SeAFusion

This is official Pytorch implementation of "Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network"

Welcome to follow the further work of our SeAFusion：Rethinking the necessity of image fusion in high-level vision tasks: A practical infrared and visible image fusion network based on progressive semantic injection and scene fidelity 【Paper】, 【Code】.

Framework

The overall framework of the proposed semantic-aware infrared and visible image fusion algorithm.

Network Architecture

The architecture of the real-time infrared and visible image fusion network based on gradient residual dense block.

To Train

Run **CUDA_VISIBLE_DEVICES=0 python train.py** to train your model. The training data are selected from the MFNet dataset. For convenient training, users can download the training dataset from here, in which the extraction code is: bvfl.

The MFNet dataset can be downloaded via the following link: https://drive.google.com/drive/folders/18BQFWRfhXzSuMloUmtiBRFrr6NSrf8Fw.

The MFNet project address is: https://www.mi.t.u-tokyo.ac.jp/static/projects/mil_multispectral/.

To Test

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

For quantitative evaluation

For quantitative assessments, please follow the instruction to modify and run . /Evaluation/test_evaluation.m .

Recommended Environment

• [ ] torch 1.7.1
• [ ] torchvision 0.8.2
• [ ] numpy 1.19.2
• [ ] pillow 8.0.1

Fusion Example

Qualitative comparison of SeAFusion with 9 state-of-the-art methods on 00633D image from the MFNet dataset.

Segmentation Results

Segmentation results for infrared, visible and fused images from the MFNet dataset. The segmentation models are re-trained on infrared, visible and fused image sets. Each two rows represent a scene.

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{TANG202228SeAFusion,
title = {Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network},
journal = {Information Fusion},
volume = {82},
pages = {28-42},
year = {2022},
issn = {1566-2535}
}