Deep Image Watermarking with Recover Module

Image watermarking based on deep learning has been proposed in the last few years. Typical framework for image watermarking consists of embedding network, extracting network, and attack stimulating module. Adversarial discriminator is sometimes used to make watermarked images much more similar to cover images. To improve the robustness of CNN-based work against attacks of different type and strength, we proposed a novel model, introducing recover module into the framework to compensate some damaged watermark information during attacks and improve extracting accuracy. For non-differentiable JPEG compression, we propose a new approximation approach based on previous methods. Experimental results show that the proposed model performs better than the state of the art in bit accuracy of message extraction while the image quality does not deteriorate.