CurveMEF: Multi-Exposure Fusion via Curve Embedding Network

Fast lightweight physical informed multi-exposure fusion model, Neurocomputing.

Introduction

This is official Pytorch implementation of CurveMEF: Multi-Exposure Fusion via Curve Embedding Network.

Framework

⚡️Quick Start

First, install requirements and setup CurveMEF following installation guide. The pre-trained weight was trained by config/curve/curvemef_rgb.yml and config/curve/curvemef_y.yml.

• Inference

python inference.py -config CONFIG_PATH

---

Install

Requirements

• Linux or Windows
• CUDA >= 10.0
• Python >= 3.6
• Pytorch >= 1.9
• Lightning >= 1.9

Step

1. Create a conda virtual environment and then activate it.

    conda create -n curvemef python=3.9
    conda activate curvemef

2. Install Pytorch and Cuda

   # For Cuda 11.3
   pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
   # For Cuda 10.2
   pip install torch==1.12.1+cu102 torchvision==0.13.1+cu102 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu102

3. Clone this repository

   git clone https://github.com/PiratePai/CurveMEF.git
   cd curvemef

4. Install requirements

   pip install lightning==2.0.4
   pip install torchmetrics==0.11.4
   pip install timm==0.9.2
   pip install termcolor==2.1.0
   pip install pytorch-lightning==1.9.0
   pip install matplotlib==3.7.1
   pip install opencv-contrib-python==4.8.0.74
   pip install termcolor==2.1.0
   pip install numpy==1.23.0
   pip install pyiqa==0.1.7

   OR

   pip install -r requirements.txt

---

---

Train

1. Prepare dataset

   You can download the dataset, or custom your own dataset. Your dataset folder needs to contain two subfolders, high and low, with the same image names in both folders

   [CustomDataset]/high/imgx
   [CustomDataset]/low/imgx

2. Prepare config file

   Copy and modify an example yml config file in config/curvemef.

   Change pro_name to your own project name.

   Change save_dir to where you want to save model.

   Change model_name to your own network.

   Change train_path and test_path to your own dataset.

   Set channel to fit your dataset.

   Set num workers and batch size in device to fit your device.

   Set total_epochs, optim_term according to your dataset and batchsize.

3. Start training

   CurveMEF is using Pytorch Lightning for training.

   python train.py -config config_path

4. Visualize Train performance

   • At the end of each epoch in the training phase, the model tests a pair of multi-exposure images, and the test results are saved in the Samples folder.
   • If you want to replace the test pair with your own images, you can store test1pngy.pt and test1png.pt as your own image pairs. The tensor in test1png.pt is of the size as [1, 6, H, W], and the tensor in test1pngy.pt is of the size as [1, 2, H, W].
   • If you want to remove this feature, you need to comment the self.test_tensor code in model_interface.py.

5. Visualize Logs

   Logs are save in wandb.

   In pro_name of the configuration file, set the project name of wandb.

   If you don't want to use it, just delete all the code that contains wandb in main.py and model_interface.py

6. Convert RGB/YCbCr to YCbCr/RGB

   Run rgb2y.m or y2rgb.m.

   Change A_dir, B_dir, Fused_dir and save_dir to your own path.

---

📖 Benchmarks

Performance Comparison

Visual Comparison

Complexity Comparison

• [1] https://github.com/SunnerLi/DeepFuse.pytorch
• [2] https://github.com/makedede/MEFNet
• [3] https://github.com/Beta-Hu/PMGI
• [4] https://github.com/1318133/LPUIF-A-Lightweight-Pixel-Level-Unified-Image-Fusion-Network
• [5] https://github.com/Linfeng-Tang/SwinFusion
• [6] https://github.com/ytZhang99/U2Fusion-pytorch
• [7] https://github.com/miccaiif/TransMEF
• [8] https://github.com/HaoZhang1018/IID-MEF (We refer to this version for implement pytorch code of IID-MEF.)

---

Citation

If you find this project useful in your research, please consider cite:

@article{PENG2024127915,
title = {CurveMEF: Multi-exposure fusion via curve embedding network},
journal = {Neurocomputing},
volume = {596},
pages = {127915},
year = {2024},
issn = {0925-2312},
doi = {https://doi.org/10.1016/j.neucom.2024.127915},
url = {https://www.sciencedirect.com/science/article/pii/S0925231224006866},
author = {Pai Peng and Zhongliang Jing and Han Pan and Yang Liu and Buer Song},
keywords = {Multi-exposure fusion, Image fusion, Curve estimation, Physical informed neural network, Deep learning}
}

---

❤️ Acknowledgement

Refer to NanoDet, Pytorch-Lightning and pytorch-lightning-template for the overall code framework

The design of CurveNet is based on GhostNetV3.

We also thanks the following public repositories: timm, torchmetrics

📧 Contact

If you have any questions, please email pengpai@sjtu.edu.cn

---