Frequency-domain attention-guided adaptive robust watermarking model

Abstract

Deep learning-based watermarking models usually take on shortcomings in visual fidelity and robustness. To address these limitations, a novel frequency-domain attention-guided adaptive robust watermarking model is explored. Frequency-domain transform and channel attention mechanism are integrated by the model, it dynamically adapts the watermark embedding process based on content features to ensure adaptability and robustness to different media types. To enhance the representation of image features, an information fusion module is designed to comprehensively capture both deep and shallow features of cover images for fusion with watermark. Additionally, the multi-scale frequency-domain attention module is deployed to generate an attention mask to guide the embedding of watermark, and the weight allocation for different frequencies are optimized during the watermark embedding. The robust feature learning is enhanced during the training by a noise layer. Furthermore, an information extraction module is devised to recover watermarks from the attacked encoded images. The experimental results indicate that the PSNR and the SSIM of the encoded image are above 44.65 dB and 0.9934 respectively. Meanwhile, the proposed model has strong robustness against JPEG attack, which achieves a bit accuracy >98.43 % for extracted messages with compression quality factor of 50. Besides, the proposed model shows strong robustness to many other distortions such as Gaussian noise, resizing, cropping, dropout and Salt & Pepper noise.