SEDD: Robust Blind Image Watermarking With Single Encoder And Dual Decoders.

This article presents a novel image watermarking framework called Single-Encoder-Dual-Decoder (SEDD) designed to enhance copyright protection of digital images by embedding imperceptible and robust watermarks. The SEDD architecture employs a flow-based encoder using normalizing flow to deeply fuse watermark and cover image features, and a parallel dual-decoder—comprising convolutional and fully connected branches—to improve watermark extraction and mitigate visual artifacts. Extensive experiments on the COCO dataset demonstrate that SEDD achieves superior robustness against various individual and combined image distortions, including JPEG compression, Gaussian blur, cropping, and social media platform processing, with bit error rates often below 0.1% and high visual quality measured by PSNR and SSIM. Ablation studies confirm the effectiveness of both the flow-based encoder and dual-decoder in improving robustness and imperceptibility. The authors note that while corner artifacts are reduced, they are not entirely eliminated, indicating a direction for future research.