Mixed Order Attention Watermark Network Against JPEG Compression
Zhengze Li, Zhang Xingshuai, Yi Yang, Gong Xing · 2024
Most images uploaded to the Internet will be automatically compressed by the system, resulting in the destruction of the watermark that is employed to protect rights and interests, which poses a new challenge to the watermark extraction work. In this paper, we propose an end-to-end mixed order attention network (MOANet) that enables watermark against Jpeg compression attacks. The network consists of three parts: watermark embedding, Jpeg compression, and watermark extraction. In comparison with prevailing methods, we incorporate a mixed-order attention mechanism during the phases of watermark embedding and extraction. Such an approach is intended to augment the effectiveness of watermark image generation and watermark content retrieval, thus refining the information features of the intermediate network layer. Furthermore, we suggest a Jpeg compression module for MOANet that carefully integrates quantization stages into the lossy compression process with convolutional layers and residual connections, ensuring the network's capability for end-to-end training. The comparative experiments conducted with varying quality factors demonstrate that MOANet outperforms several mainstream watermarking methods in terms of watermark invisibility and robustness against JPEG compression. Notably, when the quality factor is set to 90, MOANet achieves an SSIM value of 0.9969 and a PSNR value of 44.99 dB, with a watermark recovery error rate of merely 1.3%. These results highlight the superior performance of MOANet under conditions that challenge watermark integrity and visibility.