DIFNet

Introduction

This is a unofficial PyTorch implementation of Unsupervised Deep Image Fusion With Structure Tensor Representations with multi-GPU training supports.

Visible image	Infrared image	Fused image

Prerequisites

• PyTorch
• tqdm
• pillow

How to train

• For training, you should get image pairs such as RGB-thermal, RGB-NIR or multi-focus pairs. And they must be registered each other.
• Set your training phase at args.py
• For multi-GPU training, you should set parameters as follows,

### args.py
# For GPU training
world_size = -1
rank = -1
dist_backend = 'nccl'
gpu = 0,1,2,3
multiprocessing_distributed = True
distributed = None

You can see details of these parameters at tutorials of PyTorch official documents.

• If you have pretrained models, you can transfer the training to them.

### args.py
# resume = "./models/samples.pt" # Transfer learning
resume = None # Train from scratch

• Make two txt files which contain paths of RGB and NIR/thermal images, respectively. For example,

# RGB images
/home/kim/images_visible/1.jpg
/home/kim/images_visible/2.jpg
/home/kim/images_visible/3.jpg

# Thermal images
/home/kim/images_thermal/1.jpg
/home/kim/images_thermal/2.jpg
/home/kim/images_thermal/3.jpg

• Run the command below

python train.py

How to test

• Please add the path of datasets for test.

### args.py
# For testing
test_save_dir = "./"
test_visible = "./test_visible.txt"
test_lwir = "./test_lwir.txt"

• For training and testing, I recommand KAIST Multispectral Pedestrian Detection Benchmark.

To-Do list

• Add evaluation code
• Upload pretrained models and samples
• Update benchmark

Reference

@ARTICLE{8962327,  
         author={H. {Jung} and Y. {Kim} and H. {Jang} and N. {Ha} and K. {Sohn}},  
         journal={IEEE Transactions on Image Processing},   
         title={Unsupervised Deep Image Fusion With Structure Tensor Representations},  
         year={2020},  
         volume={29},  
         number={},  
         pages={3845-3858},}