High Dynamic Range Imaging via Spatial-Frequency Interaction

---

Weiyu Zhou¹, Yongqing Yang¹, Tao Hu¹, Pu Hui¹, Jian Jin², Yu Cao¹, Qingsen Yan^12*, Yanning Zhang^1
1Northwestern Polytechnical University   ²Shenzhen Research Institute of Northwestern Polytechnical University   ^*Corresponding Author

📜IEEE Xplore | 🌐Project Page (Coming Soon) | 📹Demo Video

---

📋 Abstract

High Dynamic Range (HDR) imaging aims to reconstruct high-quality HDR images from multiple Low Dynamic Range (LDR) images with different exposures. Existing methods primarily focus on spatial domain alignment and fusion, often neglecting the critical frequency information that is essential for preserving fine details and textures. To address these challenges, we introduce a Dual-Domain Parallel Fusion Network with Progressive Refinement (DDPF-PR) that explicitly models spatial-frequency interaction for HDR reconstruction. Our framework decouples feature learning into spatial and frequency branches, enabling complementary information aggregation from both domains. Extensive experiments demonstrate that DDPF-PR consistently outperforms state-of-the-art methods on standard HDR benchmarks.

---

🔥 Update Log

• [2026/02] 📢 📢 Paper accepted by IEEE TCSVT!
• [2026/03] 🚀 Code and pretrained models released.

---

📖 Method Overview

The proposed DDPF-PR framework consists of three key components:

• Spatial Branch: Captures local spatial details and structural information
• Frequency Branch: Models global frequency characteristics and texture patterns
• Progressive Refinement Module: Fuses dual-domain features with iterative enhancement

---

🛠️ Prerequisites

• Python 3.8+
• PyTorch 1.10+
• CUDA 11.0+ (for GPU training)
• Accelerate (for distributed training)

---

🌍 Create Environment

Create and activate the Conda environment named ddpf with Python 3.8:

conda create -n ddpf python=3.8 -y
conda activate ddpf

Install Python dependencies from requirements.txt:

pip install -r requirements.txt

---

⬇️ Dataset Preparation

Prepare the train and test datasets following the Datasets section in our paper.

We evaluate our method on the following HDR datasets:

Training Datasets

Dataset	Description	Download
Kalantari et al. (SIG17)	74 training scenes with 3 exposure brackets	Link
Hu et al.	Real-world challenging scenes	-
TEL	Extreme dynamic range scenes	-

Testing Datasets

Dataset	Scenes	Characteristics
Kalantari Test Set	15	Standard HDR benchmark
Tursun et al.	79	Large motion, saturated regions
Prabhakar et al.	65	Night-time HDR scenes
Hu et al.	20	Real-world challenging scenes
TEL	10	Extreme dynamic range scenes

Directory Structure

datasets/
├── train/
│   ├── kalantari/
│   │   ├── training/          # LDR image sequences (3 exposures)
│   │   └── val/               # Ground truth HDR images
│   ├── Hu/
│   │   ├── training/
│   │   └── val/
│   └── TEL/
│       ├── training/
│       └── val/
└── test/
    ├── kalantari/             # 15 test scenes
    ├── Hu/                    # 20 real-world challenging scenes
    ├── TEL/                   # 10 extreme dynamic range scenes
    ├── Tursun/                # 79 large motion scenes
    └── Prabhakar/             # 65 night-time scenes

---

🚀 Getting Started

Training

Train the DDPF-PR model using Accelerate:

CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch train.py --dataset_dir /path/to/dataset --logdir /path/to/log

For single GPU training:

CUDA_VISIBLE_DEVICES=0 accelerate launch train.py --dataset_dir /path/to/dataset --logdir /path/to/log

Key training arguments:

• --dataset_dir: Path to the training dataset
• --logdir: Directory to save checkpoints and logs
• --batch_size: Training batch size (default: 12)
• --lr: Learning rate (default: 0.0002)
• --epochs: Number of training epochs (default: 100)
• --resume: Path to checkpoint for resuming training

Testing

Run inference on test datasets:

python fullimagetest.py --checkpoint /path/to/checkpoint.pth --input /path/to/test/data --output /path/to/results

---

✨ Qualitative Results

Kalantari Dataset Results

Visual comparison on Kalantari dataset with large motion and saturation.

Tursun Dataset Results

Visual comparison on Tursun dataset with large motion.

Hu and TEL Dataset Results

Visual comparison on Hu and TEL datasets with extreme dynamic range.

---

✨ Quantitative Results

Results on Kalantari Test Set

Method	PSNR-μ↑	PSNR-l↑	SSIM-μ↑	SSIM-l↑	HDR-VDP-2↑
DHDR [41]	41.64	40.91	0.9869	0.9858	60.30
AHDR [20]	43.62	41.03	0.9900	0.9862	62.30
NHDRR [42]	42.41	41.08	0.9887	0.9861	61.21
HDR-GAN [34]	43.92	41.57	0.9905	0.9865	65.45
APNT [30]	43.94	41.61	0.9898	0.9879	64.05
CA-ViT [18]	44.32	42.18	0.9916	0.9884	66.03
HyHDR [43]	44.64	42.47	0.9915	0.9894	66.05
SCTNet [2]	44.43	42.21	0.9918	0.9891	66.64
DiffHDR [36]	44.11	41.73	0.9911	0.9885	65.52
SAFNet [44]	44.66	43.18	0.9919	0.9901	66.11
LFDiff [21]	44.76	42.59	0.9919	0.9906	66.54
AFUNet [33]	44.91	42.59	0.9923	0.9906	66.75
DDPF-PR (Ours)	44.93	42.61	0.9922	0.9908	66.78

Results on TEL Dataset

Method	PSNR-μ↑	PSNR-l↑	SSIM-μ↑	SSIM-l↑	HDR-VDP-2↑
DHDR [41]	40.05	43.37	0.9794	0.9924	67.09
AHDR [20]	42.08	45.30	0.9837	0.9943	68.80
NHDRR [42]	36.68	39.61	0.9590	0.9853	65.41
HDR-GAN [34]	41.71	44.87	0.9832	0.9949	69.57
CA-ViT [18]	42.39	46.35	0.9848	0.9948	69.23
SCTNet [2]	42.55	47.51	0.9850	0.9952	70.66
DiffHDR [36]	42.18	45.63	0.9841	0.9946	69.88
SAFNet [44]	42.68	47.46	0.9792	0.9955	68.16
AFUNet [33]	43.31	47.83	0.9876	0.9959	71.08
DDPF-PR (Ours)	43.49	48.25	0.9878	0.9961	70.96

Results on Hu Dataset

Method	PSNR-μ↑	PSNR-l↑	SSIM-μ↑	SSIM-l↑	HDR-VDP-2↑
DHDR [41]	41.13	41.20	0.9870	0.9941	70.82
AHDR [20]	45.76	49.22	0.9956	0.9980	75.04
NHDRR [42]	45.15	48.75	0.9956	0.9981	74.86
HDR-GAN [34]	45.86	49.14	0.9945	0.9989	75.19
APNT [30]	46.41	47.97	0.9953	0.9986	73.06
CA-ViT [18]	48.10	51.17	0.9947	0.9989	77.12
HyHDR [43]	48.46	51.91	0.9959	0.9991	77.24
DiffHDR [36]	48.03	50.23	0.9954	0.9989	76.22
SCTNet [2]	48.10	51.03	0.9963	0.9991	77.14
SAFNet [44]	47.18	49.35	0.9951	0.9990	76.83
LFDiff [21]	48.74	52.10	0.9968	0.9993	77.35
AFUNet [33]	48.83	52.13	0.9968	0.9991	77.44
DDPF-PR (Ours)	48.86	52.42	0.9969	0.9992	77.46

---

📦 Pretrained Models

Download our pretrained model:

Baidu Pan: 44.95sig17.pth (提取码: zxat)

Place the downloaded checkpoint in your working directory and specify the path when running fullimagetest.py.

---

📏 Troubleshooting

• CUDA / PyTorch mismatch: Verify installed torch wheel matches your CUDA toolkit version. Reinstall torch if necessary.
• Accelerate configuration: Run accelerate config to set up distributed training configuration.
• Out of memory: Reduce --batch_size or use gradient accumulation.
• Windows users: Some shell scripts use bash; run them under WSL or adapt commands for PowerShell.

---

💖 Acknowledgment

We thank Qingsen Yan and Tao Hu for their support and guidance throughout this research.

---

🤝🏼 Citation

If this code contributes to your research, please cite our work:

@article{zhou2026high,
  title={High Dynamic Range Imaging via Spatial-Frequency Interaction},
  author={Zhou, Weiyu and Yang, Yongqing and Hu, Tao and Hui, Pu and Jin, Jian and Cao, Yu and Yan, Qingsen and Zhang, Yanning},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2026},
  publisher={IEEE}
}

---

🔆 Contact

If you have any questions, please feel free to contact:

• Weiyu Zhou: weiyuzhou@mail.nwpu.edu.cn

Or open an issue in this repository.