High-fidelity Reversible Data Hiding Research Based On Multiple Histogram Modification

With the rapid development and wide application of computer network technology and multimedia technology,images are vulnerable to illegal copying and tampering.How to protect the copyright and verify the integrity of images has become a concern of the whole society.For example,in the electrical field,unauthorized third parties can easily use various image processing software to tamper with electrical images,resulting in misjudgment and other problems that will cause incalculable losses to related companies.Reversible Data Hiding(RDH)algorithm can not only extract the secret information from the secret image accurately,but also restore the secret image to the original image without loss.Therefore,with the help of RDH algorithm,copyright information can be embedded into electrical images,which can protect the copyright and verify the integrity of electrical images.In this paper,the gray image is taken as the research object,and the RDH algorithm which can be used for image copyright protection and integrity authentication is further studied.The performance of RDH algorithm based on multi-histogram modification framework is improved from two aspects:improving the predictor and proposing asymmetric multi-histogram modification.The paper mainly obtains the following two research results:(1)An RDH algorithm based on asymmetric multi-histogram modification is proposed.Different from the traditional multi-histogram modification algorithm,which directly classifies all pixels,this paper first introduces a smoothness threshold to exclude pixels located in the texture area from data embedding as much as possible.Then,with the help of 16carefully designed pixel features and combined with FCM,the pixels belonging to the smooth area are grouped into multiple categories.Next,two asymmetric predictors are used to generate two asymmetric PEHs for each type of pixel,and only the short tail part of each asymmetric PEH is modified to reduce the invalid modification of the prediction error.Finally,the IDPSO is used to search for the best embedding point for each PEH while reducing the computational complexity.(2)A multi-histogram modification framework based on improved convolutional neural network predictor(ICNNP)or U~2P is proposed.Different from the traditional predictor,which can only extract the shallow correlation between pixels,this paper combines the powerful feature learning ability of deep learning technology to extract the deep correlation between pixels,and then proposes two predictors with higher prediction performance.On the one hand,we propose an improved convolutional neural network predictor with multiple residual blocks.On the other hand,a number of RSUs are introduced into reversible information hiding for the first time,and a new predictor U~2P is designed by fusing feature maps in different receptive fields.Then,all the pixels are grouped into multiple categories by FCM algorithm,and sharp PEH is constructed for each category by ICNNP or U~2P.Finally,the best embedding point is searched for each PEH by IDPSO algorithm.Aiming at the above two RDH algorithms,this paper tests them on general datasets and electrical image dataset,and compares them with several advanced RDH algorithms.Experimental results show that the proposed two RDH algorithms have higher embedding performance.Finally,this paper summarizes the two RDH algorithms proposed,and gives the development direction of RDH algorithms in the future.