《计算机应用》唯一官方网站 ›› 2022, Vol. 42 ›› Issue (10): 3102-3110.DOI: 10.11772/j.issn.1001-9081.2021050737

• 网络空间安全 • 上一篇    

基于多尺度知识学习的深度鲁棒水印算法

樊缤, 李智, 高健   

  1. 贵州大学 计算机科学与技术学院,贵阳 550025
  • 收稿日期:2021-05-10 修回日期:2021-09-16 接受日期:2021-10-12 发布日期:2021-09-16 出版日期:2022-10-10
  • 通讯作者: 李智
  • 作者简介:第一联系人:樊缤(1996—),男,贵州遵义人,硕士研究生,主要研究方向:水印算法、计算机视觉、医学影像分析
    李智(1977—),女,贵州贵阳人,教授,博士,CCF会员,主要研究方向:信息隐藏、人工智能、医学影像分析、图像与视频分析;1392589@qq.com
    高健(1993—),男,山西大同人,硕士研究生,主要研究方向:信息隐藏、医学影像分析。
  • 基金资助:
    国家自然科学基金资助项目(62062023);贵州省科技计划项目(ZK[2021]?YB314)

Deep robust watermarking algorithm based on multiscale knowledge learning

Bin FAN, Zhi LI, Jian GAO   

  1. College of Computer Science and Technology,Guizhou University,Guiyang Guizhou 550025,China
  • Received:2021-05-10 Revised:2021-09-16 Accepted:2021-10-12 Online:2021-09-16 Published:2022-10-10
  • Contact: Zhi LI
  • About author:FAN Bin, born in 1996, M. S. candidate. His research interests include watermarking algorithm, computer vision, medical image analysis.
    LI Zhi, born in 1977, Ph. D. , professor. Her research interests include information hiding, artificial intelligence, medical image analysis, image and video analysis.
    GAO Jian,born in 1993, M. S. candidate. His research interests include information hiding, medical image analysis.
  • Supported by:
    National Natural Science Foundation of China(62062023);Guizhou Science and Technology Program(ZK [2021]-YB314)

摘要:

针对现有基于深度学习框架的水印算法无法有效保护高维医学图像版权问题,提出一种基于多尺度知识学习的医学图像水印算法用于弥散加权图像的版权保护。首先,提出一个基于多尺度知识学习的水印嵌入网络来嵌入水印,并通过微调的预训练网络提取弥散加权图像的语义、纹理、边缘以及频域信息作为多尺度的知识特征;然后,结合多尺度的知识特征来重构弥散加权图像,并在该过程中冗余地嵌入水印,从而获得视觉上与原始图像高度相似的含水印的弥散加权图像;最后,提出一个基于金字塔特征学习的水印提取网络,并通过在含有水印的弥散加权图像的不同尺度的上下文中学习水印信号的分布相关性来提高算法的鲁棒性。实验结果表明,所提算法重构出的含水印图的平均峰值信噪比(PSNR)达到57.82 dB。由于弥散加权图像在转换为弥散张量图像时需满足一定的弥散性特征,所提算法仅8个像素点的主轴方向偏转角大于5°,且这8个像素点均不在图像的感兴趣区域。此外,该算法所得图像的各项异性(FA)以及平均弥散率(MD)都接近为0,完全满足临床诊断的要求;且面对裁剪强度小于0.7,旋转角度小于15°等常见的噪声攻击,该算法的水印正确率达到95%以上,能有效保护弥散加权图像的版权信息。

关键词: 鲁棒水印, 神经网络, 弥散加权图像, 多尺度特征, 知识学习, 迁移学习

Abstract:

Aiming at the problem that existing watermarking algorithms based on deep learning cannot effectively protect the copyright of high-dimensional medical images, a medical image watermarking algorithm based on multiscale knowledge learning was proposed for the copyright protection of diffusion-weighted images. First, a watermark embedding network based on multiscale knowledge learning was proposed to embed watermarks, and the semantic, texture, edge and frequency domain information of the diffusion-weighted image was extracted by a fine-tuned pre-training network as multiscale knowledge features. Then, the multiscale knowledge features were combined to reconstruct the diffusion-weighted image, and a watermark was embedded during the process redundantly to obtain a watermarked diffusion-weighted image highly similar to the original one visually. Finally, a watermark extraction network based on pyramid feature learning was proposed to improve the robustness of the algorithm by learning the distribution correlation of watermarking signals from different scales of context in the watermarked diffusion-weighted image. Experimental results show that the average Peak Signal-to-Noise Ratio (PSNR) of the reconstructed watermarked images by the proposed algorithm reaches 57.82 dB. Since diffusion-weighted images need to meet certain diffusivity features when converting to diffusion tensor images, the proposed algorithm only has 8 pixel points with the deflection angle of the principal axis direction greater than 5°, and none of these 8 pixel points is in the region of interest of the image. Besides, both of the Fraction Anisotropy (FA) and the Mean Diffusivity (MD) of the image generated by the proposed algorithm are close to 0, which fully meets the requirements of clinical diagnosis. At the same time, facing common noise attacks such as those with cropping strength less than 0.7 and rotation angle less than 15, the proposed algorithm achieves more than 95% watermarking accuracy and can effectively protect the copyright information of diffusion-weighted images.

Key words: robust watermarking, neural network, diffusion-weighted image, multiscale feature, knowledge learning, transfer learning

中图分类号: