Accurate pixel prediction is crucial for achieving high-quality reversible data hiding (RDH). This paper proposes a hybrid prediction-based RDH framework that combines adaptive polynomial modeling with deep residual learning to enhance prediction accuracy while ensuring strict reversibility. The method begins with 2 \(\times \) 2 sub-sampling, followed by reconstruction using Lagrange and bicubic interpolation to generate multiple prediction candidates, which are fused block-wise to form a refined reference image. An adaptive polynomial oriented predictor (APOP) is then employed to further improve local prediction performance. To effectively handle complex textures that are difficult to model using polynomial predictors alone, a lightweight DnCNN-based residual network is trained to learn the remaining prediction errors. The outputs of the APOP and residual predictors are combined using a soft fusion strategy to generate the final prediction map. Based on this enhanced prediction, data embedding is performed via a threshold-adaptive prediction-error expansion scheme. Experimental results on standard grayscale images demonstrate a favorable rate–distortion trade-off over a wide range of embedding capacities, while guaranteeing perfect data extraction and exact recovery of the original image (BER = 0).
