A Practical Distortion Simulator for Screen-Shooting Resilient Image Watermarking in Consumer Electronic Applications

Nowadays screen-shooting resilient watermarking still remains as a predictive and challenging area of research for proactive data protection in consumer electronic applications. Present deep learning-based methodologies embrace end-to-end frameworks and devise specialized noise layer to simulate the distortions introduced in cross-media transmission between consumer electronics. Typically, the noise layer is designed as a deterministic image-to-image network to simulate the transformation from the clean images to the screen-captured ones, overlooking the randomness of real-world distortions, which makes it challenging to satisfy the demands of various consumer electronics and application scenarios. To address this issue, we conceptualize the screen-shooting channel as an image degradation model and accordingly develop the Resolver-Simulator framework (ReSim). The involved screen-shooting simulator is designed as a conditional image-to-image network to learn the exact degradation function. By taking advantage of the pre-trained parameter resolver, the noise components can be disentangled to compose the instance sets. Then the set sampling strategy is adopted to obtain the noise instances for realistic screen-shooting simulation for unseen images. Experimental results demonstrate that the proposed scheme outperforms the previous arts in terms of real-world robustness as well as exhibits high computational efficiency to enable real-time security solutions in consumer devices.