A Pixel Distribution Complexity Classification Enhanced Convolutional Neural Network Predictor for Reversible Data Hiding

This paper proposes a complexity classification enhanced convolutional neural network (CNN)-based high precision pixel prediction algorithm, aiming to enhance image prediction accuracy of reversible data hiding (RDH). Firstly, a new RDH image division strategy is presented, which can divide the cover image into complex blocks and smooth blocks. By using it, CNN-based predictor can be respectively optimized by using training samples with different features, thus achieving accurate pixel prediction. Additionally, the well-designed CNN predictor combined with residual attention module (RAM) and hybrid dilated convolution module (HDCM) help capture locally important features and maintain efficient capture of global information, thereby improving the performance of prediction network. Experimental results indicate that, compared to the state-of-the-art methods, the proposed method achieves superior prediction accuracy, thus enhancing the performance of RDH scheme well.