UDPNet: Unleashing Depth-based Priors for Robust Image Dehazing

TL;DR

The paper addresses robustness gaps in single-image dehazing by introducing UDPNet, a depth-prior guided framework that integrates large-scale depth cues into a multi-scale encoder–decoder. It introduces two modules, DGAM for depth-guided channel attention and DPFM for hierarchical depth–RGB fusion via cross-attention, leveraging DepthAnything V2 as a fixed depth provider. Experimental results demonstrate state-of-the-art PSNR/SSIM gains across daytime, nighttime, real-world, and remote sensing datasets, including notable improvements on SOTS, Haze4K, NH-HAZE, Dense-Haze, NHR, and SateHaze1k, as well as strong all-in-one restoration performance. The work highlights the practical significance of depth priors for robust, generalizable image restoration and suggests broad applicability beyond dehazing, with attention to depth quality and inference efficiency.

Abstract

Image dehazing has witnessed significant advancements with the development of deep learning models. However, a few methods predominantly focus on single-modal RGB features, neglecting the inherent correlation between scene depth and haze distribution. Even those that jointly optimize depth estimation and image dehazing often suffer from suboptimal performance due to inadequate utilization of accurate depth information. In this paper, we present UDPNet, a general framework that leverages depth-based priors from large-scale pretrained depth estimation model DepthAnything V2 to boost existing image dehazing models. Specifically, our architecture comprises two typical components: the Depth-Guided Attention Module (DGAM) adaptively modulates features via lightweight depth-guided channel attention, and the Depth Prior Fusion Module (DPFM) enables hierarchical fusion of multi-scale depth map features by dual sliding-window multi-head cross-attention mechanism. These modules ensure both computational efficiency and effective integration of depth priors. Moreover, the intrinsic robustness of depth priors empowers the network to dynamically adapt to varying haze densities, illumination conditions, and domain gaps across synthetic and real-world data. Extensive experimental results demonstrate the effectiveness of our UDPNet, outperforming the state-of-the-art methods on popular dehazing datasets, such as 0.85 dB PSNR improvement on the SOTS dataset, 1.19 dB on the Haze4K dataset and 1.79 dB PSNR on the NHR dataset. Our proposed solution establishes a new benchmark for depth-aware dehazing across various scenarios. Pretrained models and codes will be released at our project https://github.com/Harbinzzy/UDPNet.

Figures (12)

• Figure 1: Illustration of Representative Image Dehazing Frameworks.(a) Approaches based on deep learning dong2020multihong2020distillingwu2021contrastivecui2023imagecui2024revitalizingliu2021swinzamir2022restormerguo2022imageqiu2023mbjin2025mb or physical models he2010singlezhu2015fastberman2018singlehe2012guidedju2019idgcpzhao2021single fail in complex distortions due to idealized assumptions and limited data, resulting in unsatisfactory dehazing outcomes and noticeable artifacts. (b) The collaborative mutual promotion between depth estimation and image dehazing zhang2024depth requires alternate training of two networks, and has lower performance in outdoor and real-world scenarios. (c) The depth-consistent dehazing network wang2024selfpromer according to the differences fails to make good use of depth information. Although the perception quality has been enhanced, the distortion metric is poor. Both (b) and (c) suffer from the inevitable error in depth estimation. (d) Our UDPNet utilizes the robust and reliable depth estimation model trained on large-scale datasets and effectively integrates it into a multi-scale hierarchical network, achieving excellent generalization.
• Figure 2: Previous image dehazing methods often lack robustness to changes in scene conditions. In contrast, our general model effectively restores hazy images under varying illumination, varying scenes, haze distributions, and haze densities. Zoom in for more details.
• Figure 3: Cross-dataset statistical visualization of residual haze (i.e., haze density) across depth ranges highlighting scene-dependent variations and depth–haze correlations in indoor, outdoor, nighttime, remote sensing, and real-world scenarios.
• Figure 4: Comparative analysis of the impact of depth map assistance on various image restoration tasks (i.e., dehazing, deraining, denoising, deblurring, and low-light enhancement). Our experiments show that incorporating depth maps yields substantial gains in dehazing and low-light enhancement, demonstrating their potential for both daytime and nighttime dehazing. Moreover, the overall performance achieves state-of-the-art results. Depth maps are generated using DepthAnything-small depthanything, and our dehazing network, PromptIR potlapalli2023promptir, takes both RGB and depth inputs rather than relying solely on RGB features.
• Figure 5: The architecture of our proposed UDPNet (a) is inspired by the symmetric hierarchical encoder-decoder design cui2023image. UDPNet mainly comprises two organically integrated sub-modules, i.e., DGAM (b) and DPFM (d). DGAM focuses on channel-level fusion between depth-based priors and the input image. DPFM incorporates a multi-scale, multi-head cross-attention mechanism with overlapping windows and dual attention pathways at the spatial level. This design enables UDPNet to act as a general paradigm that leverages depth information to boost existing dehazing models, achieving a trade-off between efficiency and effectiveness.