Contrastive and distillation-assisted unpaired real-world image dehazing
Zhixuan Sun, Wenbin Xiong, Yingguang Hao, Hongyu Wang · Engineering Applications of Artificial Intelligence · 2026
Supervised image dehazing methods have shown excellent performance on synthetic datasets, achieving detailed restoration results. However, paired training strategies often struggle when applied to real-world hazy images. Existing generative methods can handle certain real-world scenarios but are often plagued by generative hallucinations. To address these issues, we propose a distillation-assisted unpaired contrastive dehazing framework. Unlike previous methods, we utilize a teacher network pre-trained in the synthetic domain to guide patch-level contrastive learning, enabling the preservation of content invariance while further capturing fine-grained mappings between paired data. Meanwhile, to thoroughly capture the characteristics of real-world hazy scenes, we perform image-level contrastive learning in the frequency domain using unpaired clean images and real hazy images, enabling the generation of visually friendly dehazing results. Furthermore, we design a brand-new Spectral Singular value decomposition Dehazing Block (SSDB), which adaptively enhances scene-relevant components while suppressing haze-related features, facilitating clearer restoration. We evaluate our method on two supervised datasets and two challenging unsupervised datasets, employing multiple full-reference and no-reference image quality metrics, along with subjective visual assessments. Experimental results demonstrate that our method outperforms state-of-the-art methods, achieving superior dehazing performance. Our code is available at: https://github.com/GaliXuan/CDA-Dehaze .