SAN_and_HDU-Net

Official repository for "Image splicing forgery detection by combining synthetic adversarial networks and hybrid dense U‐net based on multiple spaces"
If you think the repository is helpful, give me a star to let me know.

Requirements

1. Python: 3.7.0
2. CUDA: 11.4
3. Pytorch: 1.8.0

Data Preparation

1. For SAN:
   Merge CASIA v2.0 and Forensics, and put the combination dataset (1891 images) into the subdirs /img/ and /mask/.

   HDU-Net/
   ├── ...
   └── SAN/
       ├── img/
       └── mask/
   

   Generate a dataset that I call $SF-Data$ (82608 images). You can download $SF-Data$ via the link https://drive.google.com/file/d/1IoG78dAcxyw5fRPo1DjisKUoykJQTs2_/view?usp=sharing.

2. For HDU-Net:
   To generate edge information according to the subdir /mask/， run

   python SAN/generate_edge.py
   

Getting Started

Training in Command Line

1. For SAN:
   If you want to retrain SAN, run

   python SAN/train.py

2. For HDU-Net:
   To train HDU-Net, run

   python train.py

You should change diverse parameters in options.py

Evaluation in Command Line

1. For SAN:
   I provide a well-trained model weight best_model_for_SAN.pth.
   You can download the weights via https://drive.google.com/file/d/1Qbn3kCxwMA7r-VQ0mpnXaI1tKKetPn7n/view?usp=sharing, and put it into the subdir /SAN/.
   You can use it to generate dataset based on other datasets like COCO, etc. I have converted the multi-label annotations "train2017" in COCO to binary mask. I are hesitating to upload the dataset since it is too large.
   After running the following codes, you should change the path of dataset in options_GAN.py.
   Note that the well-trained weight only accept binary mask.

   python SAN/generate_data.py

2. For HDU-Net:
   You can download the weights of HDU-Net via https://drive.google.com/file/d/1XDMZdGzxSvs22j5uKx6Ywm92JlgFv_iw/view?usp=sharing.
   Run

   python inference.py

Citation

If you find this project useful for your research, please use the following BibTeX entry.

@article{Wei2022ImageSF,
  title={Image splicing forgery detection by combining synthetic adversarial networks and hybrid dense U‐net based on multiple spaces},
  author={Yang Wei and Jianfeng Ma and Zhuzhu Wang and Bin Xiao and Wenying Zheng},
  journal={International Journal of Intelligent Systems},
  year={2022}
}

License

This project is released under the MIT license.

Contact

Contact yale ywei9395@gmail.com for any further information.