Abstract

In this paper, we propose a universal screen-shooting robust image watermarking scheme, which can be used to embed extractable information into on-screen images for copyright protection or additional information acquisition. Specifically, to obtain higher robustness and watermarked image quality, we introduce the channel-attention mechanism into DCT domain and exploit the noise layer to guide the watermarking model, which integrates simulated distortions and screen-shooting distortions. After training, the universal watermark mask can be directly added on arbitrary cover image to produce a watermarked image that effectively resists multiple distortions. Experimental results show that the proposed scheme can achieve better performance than some state-of-the-art (SOTA) schemes in terms of the watermark embedding invisibility and the watermark extraction accuracy.