🐼 Hipandas: Hyperspectral Image Joint Denoising and Super-Resolution Framework

Codes for "Hipandas: Hyperspectral Image Joint Denoising and Super-Resolution by Image Fusion with the Panchromatic Image" (ICCV 2025)

Shuang Xu(徐爽), Zixiang Zhao(赵子祥), Haowen Bai(白浩闻), Chang Yu(余畅), Jiangjun Peng(彭江军), Xiangyong Cao(曹相湧), Deyu Meng(孟德宇)

Hipandas: a novel learning paradigm that reconstructs high quality HSIs from noisy low-resolution HSIs (NLRHS) and high-resolution PAN images

📌 Overview

This project implements the UHipandas method proposed in the paper "Hipandas: Hyperspectral Image Joint Denoising and Super-Resolution by Image Fusion with the Panchromatic Image" (ICCV 2025). Traditional hyperspectral image processing treats denoising and super-resolution as separate sequential tasks, leading to accumulated errors and suboptimal results. Hipandas introduces a novel joint framework that simultaneously addresses both challenges by fusing noisy low-resolution hyperspectral images (NLRHS) with high-resolution panchromatic (PAN) images, achieving superior spatial resolution and spectral fidelity.

🧠 Methodology

Core Innovation

Hipandas revolutionizes hyperspectral image restoration through a unified optimization paradigm that replaces traditional sequential processing with simultaneous denoising and super-resolution. This approach leverages the complementary strengths of:

• Noisy low-resolution hyperspectral images (rich spectral information)
• High-resolution panchromatic images (detailed spatial information)

By integrating these modalities in a joint framework, Hipandas overcomes the limitations of step-by-step processing pipelines that propagate errors between stages .

Key Technical Contributions

1. Two-Stage Training Strategy

   
   The two-stage training strategy of UHipandas

   • Stage 1 (Pretraining): Individual networks are pretrained to establish strong initial parameters
   • Stage 2 (Joint Training): Networks are optimized together with cross-network constraints to ensure consistency

2. Dual Low-Rank Priors

   • HSI Low-Rank Prior: Captures global spectral correlations to preserve spectral integrity
   • Detail-Oriented Low-Rank Prior: Enhances local spatial details from PAN images

Network Architecture

The framework comprises three interconnected components working in synergy:

1. GDN (Guided Denoising Network)

   • Utilizes low-rank matrix decomposition combined with deep learning
   • Learns spectral-spatial correlations to suppress noise while preserving critical features
   • Incorporates gated recurrent convolution units for effective spatio-spectral feature extraction
   • Takes noisy LRHS as input and produces clean low-resolution HSIs

2. GSRN (Guided Super-Resolution Network)

   • Enhances spatial resolution using both denoised HSIs and PAN image guidance
   • Implements multi-scale feature fusion to propagate fine details from PAN to HSI
   • Uses low-rank decomposition to maintain spectral consistency during upsampling

3. PRN (Panchromatic Reconstruction Network)

   • Predicts PAN images from super-resolved HSIs to enforce cross-modal consistency
   • Implements cross-layer guided attention mechanisms for effective feature alignment
   • Serves as a spectral-spatial consistency check to prevent distortion

Working Principle

1. GDN first removes noise from input NLRHS using spectral low-rank properties
2. GSRN upsamples the denoised HSI while integrating spatial details from PAN
3. PRN reconstructs PAN from the super-resolved HSI, creating a feedback loop that:
   • Ensures spatial details from PAN are appropriately transferred
   • Preserves spectral characteristics of the original HSI
   • Minimizes spectral distortion common in traditional fusion methods

🛠️ Requirements

• Python
• PyTorch
• Kornia
• NumPy
• SciPy
• Matplotlib
• scikit-image

📁 Project Structure

Hipandas/
├── model.py              # Network architectures (GDN, GSRN, PRN) 
├── simulate_data.py      # Generate noisy hyperspectral datasets with various noise models   
├── main.py               # Main training & evaluation pipeline (orchestrates model training/inference)   
├── eval_metric.py        # Calculate quantitative metrics (PSNR, SSIM, SAM, ERGAS)   
├── utils/                # Core utility modules  
│   ├── common.py         # Basic utilities (seed setup, device config, data conversion)  
│   ├── metrics.py        # Core metric calculation functions (used by eval_metric.py)  
│   ├── rsshow.py         # Visualization tools for hyperspectral/panchromatic images  
│   ├── spectral_tools.py # Spectral processing utilities (e.g., spectral response handling)  
│   └── noise_model.py    # Noise generation implementations (Gaussian, impulse, mixed noise) 
├── data/                 # Dataset storage (to be created; populated by simulate_data.py)  
└── result/               # Output directory for reconstructed images & metrics (auto-created)

📥 Dataset Preparation

1. Create a data directory in the project root
2. Download the base data Dongying_1_1.mat Link: https://pan.baidu.com/s/15hdJooFUbSzGbbz1gMdAyQ?pwd=nwpu Password: nwpu
3. Place Dongying_1_1.mat into the data directory:

   data/
   └── Dongying_1_1.mat          
   

4. Run the data simulation script to generate various noise cases:

   python simulate_data.py
   

5. The script will create organized subdirectories with simulated noise:

   data/
   ├── g10/           # Gaussian noise (σ=10)
   │   ├── Dongying_0_0.mat
   │   ├── Dongying_0_1.mat
   │   └── ... (81 images total)
   ├── g30/           # Gaussian noise (σ=30)
   ├── gni/           # Gaussian + impulse noise
   ├── mix15/         # Mixed noise 1
   ├── mix35/         # Mixed noise 2
   └── mix55/         # Mixed noise 3
   

6. Each MAT file contains:
   • I_GT: Ground truth hyperspectral image
   • I_LRHS: Low-resolution hyperspectral image
   • I_PAN: Panchromatic image
   • N_LRHS: Noisy low-resolution hyperspectral image

🚀 Running the Code

Basic Execution

Process all noise cases with default parameters:

python main.py

Key Parameters (configurable in main.py)

• noise_case: List of noise scenarios to process (['g10', 'g30', 'gni', 'mix15', 'mix35', 'mix55'])
• lr: Learning rate (default: 1e-3)
• num_epoch: Training epochs for pretraining and main training (default: [400, 600])
• rank: Rank parameter for low-rank decomposition modules

Result Output

Results are saved in the following directory structure:

result/
└── UHipandas/
    ├── g10/
    │   ├── Dongying_0_0.mat
    │   └── ...
    ├── g30/
    └── ...

Output includes:

• Reconstructed high-resolution hyperspectral images
• Evaluation metrics printed to console (PSNR, SSIM, SAM, ERGAS)
• Runtime statistics for each processing stage

📊 Evaluation

Run the evaluation script to generate comprehensive performance metrics:

python eval_metric.py

This produces an Excel file with quantitative results:

• PSNR (Peak Signal-to-Noise Ratio) - measures reconstruction fidelity
• SSIM (Structural Similarity Index) - assesses structural preservation
• ERGAS (Error Relative Global Accuracy in Synthesis) - evaluates spectral-spatial quality
• SAM (Spectral Angle Mapper) - quantifies spectral distortion

🔍 Performance Highlights

Results with Gaussian noise

Results with mixed noise

📚 Citation

If you use this code in your research, please cite our paper:

@inproceedings{UHipandas,
  author       = {Shuang Xu and 
                  Zixiang Zhao and 
                  Haowen Bai and 
                  Chang Yu and 
                  Jiangjun Peng and 
                  Xiangyong Cao and 
                  Deyu Meng},
  title        = {Hipandas: Hyperspectral Image Joint Denoising and Super-Resolution by Image Fusion with the Panchromatic Image},
  booktitle    = {International Conference on Computer Vision (ICCV)},
  pages        = {},
  year         = {2025},
}

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

⭐ If you find this project useful, please give it a star! Your support motivates us to improve further.