Neural Networks Based Sorting Order for Reversible Data Hiding in Pixel Prediction Error

Abstract

This paper presents a Reversible Data Hiding method in grayscale digital images using Pixel Prediction Error. The prediction error of the prediction error is modified to store the secret data. The prediction of the image pixels is carried out using interpolation from neighboring pixels. The high spatial correlation of image pixels in natural images lead to prediction errors close to zero. In the second step a further interpolation of prediction errors are applied to get the prediction error of prediction errors. These errors are then modified by using histogram modification procedure to carry the secret data in binary form. A novel artificial neural networks based system is trained to find the optimal sorting order of the prediction errors for embedding. In the existing work, a simple local complexity measure is used as a proxy for pixel prediction errors. However, in the proposed work the neural networks based solution gives a more optimal order to embed pixels. Experimental results and analysis are carried out with a set of eight test images with varying characteristics. It is shown that the proposed pixel sorting method gives better visual performance for the same embedding rate compared against existing procedure. The average Peak Signal to Noise Ratio of the proposed work is 56.1 dB which is better than 54.40 dB given by existing work.