed to improve feature fusion in U-Net-based dehazing while keeping computational costs low. Empirical results show that models trained on DA-HAZE with GSS achieve stronger performance on real-world benchmarks and exhibit reduced cross-distribution discrepancies, while CSC consistently boosts dehazing quality across methods and datasets.

Abstract

Single image dehazing is a challenging ill-posed problem. Existing datasets for training deep learning-based methods can be generated by hand-crafted or synthetic schemes. However, the former often suffers from small scales, while the latter forces models to learn scene depth instead of haze distribution, decreasing their dehazing ability. To overcome the problem, we propose a simple yet novel synthetic method to decouple the relationship between haze density and scene depth, by which a depth-agnostic dataset (DA-HAZE) is generated. Meanwhile, a Global Shuffle Strategy (GSS) is proposed for generating differently scaled datasets, thereby enhancing the generalization ability of the model. Extensive experiments indicate that models trained on DA-HAZE achieve significant improvements on real-world benchmarks, with less discrepancy between SOTS and DA-SOTS (the test set of DA-HAZE). Additionally, Depth-agnostic dehazing is a more complicated task because of the lack of depth prior. Therefore, an efficient architecture with stronger feature modeling ability and fewer computational costs is necessary. We revisit the U-Net-based architectures for dehazing, in which dedicatedly designed blocks are incorporated. However, the performances of blocks are constrained by limited feature fusion methods. To this end, we propose a Convolutional Skip Connection (CSC) module, allowing vanilla feature fusion methods to achieve promising results with minimal costs. Extensive experimental results demonstrate that current state-of-the-art methods. equipped with CSC can achieve better performance and reasonable computational expense, whether the haze distribution is relevant to the scene depth.
Paper Structure (14 sections, 8 equations, 6 figures, 3 tables)

This paper contains 14 sections, 8 equations, 6 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Image Dehazing datasets
  4. Single Image Dehazing
  5. Method
  6. Depth-agnostic Dataset
  7. Convolutional Skip Connection
  8. Loss Function
  9. Experiment
  10. Datasets and Metrics
  11. Implementation Details
  12. Comparison with state-of-the-art methods
  13. Abalation studies
  14. Conclusion

Figures (6)

  • Figure 1: (a) is the hazy image. (b) is the restored image generated by the model trained on OTS. (c) is the GroundTruth. (d,e) shows the difference between (b) and (c). (f) is the depth map to stnthesize (a).
  • Figure 2: Comparison of the results generated by NAF-Net trained on the OTS and our DA-HAZE and tested on the corresponding test sets SOTS and DA-SOTS (the top figure), real-world dataset NH-HAZE and O-HAZE (bottom figures). DA-HAZE($\times$n) denotes various scales of the dataset.
  • Figure 3: (a) Non-homogeneous hazy image. (b) Homogeneous hazy image.
  • Figure 4: Comparisons of previous synthetic method (top line) and ours (bottom line). Left column is the haze-free image, middle column is the depth map to synthesis the hazy image (right column).
  • Figure 5: Illustration of Convolutional Skip Connection (CSC). By introducing a single convolution layer (pink line), insufficient feature representations can be mitigated.
  • ...and 1 more figures