FreqTransNet: a Frequency-aware Transformer Network for Robust Image Watermarking

To conquer the limitations of traditional watermarking algorithms in adaptively embedding watermarks and to address the constrained feature representation capability of convolution-based methods, this paper introduces a robust deep learning watermarking model named FreqTransNet. This model effectively integrates convolutional modules, Transformer structures, and frequency-domain transformations to enhance its ability in both capturing multi-scale image features and improving robustness. Specifically, the self-attention mechanism inherent in the Transformer effectively models long-range dependencies and global context, thereby improving the model's representation of complex structures. The integration of frequency-domain transformations also mitigates spatial distortions, further amplifying resistance to interference. The extensive experimental results reveal that FreqTransNet surpasses existing state-of-the-art deep learning watermarking approaches in both visual quality and robustness. Compared to the best-performing baseline, FreqTransNet achieves a notable 3.09 dB enhancement in PSNR (from 47.46 dB to 50.55 dB). Additionally, it consistently maintains extraction accuracy exceeding 97% in the presence of various attack scenarios, highlighting its practical applicability and stability in real-world applications.