A Prediction Error Order Scheme for Reversible Data Hiding in Image

Prediction Error Ordering (PEO) can integrate the encoding gains of both prediction and sorting by replacing the PVO (Pixel Value Ordering) pixel blocks with prediction-error blocks before sorting. It remains an open topic to construct a consistent, high-accuracy, and block-based prediction method for both the cover and stego images, which are the images before and after embedding the secret data. This paper proposes a novel scheme that explicitly exploits image self-similarity through a deterministic, non-linear prediction framework. Our scheme constructs four independent but self-similar images using a down-sampling and a bicubic-based up-sampling algorithm. Then, a linear ridge regression model learns the self-similarity among pixels with the same coordinates across the four dependent images. It also provides two new 2D mappings to reduce embedding disturbances further. Experimental results demonstrate that our PEO scheme achieves average PSNRs of 59.7 and 58.6 dB at 0.04 bpp, 56.7 and 56.0 dB at 0.08 bpp on the Kodak and USC-SIPI image datasets, respectively.