Embedding Guide: Improving Watermarking Robustness and Imperceptibility based on Attention and Edge Information
Baowei Wang, Xinyu Lv, Yufeng Wu, Changyu Dai, Zhengyu Hu, Xingyuan Zhao · 2024
In the past few years, there has been an increasing focus on deep learning-based watermarking techniques. Many existing methods do not impose constraints to guide the embedding of watermarking, which leads to random embedding positions and makes watermarks vulnerable to detection and attack. In this paper, an adaptive robust watermarking technique is proposed as a solution to this issue. The proposed method employs a new embedding-guided end-to-end architecture, introducing the Embedding Guide component that utilizes attention mechanism and edge information to embed the secret message into regions that are visually insensitive and inconspicuous. This component enables adaptive embedding of the secret message in each cover image, resulting in high-quality watermarked images with improved imperceptibility. To enhance robustness, this study integrates the Efficient Channel Attention (ECA) block into both the message preprocessor and decoder, facilitating more effective secret message embedding and extraction. Furthermore, UNet++ is applied to improve performance against combined noise. The experimental findings demonstrate that the suggested algorithm surpasses current approaches.