An advanced framework for polarization image fusion using analytical attention heads to achieve high-quality fusion of RGB and polarization images with noise suppression and feature enhancement. Paper's share link (available before 2026-3-10): Elsevier - Optics and Lasers in Engineering

This repository implements AGPFusion, a novel attention-guided polarization image fusion method designed for industrial inspection applications. The framework intelligently combines RGB and polarization images through multiple analytical attention mechanisms including gradient, texture, semantic, entropy, and noise-aware heads, producing enhanced fused images with improved contrast and detail preservation.

• Multi-Head Attention Architecture: Combines 5 specialized attention heads (GAH, TAH, EAH, SAH, NAH)
• Intelligent Noise Suppression: Adaptive local variance-based noise attenuation with configurable soft/hard masking
• Multi-Scale Processing: Automatic block size adaptation for different image resolutions
• Comprehensive Evaluation: 14 quantitative metrics (CC, MI, SSIM, PSNR, NIQE, NIMA, Qabf, etc.)
• Batch Processing: Efficient processing of entire datasets with detailed logging
• Flexible Configuration: Extensive parameter tuning for domain-specific optimization

DeepFusion/
├── fusion_agp.py              # Main fusion script
├── metric.py                   # Evaluation metrics tool
├── core/
│   ├── AGPFusion.py           # Core fusion implementation
│   ├── metric.py              # Metric calculations
│   ├── tensor.py              # Tensor utilities
│   └── common.py              # Common functions
├── images/                     # Documentation figures
└── README.md                  # This file

pip install torch torchvision
pip install opencv-python scikit-image matplotlib tqdm
pip install kornia pytorch-grad-cam natsort

python fusion_agp.py \
    --rgb_input "path/to/rgb/image.png" \
    --pol_input "path/to/polarization/image.png" \
    --outputdir "path/to/output" \
    --single_img_test

python fusion_agp.py \
    --rgb_input "datasets/F12CCP/rgb/*.png" \
    --pol_input "datasets/F12CCP/pol/*.png" \
    --outputdir "results/fusion_batch" \
    --img_size 1024,768

# Single folder evaluation
python metric.py \
    --vis_src "datasets/F12CCP/rgb" \
    --pol_src "datasets/F12CCP/pol" \
    --fusion_results "results/fusion_batch"

# Batch evaluation (all subfolders)
python metric.py \
    --vis_src "datasets/F12CCP/rgb" \
    --pol_src "datasets/F12CCP/pol" \
    --fusion_root "results/parameter_sweep"

The framework computes 14 comprehensive metrics:

Results are saved to metric.csv with per-image values and summary statistics (mean & variance).

We provide the Industry-Polarization-Dataset containing 416 sets of polarization images for industrial product inspection:

• DoFP Camera: Daheng MER2-503-36U3M POL (Sony IMX264 MZR, 2448×2048)
• Time-Sequential: Custom rotating polarization imaging system
• Polarization Types: DoFP, linear polarization sequences, full Stokes images
• Challenges: Specular reflection, low contrast, no registration required

Download Link: https://pan.quark.cn/s/9c8fe6e6db79

Benchmark results on the F12CCP dataset demonstrate superior performance compared to state-of-the-art methods.

# 5 Attention Heads working collaboratively:
1. GAH: Sobel + Laplacian filters for gradient features
2. TAH: LBP + Canny edge detection for texture patterns
3. EAH: Block-wise entropy for information-rich regions
4. SAH: EfficientNet GradCAM++ for semantic saliency
5. NAH: Adaptive local variance with entropy guidance 

1. Attention Map Computation: Multi-head feature extraction
2. Weight Map Generation: Cross-source attention comparison
3. Image Decomposition: Base/detail layer separation (average pooling)
4. Guided Filtering: Edge-preserving smoothing of weight maps
5. Layer Fusion: Weighted combination of base/detail components
6. Post-processing: CLAHE contrast enhancement

F12CCP Dataset results demonstrate:

• Superior Edge Preservation: High Qabf scores
• Noise Robustness: Low Nabf values
• Contrast Enhancement: Improved AG, EN, and SF metrics
• Visual Quality: Competitive NIQE and NIMA scores

If you use this code or dataset in your research, please cite:

@article{ZHOU2026109628,
title = {Polarization image fusion via analytical attention heads: A multi-scale feature integration framework},
journal = {Optics and Lasers in Engineering},
volume = {201},
pages = {109628},
year = {2026},
issn = {0143-8166},
doi = {https://doi.org/10.1016/j.optlaseng.2026.109628},
url = {https://www.sciencedirect.com/science/article/pii/S014381662600028X},
author = {Junzhuo Zhou and Jun Zou and Ye Qiu and Zhihe Liu and Jia Hao and Wenli Li and Yiting Yu},
keywords = {Polarization imaging, Multimodal image fusion, Computer vision, Industrial application, Surface inspection},
}

Contributions are welcome! Please feel free to submit pull requests or open issues for bugs and feature requests.

This project is licensed under the MIT License.