Open Access

ARTICLE

An Unpaired Dual-Domain Image Dehazing Framework Using Unsupervised Learning

Shunpeng Yang¹, Yunpeng Wu¹, Wenwen Qin¹, Cheng Yang^2,*, Yu Qian³

1 Faculty of Transportation Engineering, Kunming University of Science and Technology, Kunming, China
2 TravelSky Technology Limited, Beijing, China
3 Department of Civil and Environmental Engineering, University of South Carolina, Columbia, SC, USA

* Corresponding Author: Cheng Yang. Email:

(This article belongs to the Special Issue: AI-Enhanced Low-Altitude Technology Applications in Structural Integrity Evaluation and Safety Management of Transportation Infrastructure Systems)

Structural Durability & Health Monitoring 2026, 20(3), 16 https://doi.org/10.32604/sdhm.2026.077878

Received 18 December 2025; Accepted 03 February 2026; Issue published 18 May 2026

Abstract

To enhance traffic infrastructure health monitoring via computer vision (CV) in adverse weather conditions, image dehazing has emerged as a critical processing step. However, current supervised dehazing models, typically trained on synthetic hazy-clean image pairs, often demonstrate limited generalization ability when deployed in real-world haze scenarios. This study proposes a novel unsupervised dehazing framework named the unpaired dual-domain dehazing network (UD³Net). Initially, a novel dual-domain convolutional mixer (DCM) is developed, which can extract local features in the spatial domain and global features in the frequency domain to achieve thorough information fusion, aiming to facilitate accurate estimation of physical parameters in haze imaging. Then, a dual-domain adaptive gating (DAG) fusion module using an attention mechanism is also designed to dynamically integrate both the spatial domain and frequency domain semanteme for image dehazing. Secondly, a newly developed multi-prior contrastive loss (MPC) is proposed to supervise intrinsic properties of unpaired data in space-frequency domains, in order to reduce the loss of semantic information in unpaired unsupervised dehazing. Finally, extensive experiments conducted on both synthetic datasets and real haze datasets validate that the proposed model surpasses the state-of-the-art unsupervised dehazing approaches yet remains while maintaining robust generalization capabilities in real haze scenarios.

---

Keywords

---