A novel robust digital image watermarking scheme based on attention U-Net++ structure

With the advancement of the internet, digital image watermarking techniques have found widespread application across various domains, including copyright protection and information security. However, traditional digital image watermarking techniques are susceptible to geometric distortions due to their limited feature extraction capabilities and reliance on manually designed watermark embedding algorithms. Recently, deep neural network-based digital watermarking has emerged as a promising approach due to its powerful nonlinear fitting ability, which has high robustness against various distortions, especially against geometric distortions. Most existing deep neural network-based digital watermarking frameworks employ U-Net style encoders, which may inadequately extract image features and exploit the correlation between secret messages and image pixels. Consequently, this results in a sub-optimal balance between visual quality and robustness. To overcome these limitations, a novel encoder called Attention U-Net++ that merges the advantages of U-Net++ and Attention U-Net is proposed. By incorporating the attention mechanism into the U-Net++ architecture, our proposed encoder effectively extracts image features and finds optimal pixel space for embedding messages, enhancing visual quality and robustness. Furthermore, a quadratic nonlinear growth loss weight setting based on the WGAN style discriminator is devised to enhance performance. Experimental results demonstrate that our proposed method achieves superior visual quality and robustness compared to the state-of-the-art schemes.