Authors:
Tymoteusz Lindner
1
;
2
;
Tomasz Hawro
2
and
Piotr Syga
3
;
2
Affiliations:
1
Poznan University of Technology, Poznan, Poland
;
2
Vestigit, Poland
;
3
Wrocław University of Science and Technology, Wrocław, Poland
Keyword(s):
Watermarking, QR Codes, Encoder-Decoder, Deep Learning.
Abstract:
The recent advancements in watermarking have indicated the capacity of deep learning for video copyright protection. We introduce a novel deep neural network architecture that uses QR-coded-based messages for video watermarking. Our framework encompasses an encoder-decoder structure, integrating two noiser components, to adeptly increase the robustness against attacks, including MPEG compression. Our solution is aimed at real-life applications; hence we focus on high-resolution videos and intend the encoded image to be indistinguishable from the cover image. To that end, we perform a subjective evaluation on a group of 72 volunteers as well as calculate objective quality metrics obtaining 0.000241 LPIPS, 1.000 SSIM, and 63.8dB PSNR for the best scenario. The obtained results improve PSNR reported by REVMark (Y. Zhang et al., 2023) by around 30dB and LPIPS by a factor of 100. Furthermore, extensive evaluation on both standard COCO dataset and high-resolution videos underlines the meth
od's high robustness against image distortion attacks, achieving over 0.9 bit accuracy for JPEG (q=90), Dropout (p=0.85) and chroma subsampling (4:2:0).
(More)