High-Capacity Robust Watermarking Technology for High-Resolution Images

As large models and AIGC content become more prevalent, the need for robust copyright protection and traceability has grown. However, existing deep learning-based watermarking techniques are primarily optimized for low-resolution images and small, fixed-length watermarks (e.g., 30–256 bits). This paper addresses the challenge of embedding high-capacity watermarks (up to 4 KB) into high-resolution images (1024x1024) without sacrificing visual quality or robustness.

The researchers introduce a block-wise strategy that partitions high-resolution images into smaller, manageable local blocks. By matching each block with a segment of the high-capacity watermark, the problem is transformed from a single large-scale embedding task into multiple, independent low-capacity embedding tasks. This design significantly reduces the computational burden, allowing the network to train effectively even under resource-constrained conditions.

The framework utilizes a reversible, symmetric architecture consisting of three convolutional layers for the encoder and three deconvolutional layers for the decoder. To ensure robustness, a noise layer is inserted between the encoder and decoder to simulate various attacks, such as JPEG compression and Gaussian noise. The training process employs a multi-component loss function that balances global and local image quality constraints with watermark extraction accuracy.

Experimental results demonstrate that the proposed method successfully embeds 4 KB of data into 1024x1024 images, achieving an embedding rate of 0.0313 bpp. The model maintains high visual imperceptibility while exhibiting strong robustness against various noise attacks. By optimizing both local and global losses, the framework ensures that the watermark remains recoverable even after the image undergoes common distortions.