A Novel Deep Video Watermarking Framework with Enhanced Robustness to H.264/AVC Compression

ACM MM 2023

Yulin Zhang, Jiangqun Ni, Wenkang Su, and Xin Liao

Introduction

This repository is a code release for the paper found here. The paper focus on deep video watermarking with temporal robustness and invisibility. The main contributions are the proposed temporal-associated feature extraction block (TAsBlock), differentiable video compression simulator(DiffH264), and spatial/temporal mask loss.

Citation

If you find our work useful, please consider citing:

@inproceedings{zhang2023novel,
  title={A Novel Deep Video Watermarking Framework with Enhanced Robustness to H. 264/AVC Compression},
  author={Zhang, Yulin and Ni, Jiangqun and Su, Wenkang and Liao, Xin},
  booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
  pages={8095--8104},
  year={2023}
}

License

The models are free for non-commercial and scientific research purpose. Please mail us for further licensing terms.

References

1. The optic flow estimation code is based on sniklaus/pytorch-spynet. The original paper is

@inproceedings{Ranjan_CVPR_2017,
    author = {Ranjan, Anurag and Black, Michael J.},
    title = {Optical Flow Estimation Using a Spatial Pyramid Network},
    booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
    year = {2017}
}

2. The code for intra compression and residual compression is based on mlomnitz/DiffJPEG.