Learning based vertex prediction for high capacity reversible data hiding in 3D meshes

초록

Reversible data hiding (RDH) is a prominent information hiding method that enables the lossless embedding of additional data within digital multimedia cover files. RDH guarantees perfect reversibility, enabling the receiver to reconstruct the original cover media following data extraction. RDH in 3D meshes has gained increasing attention due to its widespread applications. A critical challenge in RDH for 3D meshes is to accurately predict vertex positions to minimize distortion during data embedding using prediction error expansion (PEE). In this paper, we propose a novel multilayer perceptron based predictor (MLPP) by dividing the vertices of a 3D mesh model into two sets and using one set (reference set) to predict the other set (embedding set) for data embedding. The 1-ring neighbor vertices of the reference set are arranged into a fixed dimension feature vector to train a lightweight and computationally efficient multilayer perceptron network. The proposed network learns from the local geometric structure of 3D meshes to predict embedding vertices and produces sharp prediction errors histogram centered at zero. Furthermore, the small prediction errors are expanded for data embedding, leading to higher capacity and lower distortion. Experimental results demonstrate that the proposed MLPP attains better performance in terms of prediction accuracy, embedding capacity and embedding distortion.

키워드