UMANeT: A two-stage interpolation-based reversible data hiding framework with attention-enhanced prediction

Interpolation-based reversible data hiding (RDH) techniques have recently attracted significant attention due to their ability to enhance image resolution while ensuring secure data embedding. However, the effectiveness of these methods heavily depends on the quality of the interpolated cover images. Conventional interpolation techniques, typically based on linear models and limited local pixel contexts, often fail to generate high-quality cover images, thereby compromising the visual quality of the resulting stego images and limiting embedding capacity. To address these limitations, this paper introduces a novel hybrid interpolation framework that combines bicubic interpolation with a deep learning-based predictor to construct a high-fidelity two-stage interpolation mechanism. Central to this framework is a newly proposed predictor, termed UMANeT, which leverages a broader contextual region for improved pixel prediction accuracy. By effectively capturing non-linear and long-range dependencies, UMANeT enhances the overall image quality used for data embedding. Experimental results demonstrate that the proposed method not only achieves superior embedding capacity but also generates cover and stego images of significantly higher visual quality compared to existing interpolation-based RDH techniques.