BLFusion-MRLL

Breaking Low-Light Fusion Barrier: Unsupervised Darkness and Noise-Aware Visible and Infrared Image Fusion Network

Multi-modal Real-world Low-light Dataset

Official repository for BLFusion and the MRLL dataset.

---

📌 Overview

BLFusion is an unsupervised darkness- and noise-aware infrared-visible image fusion framework for challenging nighttime scenes. It is designed to jointly handle two major low-light barriers: insufficient illumination and complex real-world noise.

Existing low-light fusion methods often focus on brightness enhancement while overlooking noise suppression. Degradation-aware supervised methods can model multiple degradations, but they usually rely on paired high-quality references and may suffer from imbalanced optimization between enhancement and denoising.

BLFusion addresses these issues with a customized two-stage unsupervised framework. The first stage enhances dark visible images through Retinex-guided decomposition with state space modeling. The second stage performs denoising and fusion jointly, producing fused images with clearer structures, richer textures, and stronger robustness under real low-light noise.

✨ Highlights

• Darkness- and noise-aware fusion: simultaneously improves illumination and suppresses complex noise in low-light infrared-visible fusion.
• Retinex-guided SSM decomposition: combines Retinex theory with state space modeling to enhance dark visible images while preserving structural details.
• Unsupervised denoising fusion: integrates pixel-shuffle downsampling, blind-spot networks, and dilated convolutions to learn noise-free features without clean references.
• Maximum replacement refinement: reduces visual artifacts and further improves fused image quality during inference.
• MRLL dataset: provides 500 aligned infrared-visible image pairs captured from real-world indoor and outdoor nighttime scenes.

🧩 Framework

BLFusion contains two dedicated stages:

1. State Space Model-Based Decomposition Network: decomposes low-light visible images into illumination and reflectance components. The Retinex-inspired design models illumination degradation, while efficient state space modules capture long-range dependencies and preserve fine visual details.
2. Unsupervised Denoising and Fusion Network: extracts noise-suppressed features from enhanced visible and infrared images. Pixel-shuffle downsampling disrupts spatial noise correlation, blind-spot networks avoid identity mapping, and dilated convolutions expand contextual perception for robust denoising and fusion.

🗓️ Release Plan

The full project will be released after the official paper process is complete. Planned materials include:

• Source code for the official BLFusion implementation
• Pretrained model checkpoints
• Testing and evaluation scripts
• Training scripts and configuration files
• MRLL dataset download links
• Visualization tools and qualitative comparison results
• Detailed usage instructions and reproducibility notes

🛠️ Installation

Create a conda environment named BLFusion and install the core dependencies:

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

pip install torch==2.4.1 torchvision==0.19.1
pip install -r requirements.txt

The project uses Mamba/SSM modules. If mamba-ssm or causal-conv1d fails to build, please make sure the local CUDA toolkit and PyTorch CUDA version are compatible before reinstalling the requirements. I previously referred to the following tutorial for installation guidance: https://blog.csdn.net/qq_45100200/article/details/139754231.

📦 Pretrained Models

Pretrained checkpoints are available from Baidu Drive:

Source	Link	Access Code
Baidu Drive	Download	yvkz

After downloading BLFusion_weights, place the checkpoint files into the corresponding folders:

Enhancement_Network/
└── weights/
    └── checkpoint_Diff_TDN.pth

Fused_Denoise_Network/
└── weights/
    ├── double_noise/
    │   └── checkpoint_Diff_TDN.pth
    ├── mul_noise/
    │   └── checkpoint_Diff_TDN.pth
    ├── no_noise/
    │   └── checkpoint_Diff_TDN.pth
    └── ran_noise/
        └── checkpoint_Diff_TDN.pth

These paths match the default checkpoint locations used by the provided testing scripts.

🧪 Testing

BLFusion follows a two-stage testing workflow. First, run the enhancement network to generate enhanced visible images. Then copy the enhancement results into the fusion network test set and run the denoising-fusion network.

Stage 1: Visible Image Enhancement

Enter the enhancement network folder:

cd Enhancement_Network

Check test.py and modify the dataset path, checkpoint path, and result-saving path if needed:

read_data("dataset/MRLL")
model_weight_path = "./weights/checkpoint_Diff_TDN.pth"
root = "./results/MRLL"

Run the enhancement script:

python test.py

The enhancement results will be saved under:

Enhancement_Network/results/MRLL/

Stage 2: Denoising and Fusion

Copy the enhanced visible results to the test directory used by the fusion network. For example:

cd ..
cp -r Enhancement_Network/results/MRLL/vi_l Fused_Denoise_Network/dataset/MRLL/test/
cp -r Enhancement_Network/results/MRLL/vi_r Fused_Denoise_Network/dataset/MRLL/test/
cp -r Enhancement_Network/results/MRLL/vi Fused_Denoise_Network/dataset/MRLL/test/
cp -r Enhancement_Network/results/MRLL/ir Fused_Denoise_Network/dataset/MRLL/test/

Then enter the fusion network folder:

cd Fused_Denoise_Network

Check test.py and modify the dataset path, checkpoint path, and fusion result-saving path if needed:

read_data_test("dataset/MRLL")
model_weight_path = "./weights/mul_noise/checkpoint_Diff_TDN.pth"
root = "./results/MRLL/"

Select the checkpoint according to the target noise setting:

weights/double_noise/checkpoint_Diff_TDN.pth  # both visible and infrared modalities contain multiplicative noise
weights/mul_noise/checkpoint_Diff_TDN.pth     # visible images contain multiplicative noise
weights/no_noise/checkpoint_Diff_TDN.pth      # no noise setting
weights/ran_noise/checkpoint_Diff_TDN.pth     # random noise setting

Run the fusion script:

python test.py

The fused images will be saved to the result directory specified in test.py, for example:

Fused_Denoise_Network/results/MRLL/fused/

🚀 Training

Similar to Testing, run python train.py !

---

MRLL Dataset

MRLL stands for Multi-modal Real-world Low-Light Dataset. It is a synchronized dual-spectral dataset built for infrared-visible image fusion under realistic nighttime degradation.

MRLL contains 500 well-aligned infrared-visible image pairs with a resolution of 960 x 720. The dataset covers diverse indoor and outdoor low-light scenes, varying illumination levels, and complex real-world noise patterns, making it suitable for evaluating fusion models under practical nighttime conditions.

📥 Dataset Availability

The full MRLL dataset is available at:

Source	Link	Access Code
Baidu Drive	Download	ek7j
Google Drive	Download	-

🔍 Synchronized Dual-Spectral Imaging and Registration

MRLL is collected with synchronized visible and infrared imaging devices (DJI Matrice 4T). The visible and infrared images are aligned through cross-modal registration to obtain paired dual-spectral samples for fusion research.

🌃 Scenario Schematic

MRLL covers diverse real-world low-light scenarios, including indoor and outdoor nighttime scenes with weak illumination, complex lighting, texture degradation, and challenging visibility.

📊 Comparison with Existing Low-Light Fusion Datasets

Compared with existing low-light infrared-visible fusion datasets, MRLL provides synchronized real-world paired data, indoor and outdoor coverage, and multiple realistic noise patterns.

📚 Citation

If you find this work or dataset useful for your research, please consider citing our paper:

@article{chen2026breaking, 
   title={Breaking Low-Light Fusion Barrier: Unsupervised Darkness and Noise-Aware Visible and Infrared Image Fusion Network},
   author={Chen, Jiajun and Xu, Han and Liu, Guangcan and Ma, Jiayi},
   journal={IEEE Transactions on Image Processing},
   year={2026},
   publisher={IEEE}
}

📬 Contact

For questions about BLFusion or the MRLL dataset, please open an issue in this repository.