Efficient CNN Prediction With Smoothness Factor for Reversible Data Hiding

Abstract

In reversible data hiding (RDH) community, researchers often train the CNN-based predictors with the Mean Square Error (MSE) loss function to evaluate the differences between original and predicted images. This will make the prediction network parameters optimized for all pixels without difference. Considering that the prediction errors in smooth areas are prioritized from the prediction error set for reversible data hiding, in this letter we propose to apply a smoothness factor into the MSE loss function. The smoothness factor used to evaluate the pixel smoothness of an image in steganography is adopted as the loss weight in the new loss function, corresponding to large values in the smooth areas and small values in the texture areas. Experimental results have shown that the CNN-based predictors trained with the proposed loss function can predict pixels more accurately in the smooth areas than using the original loss function. As a bonus, better embedding performance can be achieved by comparing with recent typical CNN-based RDH methods.