Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Addressing Domain Discrepancy: A Dual-branch Collaborative Model to Unsupervised Dehazing

Shuaibin Fan, Minglong Xue, Aoxiang Ning, Senming Zhong

TL;DR

This work tackles domain discrepancy in unsupervised image dehazing by introducing DCM-dehaze, a dual-branch framework that couples a dehazing pathway with a contour-constrained pathway. Key innovations include Dense Flow Residual Enhancer (DFRE) to remove redundant high-frequency information, Dual Depthwise Separable Convolutional Module (DDSCM) to deepen feature representations, and Bidirectional Contour Analysis (BCA) to sharpen edges and textures. The model optimizes a combined loss with cycle-consistency, adversarial, and contour terms, promoting realistic dehazing while preserving boundaries. Empirical results on RESIDE datasets and real-world hazy images demonstrate state-of-the-art performance for unsupervised dehazing and robustness to domain shifts, with clear improvements in edge fidelity and detail preservation.

Abstract

Although synthetic data can alleviate acquisition challenges in image dehazing tasks, it also introduces the problem of domain bias when dealing with small-scale data. This paper proposes a novel dual-branch collaborative unpaired dehazing model (DCM-dehaze) to address this issue. The proposed method consists of two collaborative branches: dehazing and contour constraints. Specifically, we design a dual depthwise separable convolutional module (DDSCM) to enhance the information expressiveness of deeper features and the correlation to shallow features. In addition, we construct a bidirectional contour function to optimize the edge features of the image to enhance the clarity and fidelity of the image details. Furthermore, we present feature enhancers via a residual dense architecture to eliminate redundant features of the dehazing process and further alleviate the domain deviation problem. Extensive experiments on benchmark datasets show that our method reaches the state-of-the-art. This project code will be available at \url{https://github.com/Fan-pixel/DCM-dehaze.

Addressing Domain Discrepancy: A Dual-branch Collaborative Model to Unsupervised Dehazing

TL;DR

This work tackles domain discrepancy in unsupervised image dehazing by introducing DCM-dehaze, a dual-branch framework that couples a dehazing pathway with a contour-constrained pathway. Key innovations include Dense Flow Residual Enhancer (DFRE) to remove redundant high-frequency information, Dual Depthwise Separable Convolutional Module (DDSCM) to deepen feature representations, and Bidirectional Contour Analysis (BCA) to sharpen edges and textures. The model optimizes a combined loss with cycle-consistency, adversarial, and contour terms, promoting realistic dehazing while preserving boundaries. Empirical results on RESIDE datasets and real-world hazy images demonstrate state-of-the-art performance for unsupervised dehazing and robustness to domain shifts, with clear improvements in edge fidelity and detail preservation.

Abstract

Although synthetic data can alleviate acquisition challenges in image dehazing tasks, it also introduces the problem of domain bias when dealing with small-scale data. This paper proposes a novel dual-branch collaborative unpaired dehazing model (DCM-dehaze) to address this issue. The proposed method consists of two collaborative branches: dehazing and contour constraints. Specifically, we design a dual depthwise separable convolutional module (DDSCM) to enhance the information expressiveness of deeper features and the correlation to shallow features. In addition, we construct a bidirectional contour function to optimize the edge features of the image to enhance the clarity and fidelity of the image details. Furthermore, we present feature enhancers via a residual dense architecture to eliminate redundant features of the dehazing process and further alleviate the domain deviation problem. Extensive experiments on benchmark datasets show that our method reaches the state-of-the-art. This project code will be available at \url{https://github.com/Fan-pixel/DCM-dehaze.
Paper Structure (13 sections, 11 equations, 7 figures, 3 tables)

This paper contains 13 sections, 11 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Prior-based Image Dehazing Methods
  4. Learning-based Image Dehazing Methods
  5. Methodology
  6. Dense Flow Residual Enhancer
  7. Bidirectional Contour Analysis
  8. Dual Depthwise Separable Convolutional Module
  9. Loss Function
  10. EXPERIMENTS
  11. Comparisons on real-world hazy images
  12. Ablation Study
  13. Conclusion

Figures (7)

  • Figure 1: A single image haze removal example. (a) The hazy image. (b) Result of Yang et al. yang2022self proposed network. (c) The dehazed image of our method. (d) The Ground-truth image.
  • Figure 2: The concept of dual-task coordination and mutual promotion. DP-GT and DP-HA denote depth maps of ground-truth and haze removal images. BCA denotes the Bidirectional Contour Analysis module.
  • Figure 3: The architecture of the method consists of the dehazing network, DFRE and DDSCM. The FFM is the central fusion part of the dual-task interaction mechanism based on the attention mechanism. This mechanism allows seamless dehazing and contour constraints integration into a unified framework and improves the model's performance by mutual reinforcement.
  • Figure 4: The structure of single Residual Dense Block.
  • Figure 5: The structure of Dual Depthwise Separable Convolutional Module.
  • ...and 2 more figures