Depth-guided hierarchical network for single-image dehazing

Image dehazing has achieved remarkable progress with deep learning. However, most existing approaches rely on RGB representations, which limits their capability to characterize spatially varying haze degradation in complex environments. Although depth provides valuable geometric priors for haze-aware restoration, depth information inferred from hazy images can be unreliable in low-contrast and severely degraded regions, leading to inaccurate guidance. To address these challenges, we propose a hierarchical dehazing network with reliability-aware depth priors. Specifically, a pretrained Depth Anything V2 model is employed to extract depth priors from hazy images. Based on these priors, three complementary components are developed: the depth-guided modulation module transforms the depth map into stage-specific scaling and offset parameters to modulate multistage image features; the depth reliability attention module evaluates the structural consistency between depth priors and image features and adaptively regulates feature responses; and the depth balanced loss provides equally weighted reconstruction supervision across three regions defined by the depth distribution of each image. The proposed framework exploits depth priors throughout restoration and recalibrates auxiliary attention using the structural consistency between the depth prior and image features. Experiments on the synthetic ITS and OTS benchmarks and the real-world NH-HAZE and Dense-HAZE datasets demonstrate the effectiveness of the proposed method in preserving fine structural details and enhancing overall restoration quality.