Research On Performance Optimization In Robust Digital Watermarking

In the past years,with the rapid development of Internet and multimedia technology,digital information has been widely spread with the forms of image,video,text,etc.Multimedia data is vulnerable to illegal access and tampering,which leads to seriours information security problems,such as copyright disputes and data leakage.As an effective multimedia data copyright protection method,digital watermarking technology embeds identification information into the multimedia data without affecting the visual effect,and then uses the extracted embedded information to comfirm the copyright ownership.For the most widely used multimedia data carriers,i.e.,image and video,this thesis studies the robust blind watermarking methods with different embedding rules,aiming to optimize its performance on marked image quality,extraction accuracy and real-time performance,and the main research results obtained are as follows:(1)A visually optimized image watermarking method based on the spatial justnoticeable difference(JND)model is proposed,which solves the significant visual distortion in the non-textured area of marked image.Firstly,the spatial JND model is used to constrain the change of each spatial pixel.Meanwhile,the difference image variance is introduced to constrain the change of adjacent pixels to ensure the marked image quality.Finally,an optimization problem is constructed to solve the optimal watermark embedding strength,which is consistent with the human visual properties of the spatial domain.Compared with the classical image watermarking method based on the DCT domain JND model,the SSIM between the marked image and the host image of the proposed method is higher with the same PSNR,and higher visual quality of the marked image is achieved.(2)A robust image watermarking method based on multi-scale features is proposed to improve the visual quality of the marked image as well as the watermark extraction accuracy by exploiting deep learning.Firstly,the watermark information is redundantly embedded into the host image in the encoder.In addition,the Inception-Res Net network with convolution kernels of different sizes is introduced to better fuse the watermark information and host image.Futhermore,to improve the watermark extraction accuracy,the Inception-Res Net is intergrated into the decoder to obtain multi-scale features.Finally,a two-stage training method is adopted to enhance the robustness while ensuring the marked image quality.Experiments on the COCO dataset prove that the proposed method improves the robustness against attacks,and the watermark extraction accuracy is improved by 6.4% on average compared to the classical Hi DDe N method.(3)An efficient video watermarking method based on frame difference is proposed,which solves the problem that the current video watermarking methods cannot resist geometric attacks with low real-time performance.Firstly,one bit data is embedded by modifying each pixel in the U channel of adjacent frames.Furthermore,the spatial JND model is introduced to restrict the modification to ensure the marked video quality.The watermark is extracted blindly by calculating the frame difference without geometric calibration in the proposed method,and the computation complexity is reduced as well.Experiments on standard video sequences prove that the proposed method can resist severe geometric attacks.Moreover,compared with the transform domain video watermarking method,for a 1080 p video,the watermark embedding and extraction speed of the proposed method is increased by more than 3 times with the same embedding capacity and PSNR.