Abstract: The watermark image generated by the current watermarking algorithm exhibits poor robustness to non-differentiable noise processing and inadequate general image quality. This study introduces an improved model for a robust image watermarking based on generative adversarial networks. The model integrates the spatial and channel attention mechanism blocks into an encoder with dense structure type to enhance feature extraction ability. A high-pass filter is added to the front of the discriminator to enhance the imperceptibility of the generated watermark image. For end-to-end training, mini-batch of real and simulated (MBRS) JPEG processing is performed in the noise layer, adding various types of noise of different strengths. The experimental results show that compared with HiDDeN and two-stage deep learning robust watermarking model, the watermarked images generated by this model have better robustness and imperceptibility. The average peak signal-to-noise ratio (PSNR) increased by 1.20 dB; average structural similarity increased by 4.71%; and average information extraction bit error rate was reduced by 13.55%.

Key words: generative adversarial networks, robust watermarking, attention mechanism, feature extraction, imperceptibility