Screen-shooting Resilient Watermarking Based on Invertible Neural Network

As an invisible mark embedded in multimedia objects, the original watermark is only utilized for copyright protection and authentication in digital domain, but nowadays it has been applied as an invisible QR code for human-machine information exchange. This kind of physical watermarking will undergo a re-imaging process, which introduces noise and distortion that are different from the digital watermarking. How to effectively improve its invisibility and robustness is a new challenge. Facing the challenges, this paper proposes a novel end-to-end screen-shooting resilient watermarking framework that seamlessly integrates an invertible neural network into an encoder-noise layer-decoder scheme. A two-stage training strategy and some elaborately designed loss functions are used to make the algorithm be stable and robust. Extensive experiments show that the proposed scheme outperforms existing algorithms in terms of robustness and invisibility. In most of the real screen-shooting experiments, the BER of our watermarking scheme is below 1% when its capacity reaches up to 196 bits. Meanwhile, the visual quality of the watermarked image still has a PSNR of 37.65dB and an SSIM of 0.9803.