DoBMark: A double-branch network for screen-shooting resilient image watermarking

Abstract

The dramatic changes in cross-media information transmission modes, especially screen-shooting, have made traditional robust image watermarking for digital channels less resistant to various physical noises from the real world. To address this problem, we propose a double-branch network for screen-shooting resilient image watermarking. Specifically, two weights conforming to a Gaussian distribution are assigned to the double-branch encoder, and by jointly training, two residual images containing watermark information are produced to generate the fused watermarked image. A dual frame alternate-fusion mechanism and a just noticeable difference (JND) constraint function on residual images are employed to improve the watermark invisibility and the quality of watermarked images. A differentiable distortion network is introduced to enhance the robustness of the end-to-end network. Additionally, the decoder integrates a quality enhancement module (QEM) that can correct distortions and apply de-noising to the distorted images, thereby improving the accuracy of watermark information extraction. Experimental results show that the proposed scheme can achieve high robustness against the screen-shooting process while maintaining a satisfactory watermark capacity and visual quality of the watermarked image.