:fire: UDPNet: Unleashing Depth-based Priors for Robust Image Dehazing

August 6, 2026 · View on GitHub

This repository contains the official implementation of the following paper:

UDPNet: Unleashing Depth-based Priors for Robust Image Dehazing
Zengyuan Zuo, Junjun Jiang^*, Gang Wu, Xianming Liu

AIIA Lab, Faculty of Computing, Harbin Institute of Technology, Harbin 150001, China.

Overview

overall_structure

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.

:star: If UDPNet is helpful to your projects, please help star this repo. Thank you! :point_left:

Installation

The project is built with PyTorch 3.8, PyTorch 1.8.1. CUDA 10.2, cuDNN 7.6.5 For installing, follow these instructions:

conda install pytorch=1.8.1 torchvision=0.9.1 -c pytorch
pip install tensorboard einops scikit-image pytorch_msssim opencv-python
conda install pillow

Please use the pillow package downloaded by Conda instead of pip.

Install warmup scheduler:

cd pytorch-gradual-warmup-lr/
python setup.py install
cd ..

Pretrained models

Download Links (Dehazing and All-in-One Image Restoration): Baidu Netdisk password: 2026

Results

📊 Synthetic Benchmarks

Table 1. SOTS-Indoor & SOTS-Outdoor

Method	Venue	Indoor PSNR	Indoor SSIM	Outdoor PSNR	Outdoor SSIM
DCP	TPAMI'10	16.61	0.855	19.14	0.861
FFA-Net	AAAI'20	36.39	0.989	33.57	0.984
AECR-Net	CVPR'21	37.17	0.990	-	-
DeHamer	CVPR'22	36.63	0.988	35.18	0.986
DehazeFormer-L	TIP'23	40.05	0.996	-	-
FSNet (Baseline)	TPAMI'23	42.45	0.997	40.40	0.997
MB-TaylorFormer-L	ICCV'23	42.64	0.994	38.09	0.991
FocalNet	ICCV'23	40.82	0.996	37.71	0.995
C²PNet	CVPR'23	42.56	0.995	36.68	0.990
DEA-Net-CR	TIP'24	41.31	0.995	36.59	0.990
DCMPNet	CVPR'24	42.18	0.997	36.56	0.993
GridFormer	IJCV'24	42.34	0.994	-	-
ConvIR-B (Baseline)	TPAMI'24	42.72	0.997	39.42	0.996
SFMN	TIP'25	41.44	0.995	37.72	0.991
PoolNet-B	TIP'25	42.01	0.997	-	-
PGH²Net	AAAI'25	41.70	0.996	37.52	0.989
MB-TaylorFormerV2-L	TPAMI'25	42.84	0.995	39.25	0.992
ConvIR + UDP (Ours)	–	43.12	0.997	40.32	0.996
FSNet + UDP (Ours)	–	43.30	0.997	40.53	0.997

Table 2. Haze4K

Method	PSNR	SSIM
DehazeNet	19.12	0.84
GridDehazeNet	23.29	0.93
FFA-Net	26.96	0.95
DMT-Net	28.53	0.96
PMNet	33.49	0.98
MB-TaylorFormer-L	34.47	0.99
FSNet (Baseline)	34.12	0.99
GridFormer	33.27	0.99
ConvIR-B (Baseline)	34.15	0.99
DEA-Net-CR	34.25	0.99
MB-TaylorFormerV2-B	34.92	0.99
ConvIR + UDP (Ours)	34.82	0.99
FSNet + UDP (Ours)	35.31	0.99

🌙 Nighttime Dehazing

Table 3. GTA5

Method	PSNR	SSIM
GS	21.02	0.639
MRP	20.92	0.646
Ancuti et al.	20.59	0.623
Yan et al.	27.00	0.850
CycleGAN	21.75	0.696
Jin et al.	30.38	0.904
ConvIR-B (Baseline)	31.83	0.921
PoolNet-B (Baseline)	31.53	0.921
PoolNet + UDP	32.78	0.930
ConvIR + UDP	33.12	0.933

Table 4. NHR

Method	PSNR	SSIM
NDIM	14.31	0.526
GS	17.32	0.629
MRPF	16.95	0.667
MRP	19.93	0.777
OSFD	21.32	0.804
HCD	23.43	0.953
FocalNet	25.35	0.969
Jin et al.	26.56	0.890
FSNet (Baseline)	26.30	0.976
ConvIR-B (Baseline)	29.49	0.983
PoolNet-B	28.28	0.980
FSNet + UDP	28.09	0.980
ConvIR + UDP	29.54	0.983

🏞️Real-World Image Dehazing

Table 5. Dense-Haze & NH-HAZE

Method	Dense PSNR	Dense SSIM	Dense LPIPS	NH PSNR	NH SSIM	NH LPIPS
DehazeNet	13.84	0.43	-	16.62	0.52	-
MSBDN	15.37	0.49	-	19.23	0.71	-
DeHamer	16.62	0.56	0.6346	20.66	0.68	0.3837
PMNet	16.79	0.51	-	20.42	0.73	-
MB-TaylorFormer-B	16.66	0.56	0.6125	-	-	-
C²PNet	16.88	0.57	-	20.24	0.69	-
FocalNet	17.07	0.63	0.6087	20.43	0.79	0.3780
SFNet	17.46	0.58	0.5689	20.46	0.80	-
FSNet (Baseline)	17.13	0.65	0.5756	20.55	0.81	0.3624
ConvIR-S (Baseline)	17.45	0.65	0.6000	20.65	0.80	0.3669
ConvIR + UDP	17.55	0.67	0.5813	20.98	0.82	0.3567
FSNet + UDP	17.85	0.65	0.6033	20.94	0.82	0.3732

🌍Remote Sensing Image Dehazing

Table 6. Remote Sensing Dehazing

Method	Thin PSNR	Thin SSIM	Moderate PSNR	Moderate SSIM	Thick PSNR	Thick SSIM
AOD-Net	19.54	0.854	20.10	0.885	15.92	0.731
H2RL-Net	20.91	0.880	22.34	0.906	17.41	0.768
FCFT-Net	23.59	0.913	22.88	0.927	20.03	0.816
C²PNet	19.62	0.880	24.79	0.940	16.83	0.790
Restormer	23.08	0.912	24.73	0.933	18.58	0.762
Trinity-Net	21.55	0.884	23.35	0.895	20.97	0.823
UMWTransformer	24.29	0.919	26.65	0.946	20.07	0.825
FocalNet	24.16	0.916	25.99	0.947	21.69	0.847
ConvIR-S (Baseline)	25.11	0.978	26.79	0.978	22.65	0.950
PoolNet-S (Baseline)	25.02	0.979	27.02	0.979	22.73	0.955
ConvIR + UDP	25.48	0.979	28.07	0.981	22.95	0.953
PoolNet + UDP	26.20	0.980	28.26	0.979	23.13	0.951

🧪 All-in-One Image Restoration Benchmarks

Table 7. Performance on Five Challenging Benchmarks

Method	Dehaze		Derain		Denoise		Deblur		Low-Light		Average
	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
DehazeFormer (TIP’23)	25.31	0.937	33.68	0.954	30.89	0.880	25.93	0.785	21.31	0.819	27.42	0.875
Retinexformer (ICCV’23)	24.81	0.933	32.68	0.940	30.84	0.880	25.09	0.779	22.76	0.863	27.24	0.873
SwinIR (ICCVW’21)	21.50	0.891	30.78	0.923	30.59	0.868	24.52	0.773	17.81	0.723	25.04	0.835
Restormer (CVPR’22)	24.09	0.927	34.81	0.960	31.49	0.884	27.22	0.829	20.41	0.806	27.60	0.881
FSNet (TPAMI’23)	25.53	0.943	36.07	0.968	31.33	0.883	28.32	0.869	22.29	0.829	28.71	0.898
TransWeather (CVPR’22)	21.32	0.885	29.43	0.905	29.00	0.841	25.12	0.757	21.21	0.792	25.22	0.836
AirNet (CVPR’22)	21.04	0.884	32.98	0.951	30.91	0.882	24.35	0.781	18.18	0.735	25.49	0.846
PromptIR (Baseline)	26.54	0.949	36.37	0.970	31.47	0.886	28.71	0.881	22.68	0.832	29.15	0.904
AdaIR (Baseline)	30.53	0.978	38.02	0.981	31.35	0.889	28.12	0.858	23.00	0.845	30.20	0.910
DCPT-PromptIR	30.72	0.977	37.32	0.978	31.32	0.885	28.84	0.877	23.35	0.840	30.31	0.911
DA-RCOT	30.96	0.975	37.87	0.980	31.23	0.888	28.68	0.872	23.25	0.836	30.40	0.911
Perceive-IR	28.19	0.964	37.25	0.977	31.44	0.887	29.46	0.886	22.88	0.833	29.84	0.909
Pool-AIO	30.25	0.977	37.85	0.981	31.24	0.887	27.66	0.844	22.66	0.841	29.93	0.906
DPPD-PromptIR	30.31	0.980	37.32	0.980	31.33	0.885	28.74	0.875	22.73	0.846	30.09	0.913
VLU-Net	30.84	0.980	38.54	0.982	31.43	0.891	27.46	0.840	22.29	0.833	30.11	0.905
PromptIR + UDP (Ours)	31.33	0.980	37.63	0.980	31.25	0.883	28.34	0.868	23.18	0.851	30.35	0.912
AdaIR + UDP (Ours)	31.41	0.980	37.85	0.980	31.28	0.888	28.62	0.870	23.53	0.854	30.55	0.915

Citation

If you find our repo useful for your research, please consider citing our paper:

@article{zuo2026udpnet,
  title={{UDPNet}: Unleashing Depth-based Priors for Robust Image Dehazing},
  author={Zuo, Zengyuan and Jiang, Junjun and Wu, Gang and Liu, Xianming},
  journal={arXiv preprint arXiv:2601.06909},
  year={2026}
}

Acknowledgments

This code is based on FSNet, ConvIR, PoolNet and AdaIR.

Contact

📮📮📮 Should you have any problem, please contact Zengyuan Zuo3565741165@qq.com. We will response to your request as soon as possible!