Enhanced Image Watermarking via Nested U-Net with Residual U-Blocks and Adversarial Training
Chaoen Xiao, Sirui Peng, Lei Zhang, Furong Yu, Ding Ding · 2024
In the digital information era, image watermarking technology provides an effective means of copyright protection and information authentication by embedding watermark information into images. However, existing deep learning-based watermarking methods have shortcomings in extracting edge details and textures from carrier images, resulting in more noticeable watermark embedding positions. Additionally, simply enhancing feature extraction capabilities can lead to larger model sizes and increased training complexity. To address these issues, this paper proposes a watermark steganography method based on an improved U-Net architecture. Specifically, we introduce a nested U-Net structure and Residual U-blocks (RSU) to optimize multi-level feature fusion through residual connections, enhancing the model's ability to learn edge details and textures, thereby improving watermark invisibility and image quality. Additionally, we replace fully connected layers with one- dimensional convolutions and global pooling, and introduce an adversarial training strategy, significantly reducing the number of model parameters, making the model easier to train, and further improving its ability to handle details and enhance watermark imperceptibility. Experimental results show that the proposed method performs well on multiple metrics, including PSNR, SSIM, and LPIPS, effectively learning fine-grained image features and preserving spatial detail information. This leads to significantly improved visual quality of images with embedded watermarks while maintaining a high watermark extraction accuracy. Moreover, the loss function converges faster, resulting in more efficient training.