Gradient Based Prediction for High Fidelity Reversible Data Hiding with Pairwise Embedding

This paper investigates the use of gradient based prediction for pairwise reversible data hiding. Pairwise based approaches split the host image into two pixel sets. The pixels in each set are processed based on the information from the other set. Host pixels from the same set are grouped into pairs based on their prediction error. Hidden data is then embedded into each pair. The proposed approach introduces a five set distribution combined with a twelve pixel rhombus shaped prediction context, improving the prediction precision and maintaining the independent prediction required for pixel pairing. This approach shows promising results on images with large textured regions.