davin-qi530/divfusion

DIVFusion: Darkness-free infrared and visible image fusion

DIVFusion

DIVFusion: Darkness-free infrared and visible image fusion (^▽^)

This is official Tensorflow implementation of "DIVFusion: Darkness-free infrared and visible image fusion"

Framework

The overall framework of the proposed DIVFusion. SIDNet is a network to seperate illumination degradation. TCEFNet integrates and enhances the complementary information of source images.

Network Architecture1

The framework of the scene-illumination disentangled network (SIDNet).

Network Architecture2

The detailed structure of (a) gradient retention module (GRM) and (b) contrast block.

Before Train

**conda env create -f XXY_DIVFusion.yaml**

To Train

Add VGG16.npy from to the file. Link:here, in which the extraction code is: 1xo5.

First Run **CUDA_VISIBLE_DEVICES=0 python decomposition.py** to train your model(SIDNet).

Second Run **CUDA_VISIBLE_DEVICES=0 python fusion_enhancement_new.py** to train your model(TCEFNet).

The training data are selected from the LLVIP dataset. For convenient training, users can download the training dataset from here, in which the extraction code is: he31. Dataset should be send in ./ours_dataset_240/train/ ./ours_dataset_240/test/

The LLVIP dataset can be downloaded via the following link: here.

To Test

The checkpoint can be found via the following link: here, in which the extraction code is: dv3s. The testing data are selected from the LLVIP dataset. link: here The extraction code is: 6hvf

Run **CUDA_VISIBLE_DEVICES=0 python test.py** to test the whole model.

Recommended Environment

numpy=1.19.2
opencv=3.4.2
python=3.6.12
tensorflow-gpu=1.14.0
scipy==1.5.4

Demo

Nighttime infrared and visible image fusion results.

Fusion Example

Vision quality comparison of our method with seven SOTA fusion methods on #010064 and #060193 images from LLVIP dataset.

Two-stage Fusion Example

Vision quality comparison of two-stage fusion experiments. Each row represents a scene, and from top to bottom is #21006, #220312, and #260092 images from LLVIP dataset. ((a)-(b): source images, (c)-(i): two-stage fusion results by different enhancement methods and fusion methods, (j): our fusion result).

Generalization Example

Segmentation results for infrared, visible and fused images from the MFNet dataset. The segmentation model is Deeplabv3+, pre-trained on the Cityscapes dataset. Each two rows represent a scene.

Detection Results

Object detection results for infrared, visible and fused images from the MFNet dataset. The YOLOv5 detector, pre-trained on the Coco dataset is deployed to achieve object detection.

If this work is helpful to you, please cite it as：

@article{Tang2022DIVFusion,
  title={DIVFusion: Darkness-free infrared and visible image fusion},
  author={Tang, Linfeng and Xiang, Xinyu and Zhang, Hao and Gong, Meiqi and Ma, Jiayi},
  journal={Information Fusion},
  volume = {91},
  pages = {477-493},
  year = {2023},
  publisher={Elsevier}
}