Research on Innovation and Application of Digital Image Watermarking Technology Driven by Deep Learning

Abstract

Digital image watermarking technology realizes copyright traceability and content authentication by embedding copyright identification and authentication data in the host image, but traditional methods rely on manual design features, and it is difficult to balance robustness, invisibility and embedding capacity. In this paper, the deep learning watermarking technology in recent years is classified and summarized, and the advantages and disadvantages of representative methods are compared and analyzed from three dimensions: network architecture, training strategy and application scenario. Aiming at the problem that it is difficult to balance the robustness and invisibility of the existing methods, this paper proposes an improved end-to-end watermarking network of U-Net encoder which integrates the SE (Squeeze-and-Exclusion) attention mechanism. By embedding the SE channel attention module in the U-Net hop connection path, the watermark is guided to be embedded in the texture-rich area, and the reconstruction loss, watermark extraction loss and confrontation loss are jointly optimized. The experiment on the COCO dataset shows that the proposed method achieves a peak signal-to-noise ratio (PSNR) of 41.56 dB and a structural similarity (SSIM) of 0.9812, which is 2.61 dB higher than HiDDeN. Under attacks such as JPEG compression (QF=50), the bit error rate (BER) is reduced by about 24% compared to the baseline, achieving a better balance between invisibility and robustness. The ablation experiment further verified the effectiveness of SE module and anti-loss, and provided a new technical path for digital copyright protection.