Research On Image Digital Watermarking Based On Generative Adversarial Network

With the development of the Internet,the types and modes of transmission of digital products are more diverse.In order to avoid tampering and illegal use of digital products in the dissemination process,digital watermarking technology comes up.Text,images,audio,video,etc,all can be used as the carrier of digital watermark.This article uses image as the research object to embed watermark information in image.The watermark information can be transmitted reliably without affecting the visual effect of the image.In the process of transmission,the watermark can still be completely extracted after distortion or certain degree of attack.Watermarking can be effectively applied to leak tracing,copyright authentication,anti-counterfeiting traceability and so on.The design of traditional digital watermarking algorithm requires a large amount of knowledge,the design is complicated,and can not dynamically adjust the algorithm.In this paper,generative adversarial network is introduced,which has far-reaching research significance.The main research contents of this paper are as follows:1.A digital watermarking model based on Generative Adversarial Networks(GAN)and Nonsubsampled Contourlet Transform(NSCT)is proposed.The watermarking image with GAN is almost the same as the original image,and can be extracted from the watermarking image smoothly.The watermark information is embedded into the lowfrequency component of the carrier image by using the two-level NSCT to improve the robustness of the watermark.The image is converted from RGB space to YUV space,and the watermark is embedded in Y space to improve imperceptibility.The loss function of LSGAN is used to alleviate the problem of gradient disappearance of traditional generative adversarial networks.An attack layer is added to the generator,which is trained with embedded and extracted structures at the same time to simulate various common attacks,so that the watermark can be recovered smoothly after a certain degree of attack.Different data sets are used to train.Detect and analyze the influence of different types of attacks on watermark extraction.Using this model,the process of watermark extraction does not need the participation of the original image,and can be directly extracted from the image containing watermark.2.The Oriented FAST and Rotated BRIEF algorithm(ORB)is used to select feature points and form feature regions.The watermark is embedded in the local area around the feature point.Compared with embedding in the whole image,this method has better invisibility.Using the characteristics of ORB algorithm and NSCT,the watermark can resist rotation,resizing,collage and combination attack.