Reversible Image Watermarking Based on Deep Learning

Reversible image watermarking refers to technology that can restore an image to its original state after extracting the watermark. The scheme based on prediction error expansion (PEE) can achieve greater embedding capacity and less image distortion than other methods, so it has been widely researched in recent years. Prediction results of the predictor used by PEE are still not accurate enough, which limits the development of PEE. In this paper, a reversible watermarking predictor based on a deep neural network is proposed. Compared with other predictors, the prediction error histogram generated by our proposed predictor distributes more sharply. At the same time, because of the better prediction results, the watermarked image is closer to the original image. Experimental results show that the proposed method is effective and superior to the existing methods.