Research On High Robust Digital Watermarking Algorithm With Image Structure Preservation

With the development of Internet technology and the popularization of electronic visual devices,more and more digital image content can be efficiently transmitted and shared through the Internet platform,which not only provides great convenience to users but also brings security issues to digital images.How to realize copyright protection,digital authentication and data tracking of the digital image has received more and more attention.As an effective tool for protecting digital image copyright,the digital watermarking algorithm is widely used in the traceability of image files and has become a research hotspot in the field of cyberspace security.Digital watermarking technology is an information hiding technology that uses copyright information or source information as a watermark and embeds it in a digital image carrier.Usually,it can carry identification information in a way that is invisible to human vision.Although digital images may be subject to inevitable signal processing attacks,people can still extract watermark information from digital image carriers for digital copyright attribution,authentication,and traceability.The quality of the watermarked image,the embedding capacity of the watermark information and the robustness of the algorithm are three important indicators to measure the digital watermarking technology,there are intractable checks and balances between various metrics,and how to balanced and steadily improve the performance of all aspects of the algorithm has always been a technical issue in the development of watermarking technology.This thesis mainly studies the high robustness of digital image watermarking algorithms,including the traditional difference modulation algorithm and the watermarking algorithm based on a deep learning framework.Combined with the structural characteristics of the image,two suitable improvements are devised,which effectively improve the robustness of digital image watermarking algorithms.The main contents are as follows:1.This thesis proposes a highly robust watermarking algorithm guided by image structure masking characteristics:First,the strong correlations inside and outside the block in the Discrete Cosine Transform(DCT)domain of the image are adopted to define the horizontal,vertical and diagonal directions as the inter-block comparison direction,and the defined coefficient vectors that are in the same direction among the adjacent direction block are projected to get the projection differences as the watermarking residual of the watermark.Then,eight-interval modulations are set up according to the distribution range of the difference,and the Just Noticeable Difference(JND)visual model is applied to construct the interval threshold perceptual offset,which effectively utilizes the visual redundancy and further improves the robustness of the algorithm.Finally,the algorithm designs two embedding rules.The watermark information embedding is realized by moving the projection difference of the adjacent block vectors to a specific interval.Meanwhile,the adaptive matching adjustment of the DCT coefficients is realized through the JND parameter,and the embedding of the watermark information is completed.The robustness of the proposed method against common signal processing attacks is validated via experiments,including ablation experiments on the proposed multi-directional correlation comparison method,multi-coefficients spread method and JND modulation method,which verifies the effectiveness of each proposed improved strategy.The comparisons between the proposed method and the traditional watermarking algorithm demonstrate that the proposed algorithm is more robust.2.This thesis proposes a highly robust watermarking algorithm based on the framework of image structure feature learning:In order to further improve the robustness of the algorithm to joint processing attacks and preserve the structural details of the watermarked image,a robust watermarking algorithm based on image structural feature learning is proposed.The algorithm uses Convolutional Neural Networks(CNN)to train the encoder and decoder in two stages.The first stage is a noise-free endto-end training stage,using a Generative Adversarial Network(GAN)improves encoder performance.In the above training process,in order to effectively preserve the structural details of the watermarked image,the input image is divided into different image channel,and wavelet transform is performed on the luminance channel to obtain the high and low-frequency bands,design loss functions for different channels and frequency bands of the image.In the second training phase,the encoder parameters are fixed and a noise layer is added to train the decoder.In order to make the algorithm have sufficient robustness in online network transmission,the noise layer is designed according to the common joint attack of the online network platform.Experimental results show that the proposed algorithm is highly robust against common image processing operations and joint attacks.