【作者】 王锋；

【导师】 俞能海；

【作者基本信息】 中国科学技术大学 ， 信息与通信工程， 2022， 硕士

【摘要】 随着新一轮工业革命的到来,3D(Three Dimensions)产业已经成为社会的重要产业。作为3D技术的核心数据形式,3D网格模型具有巨大的商业价值。当这一类模型被商业间谍窃取并非法复制,就会对模型拥有者的版权造成侵犯。因此,针对3D网格模型的版权保护逐渐成为3D技术相关研究中的关键科学问题。数字水印技术,是信息隐藏领域的一大分支。它将一些重要标识以某种方式嵌入到数字载体中,达到防伪溯源、版权保护的效果。3D网格模型的版权保护问题可以通过3D网格水印技术来解决。近年来,3D网格水印技术的相关研究层出不穷,主要分为基于空间域的方法和基于变换域的方法。然而,现有方法普遍存在鲁棒性不足、算法设计的成本较高等局限性,无法作为有效的解决方案。为了解决这一问题,本文提出采用深度神经网络来设计3D网格水印算法。3D网格数据主要包含顶点位置和拓扑连接信息。本文首先研究在特定拓扑连接的3D网格上设计水印网络。进一步地,本文研究了面向多样拓扑连接的3D网格的水印网络设计方法。主要研究工作归纳如下:1.提出了面向特定拓扑连接的3D网格的水印网络设计方法特定拓扑连接的3D网格都是基于同一个拓扑模板,并在模板上进行各种变形而形成不同的形状,其广泛用于人体建模、形态渲染等应用。对于这一类数据,本文提出针对模板网格构建出多尺度的网格分层结构,从而实现对3D网格不同尺度下的特征提取;以基于切比雪夫多项式的谱卷积来构建水印嵌入网络和水印提取网络;基于对抗训练的思想设计了攻击子网络,在训练时使其与主网络进行对抗,从而提升网络对未知噪声的鲁棒性。实验结果表明,本文提出的特定拓扑连接的3D网格水印网络性能优于传统的3D网格水印方法。2.提出了面向多样拓扑连接的3D网格的水印网络设计方法相较于特定拓扑连接的3D网格,多样拓扑连接的3D网格有着更普遍的应用场景。而因为拓扑结构的多样性,针对这一类网格的水印网络设计也具有更大的挑战性。为此,本文提出采用各向同性的空域图卷积网络来搭建主体网络,保证网络参数与网格参数相对独立,从而使得网络可以适用于不同拓扑结构的3D网格;在训练时,引入自适应的攻击层,将应用场景中可能面对的攻击进行整合,实现鲁棒性增强;设计了基于网格曲率一致性的损失函数,从而保证了嵌水印网格的视觉质量。实验结果表明,相比于传统的网格水印算法,本文方法在维持一定的视觉质量条件下,在鲁棒性方面具有明显优势。更多还原

【Abstract】 With the advent of the new industrial revolution,the 3D(Three Dimensions)industry has become an important industry for society.As the core form of data for the 3D techniques,3D mesh models are of great commercial value.When such models are stolen and illegally copied by commercial spies,the copyright of the model owner is infringed.As a result,copyright protection for 3D mesh models is becoming a key scientific issue in related researches to 3D techniques.Digital watermarking is a major branch of the information hiding field.It embeds some important logos into the digital carrier in some way to achieve anti-counterfeit traceability and copyright protection.The copyright protection of 3D mesh models can be solved by using the 3D mesh watermarking techniques.In recent years,there have been numerous researches related to 3D mesh watermarking techniques,which are mainly divided into spatial domain-based methods and transform domain-based methods.However,the existing methods generally have limitations such as insufficient robustness and high costs of algorithm design,and cannot be used as effective solutions.In order to solve this problem,deep neural network is used to design 3D mesh watermarking algorithm in this dissertation.The 3D mesh data mainly contains vertex positions and topology connection information.This dissertation first investigates the design of watermarking network for 3D meshes with specific topological connections.Further,the design method of watermarking network for 3D meshes with diverse topological connections is also investigated.The main research work is summarised as follows:1.The watermarking network design method for 3D meshes with specific topological connections is proposed3D meshes with specific topological connections are all based on the same topological template,and various deformations are performed on the template to form different shapes,which are widely used in applications such as human modelling and morphological rendering.For this type of data,this dissertation proposes to construct a multi-scale mesh hierarchy structure for the template mesh,so as to achieve feature extraction for 3D meshes at different scales;construct the watermark embedding network and watermark extraction network with the Chebyshev polynomial-based spectral convolution;the attack sub-network is designed based on the idea of adversarial training,and is made to fight against the main network during training,thus improving the robustness of the network against agnostic noises.The experimental results show that the proposed watermarking network for 3D meshes with specific topological connections outperforms the traditional 3D mesh watermarking methods.2.The watermarking network design method for 3D meshes with diverse topological connections is proposedCompared with 3D meshes with specific topological connections,3D meshes with diverse topological connections have more general application scenarios.Because of the diversity of the topologies,the design of the watermarking network for this type of meshes is also more challenging.To this end,this dissertation proposes to employ an isotropic spatial convolutional network to form the backbone network,ensuring that the network parameters are independent of the mesh parameters,thus making the network applicable to 3D meshes with various topologies;during training,an adaptive attack layer is introduced to integrate the possible attacks faced in the application scenario to enhance the robustness;a loss function based on the the mesh curvature consistency is designed,thus ensuring the visual quality of the watermarked mesh.The experimental results show that,compared with traditional 3D mesh watermarking algorithms,the proposed method has obvious advantages in terms of robustness while maintaining a certain visual quality.更多还原