Error Prediction Of High Precision Reversible Data Hiding Based On Convolutional Neural Network

Reversible data hiding is a technique for hiding data in digital media without permanently damaging or distorting the media content.In recent years,deep learning has made great progress.The combination of reversible data hiding and convolutional neural network has improved the global optimization and prediction accuracy.However,with the development of science and technology,higher requirements are put forward for reversible data hiding in military,medical and other fields.Therefore,it is imminent to improve the accuracy of the existing error prediction.This thesis studies the error prediction of high-precision reversible data hiding based on convolutional neural network,and discusses the problem of improving the prediction accuracy of target images.The main content of this article is as follows:Two different error prediction algorithms are proposed based on the study of convolutional neural network error prediction algorithms.For the classical error prediction algorithm that does not make full use of the target pixel correlation,the attention enhancement-based deep convolutional neural network reversible information hiding error prediction algorithm is proposed;for the problem that the loss decreases slowly and the prediction accuracy can be further improved in the convolutional neural network error prediction process,the Inception performance optimization-based multiple convolutional neural network reversible Information hiding error prediction algorithm based on Inception performance optimization is proposed.Both error prediction algorithms proposed in this thesis greatly improve the prediction ability of the target pixels and effectively improve the quality of the predicted images.For the urgent need to improve the prediction accuracy of target pixels,this thesis proposes an attention enhancement based convolutional neural network error prediction algorithm,which obtains multi-sensory domains by convolutional kernels of different sizes after the gray-scale image enters the convolutional neural network,and introduces spatial and channel attention mechanisms after the convolutional neural module to assign weights of different degrees of importance to the space and channels after convolution,respectively,to generate more The spatial and channel attention mechanisms are introduced after the convolutional neural module to assign different weights of importance to the convolved space and channels,respectively,to generate more accurate prediction images.A residual network is introduced between the convolutional modules of different dimensions to fuse the feature information of different stages,so as to enhance the network’s representation of the target information,and to promote the attention-based enhancement network to generate prediction images of higher visual quality by making mean square error loss,denoted as MSE＿loss,between the generated prediction images and the target images.Aiming at the problem that the prediction accuracy and convergence speed of the convolutional neural network error prediction algorithm based on attention enhancement need to be improved.This thesis proposes a multiple convolutional neural network prediction algorithm based on Inception structure optimization.Use multiple parallel convolutional networks to increase the breadth of the network,increase the depth of the network through the serial Inception structure,and use ECA networks between different sequences of Inception structures to increase the information expression of important channels.In the Inception network structure and ECA network,the joint training of the residual network is introduced to speed up the convergence speed of the network;the auxiliary predictor is introduced after different Inception structures to improve the prediction accuracy.The prediction image generated by different auxiliary predictors and the target image are used as auxiliary mean square error loss,which is recorded as Aux＿loss,and the mean square error loss of the prediction image and target image generated by the backbone network is recorded as Main＿loss.The final loss function is composed of Aux＿loss and Main＿loss.Experiments show that,compared with the classical prediction algorithm and the error prediction algorithm based on convolutional neural network,the model in this thesis can effectively improve the accuracy of the prediction algorithm.