DSH-Net: dual-stage nighttime image dehazing via dynamic histogram attention and soft–hard pseudo-label semi-supervised learning

Nighttime hazy images often exhibit spatially non-uniform haze, coexisting overexposed and underexposed regions, and color distortion due to the coupled effects of haze scattering, halo effects, and low-light noise, posing substantial challenges to image restoration. To address these issues, we propose DSH-Net, a dual-stage framework for nighttime image dehazing. DSH-Net comprises a dynamic histogram attention restoration network (DHARNet) and a detail enhancement network (DENet), which perform global structural restoration and local detail enhancement, respectively. To address the spatially non-uniform degradation caused by the coupling of haze and complex illumination, DHARNet incorporates a sorting-based dynamic-range histogram self-attention mechanism to model the non-uniform distribution of nighttime haze. This mechanism is combined with a frequency-modulated feed-forward network that adaptively modulates multi-frequency features, thereby enabling joint modeling in the spatial and frequency domains. DENet progressively models multi-scale spatial features to further enhance local textures and structural details. To mitigate the domain gap between synthetic data and real-world nighttime scenes, we use DHARNet and DENet to generate complementary pseudo-labels and introduce a confidence-aware adaptive weighting mechanism that dynamically balances structural consistency and detail preservation. The experimental results indicate that the proposed method achieves competitive performance across the evaluated synthetic datasets and exhibits a favorable trend in no-reference image-quality metrics on held-out images from the real-world nighttime haze (RWNH) dataset after target-domain adaptation.