Abstract Haze significantly degrades image quality and adversely affects the performance of downstream vision applications. Thus, dehazing is essential for restoring visual clarity. Most existing data-driven dehazing methods rely heavily on synthetic hazy-clean image pairs due to the scarcity of real-world paired data. However, these methods often suffer from limited generalization in real-world settings owing to the inherent domain shift between synthetic and real haze distributions. To address this challenge, we propose a Domain-Invariant Dehazing Network (DID-Net) which comprises two core components: a Depth-Guided Transmission Map Estimation Network (DTME-Net) and a Physics-Aware Dehazing Network (PDD-Net). DTME-Net learns from real-world depth-transmission mappings to generate synthetic hazy images with realistic distributions, providing reliable data for enhanced cross-domain generalization. PDD-Net leverages depth-aware attention to modulate features by spatial depth, improving dehazing in complex scenes. We further employ post-optimization to refine its parameters for superior results. Extensive experimental results on real-world benchmarks demonstrate that the proposed method significantly mitigates the synthetic-to-real domain gap and outperforms state-of-the-art dehazing approaches both quantitatively and qualitatively.
Ling et al. (Wed,) studied this question.