Multi-scale adaptive image watermarking algorithm based on cross-attention

With the advancement of digital media in recent years, copyright protection and image information security issues have become increasingly significant. Image watermarking algorithms are crucial for protecting image copyright and information security. Traditional watermarking algorithms fail to consider the problem of imperceptibility and robustness, so this article proposes a Multi-scale Adaptive Watermarking (MSAW) algorithm based on cross-attention. By exploiting the energy concentration characteristic of Discrete Cosine Transform (DCT) coefficients to embed watermarking information in the frequency domain, MSAW effectively balances imperceptibility and robustness. Meanwhile, the proposed algorithm uses a multi-scale U-Net encoder to optimize the traditional U-Net structure, and injects multi-scale watermarking information in the up/down sampling process, which enables the algorithm to better resist geometric attacks such as isometric scaling. Additionally, MSAW introduces a cross-attention adapter to improve image quality. Based on the dynamic interaction between the cover image and the watermarking features, the intensity coefficient matrix is generated to realize pixel-by-pixel adaptive adjustment. The algorithm combines multi-scale feature injection with dynamic intensity adjustment strategies while ensuring the quality of the watermarked image. Moreover, through the adversarial training strategy of noise layer simulation, the MSAW algorithm can stably extract watermarks in complex distortion scenarios, leading to significantly higher robustness to geometric attacks. Experiments indicate that MSAW performs well on the COCO2014, Flickr, and ImageNet datasets: The decoding accuracy exceeds 90% in the face of JPEG compression, Gaussian noise, median filtering, and other attacks. Also, in scenarios such as scaling, MSAW has higher robustness than existing methods such as Universal Deep Hiding (UDH), Adaptor, and Deep Independent Template-based Watermarking (DIWT).