Abstract:

Diffusion Tensor Imaging (DTI) is a commonly used nuclear magnetic resonance imaging technology. To protect the copyright of diffusion tensor images and ensure the integrity of diffusion tensor data, a two-stage robust reversible watermarking algorithm based on the Riemann metric is proposed. To avoid errors in conventional Euclidean operations in tensor space, the diffusion tensor is mapped from manifold space to logarithmic Euclidean space through logarithmic Euclidean transformation, and the features of the diffusion tensor image are extracted and embedded with watermark information using a deep learning model. Subsequently, combined with the reversible watermarking algorithm, a high-quality diffusion tensor image is restored after embedding the robust watermark. Experimental results show that the algorithm can resist various attacks, such as clipping, rotation, and Gaussian noise attacks. Further, the dispersion tensor error recovered by the algorithm is less than 2×10^-8, and the Peak Signal-to-Noise Ratio (PSNR) is 21% higher than that of the VSTNet algorithm.

Key words: robust watermark, reversible watermark, diffusion tensor image, Riemannian metric, deep learning

摘要：

弥散张量成像(DTI)是一种常用的核磁共振成像技术, 为了对弥散张量图像进行版权保护, 同时保证弥散张量数据的完整性, 提出一种基于黎曼度量的两阶段鲁棒可逆水印算法。为了避免常规欧几里得运算在张量空间中的误差, 通过对数-欧几里得变换, 将弥散张量从流形空间映射到对数-欧几里得空间, 利用深度学习模型提取弥散张量图像的特征并嵌入水印信息。然后, 结合可逆水印算法, 在嵌入鲁棒水印后恢复出高质量的弥散张量图像。实验结果表明, 该算法能够抵御裁剪、旋转、高斯噪声等攻击, 且算法恢复出的弥散张量误差不超过2×10^-8, 峰值信噪比(PSNR)相较于VSTNet算法提高了21%。

关键词: 鲁棒水印, 可逆水印, 弥散张量图像, 黎曼度量, 深度学习