Watermark information processing method and device based on deep learning, equipment, storage medium and computer program product

Through the deep learning-based watermark information processing method, comprehensively considering the geometric and texture parts of the three-dimensional model, the end-to-end neural network framework is designed, which solves the problem of watermark extraction failure caused by the missing texture or geometric parts, and realizes the copyright traceability and tracing of the three-dimensional model. Improved robustness.

CN118898535APending Publication Date: 2024-11-05CHINA MOBILE COMM LTD RES INST +1
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410903361.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-05
Publication Date
2024-11-05

AI Technical Summary

Technical Problem

When the textured or geometric parts of the three-dimensional model are missing or deleted, the watermark information is failed and copyright traceability cannot be achieved.

Method used

Using deep learning-based watermark information processing method, we design an end-to-end neural network framework by comprehensively considering the geometric parts and texture parts of the three-dimensional model, simulate various attacks, optimize watermark embedding and extraction, and ensure that even if the texture parts or geometric parts are missing or If deleted, complete target watermark information can also be extracted from the corresponding part.

Benefits of technology

It improves the success rate of watermark information extraction, realizes copyright tracing of three-dimensional models, enhances the robustness of watermarks to various attack forms, and ensures the copyright protection of digital human assets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118898535A_ABST
    Figure CN118898535A_ABST
Patent Text Reader

Abstract

The invention discloses a watermark information processing method and device based on deep learning, equipment, a storage medium and a computer program product. The method comprises the following steps: acquiring an original three-dimensional model and embedded watermark information; inputting the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system to obtain a three-dimensional model embedded with a watermark; the three-dimensional model embedded with the watermark represents that the geometric part and the texture part of the original three-dimensional model are embedded into the three-dimensional model embedded with the watermark information; extracting target watermark information from the attacked three-dimensional model embedded with the watermark through the watermark processing system; the target watermark information is obtained based on a geometric part and / or a texture part in the attacked three-dimensional model embedded with the watermark, and is consistent with the embedded watermark information.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of information security technology, and in particular to a watermark information processing method, device, equipment, storage medium and computer program product based on deep learning. Background Art

[0002] Digital watermarking technology is an important research direction of information hiding technology, and it is considered to be an effective solution for protecting the copyright of 3D models. Among them, digital watermarking technology is to embed the copyright-related information of 3D models into the carrier data without affecting the use value of the carrier data (such as multimedia information, documents, etc.), which is used for copyright identification and usage tracking in practical applications.

[0003] However, the digital watermark technology in the related art will lead to failure in extracting the watermark information if any part of the texture part or the geometric part of the three-dimensional model is missing or deleted manually. It can be seen that since it is difficult to extract the complete watermark information, the purpose of copyright tracing cannot be achieved. Summary of the invention

[0004] In order to solve the technical problems existing in the related technologies, the embodiments of the present application provide a watermark information processing method, device, equipment, storage medium and computer program product based on deep learning.

[0005] To achieve the above purpose, the technical solution of the embodiment of the present application is implemented as follows:

[0006] In a first aspect, an embodiment of the present application provides a watermark information processing method based on deep learning, the method comprising:

[0007] Obtaining the original 3D model and embedded watermark information;

[0008] Inputting the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system to obtain a three-dimensional model embedded with a watermark; the three-dimensional model embedded with a watermark represents a three-dimensional model in which both the geometric part and the texture part of the original three-dimensional model are embedded with the embedded watermark information;

[0009] The target watermark information is extracted from the attacked three-dimensional model with embedded watermark by the watermark processing system; the target watermark information is obtained based on the geometric part and / or texture part of the attacked three-dimensional model with embedded watermark and is consistent with the embedded watermark information.

[0010] In a second aspect, an embodiment of the present application further provides a watermark information processing device based on deep learning, the device comprising:

[0011] An acquisition unit, used for acquiring the original three-dimensional model and embedding watermark information;

[0012] A first processing unit is configured to input the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system to obtain a three-dimensional model embedded with a watermark; the three-dimensional model embedded with a watermark represents a three-dimensional model in which both the geometric part and the texture part of the original three-dimensional model are embedded with the embedded watermark information;

[0013] The second processing unit is used to extract target watermark information from the attacked three-dimensional model with embedded watermark through the watermark processing system; the target watermark information is obtained based on the geometric part and / or texture part in the attacked three-dimensional model with embedded watermark, and is consistent with the embedded watermark information.

[0014] In a third aspect, an embodiment of the present application further provides a watermark information processing device based on deep learning, comprising: a processor and a memory for storing a computer program that can be run on the processor;

[0015] In which, when the processor is used to run the computer program, it executes the steps of the watermark information processing method based on deep learning described in the embodiment of the present application.

[0016] In a fourth aspect, an embodiment of the present application further provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the watermark information processing method based on deep learning described in the embodiment of the present application are implemented.

[0017] In a fifth aspect, an embodiment of the present application further provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the watermark information processing method based on deep learning described in the embodiment of the present application.

[0018] The deep learning-based watermark information processing method, apparatus, device, storage medium and computer program product provided in the embodiments of the present application obtain an original three-dimensional model and embedded watermark information; input the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system to obtain a three-dimensional model embedded with a watermark; the three-dimensional model embedded with a watermark represents a three-dimensional model in which both the geometric part and the texture part of the original three-dimensional model are embedded with the embedded watermark information; the target watermark information is extracted from the attacked three-dimensional model embedded with a watermark through the watermark processing system; the target watermark information is obtained based on the geometric part and / or texture part in the attacked three-dimensional model embedded with a watermark, and is consistent with the embedded watermark information. By adopting the solution of the embodiment of the present application, the geometric part and texture part of the three-dimensional model are comprehensively considered in the watermark extraction stage. In this way, the target watermark information consistent with the embedded watermark information can be extracted based on the geometric part and / or texture part of the attacked three-dimensional model with embedded watermark. It can be seen that even if any part of the texture part or the geometric part of the three-dimensional model is missing or deleted artificially, the complete target watermark information can be extracted from the corresponding complete part, thereby improving the success rate of extracting the target watermark information, thereby effectively realizing the copyright traceability of the three-dimensional model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The process diagram of the watermark information processing method based on deep learning in the embodiment of the present application is as follows Figure 1 ;

[0020] Figure 2 A schematic diagram of a process of generating a three-dimensional model embedded with a watermark according to an embodiment of the present application;

[0021] Figure 3 A schematic diagram of a process for generating target watermark information according to an embodiment of the present application;

[0022] Figure 4 The process diagram of the watermark information processing method based on deep learning in the embodiment of the present application is as follows Figure 2 ;

[0023] Figure 5 This is a schematic diagram of the architecture of a watermark processing system based on deep learning according to an embodiment of the present application;

[0024] Figure 6 A schematic diagram of the composition structure of a watermark information processing device based on deep learning according to an embodiment of the present application;

[0025] Figure 7 This is a schematic diagram of the hardware composition structure of a watermark information processing device based on deep learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] The present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0028] At present, with the rapid development of virtual reality (VR) technology, augmented reality (AR) technology, metaverse, digital twins and other fields, users' immersive visual experience needs are constantly increasing, thus promoting the mass production and circulation of immersive media data represented by digital humans. In practical applications, since the design and production of three-dimensional (3D) digital human assets in animation, games and other fields is a labor-intensive process, how to protect the copyright of 3D digital human assets has become an urgent problem to be solved in the industry.

[0029] In order to protect the copyright of 3D digital human assets, an effective method proposed in the related art is digital watermark technology. This digital watermark technology uses a specific algorithm to covertly embed the copyright-related information of the three-dimensional model into the carrier data without affecting the value and use of the original media data, and can resist various attacks during data transmission and use, so as to extract copyright information as evidence or trace the media data circulation process in the event of a copyright dispute. Since the three-dimensional mesh (3D mesh) is the main data format for the production and circulation of digital human models (i.e., three-dimensional models), most of the digital watermark technologies currently used are digital watermark technologies for three-dimensional models containing three-dimensional meshes (hereinafter referred to as digital watermark technologies). The digital watermark technology in the related art can be divided into two categories: traditional watermark methods and watermark methods based on deep learning.

[0030] Traditional watermarking methods manually design watermark embedding and watermark extraction algorithms in the spatial or frequency domain to resist certain forms of attacks. However, the robustness of spatial watermarking technology to attacks such as adding noise is relatively weak, and it may also cause distortion of the three-dimensional grid. For example, Solution 1 in the related art proposes an algorithm based on the distribution of distances between vertices and mesh centroids, which groups mesh vertices and embeds a watermark bit in each group; for another example, Solution 2 in the related art proposes a three-dimensional model watermark embedding and extraction method based on facet vertex sorting, which determines the facet sorting by calculating geometric properties, and then embeds watermark information in the faces in sequence, which can resist vertex reordering attacks, but its robustness against attacks such as cropping and noise addition will deteriorate, and it is necessary to manually set the threshold; for another example, Solution 3 in the related art proposes frequency domain watermarking technology, which uses Fourier or wavelet analysis to convert the three-dimensional mesh into the frequency domain, and embeds the watermark information into the Fourier / wavelet coefficients. However, the computational complexity of this type of method increases with the number of vertices, and there may be a problem of decreased robustness due to the increase in the number of watermark bits.

[0031] It can be seen that the traditional watermarking methods in the relevant technology are difficult to have universal robustness against various forms of attacks (especially clipping, smoothing, etc.), and manually designing a watermarking algorithm for each specific scenario is very time-consuming and labor-intensive.

[0032] The watermarking method based on deep learning can achieve end-to-end optimization of watermark embedding and watermark decoding neural network by simulating various forms of attacks, which effectively improves the universality of watermarks. For example, the fourth scheme in the related technology proposes to reduce the dimension of the three-dimensional model based on principal component analysis to convert it into a three-channel image, use the encoder to convert the image and watermark information into the latent space for fusion, and then decode it through the decoder to obtain the watermark image, and then restore it to a three-dimensional model; when extracting the watermark, the three-dimensional model embedded with the watermark is also converted into an image and then the watermark information is decoded. However, this method may cause the three-dimensional model after the watermark is embedded to be distorted, and the robustness to noise and smoothing attacks is weak; for example, the fifth scheme in the related technology uses a graph neural network to directly encode the three-dimensional geometric structure, introduces various geometric losses, and generates adversarial losses to ensure the fidelity of the model after the watermark is embedded, and has a more universal robustness to common attacks such as transformation, noise, and cropping.

[0033] However, most of the traditional watermarking methods and deep learning-based watermarking methods in the related art only focus on embedding watermarks into the geometric part of the three-dimensional model, that is, the vertices / facets of the three-dimensional mesh, without considering the texture part of the three-dimensional model (that is, texture maps). In practical applications, texture maps are also an important part of 3D digital human assets, and infringement of 3D digital human assets also includes illegal copying and tampering of their texture maps. To this end, Scheme 6 in the related art proposes a 3D-to-2D (2-Dimension) watermarking technology, which encodes and embeds watermarks into mesh vertices and texture maps respectively, and extracts the watermark from the 2D rendering image of the three-dimensional model through an image decoder in the watermark extraction stage. It can be seen that although this method comprehensively considers the geometry and texture of the three-dimensional model in the watermark embedding stage, if the texture part of the three-dimensional model is directly deleted or completely replaced, it may cause the watermark extraction to fail. In other words, if any part of the texture part or the geometric part of the three-dimensional model is missing or manually deleted, the watermark information extraction will fail. It can be seen that since it is difficult to extract the complete watermark information, the purpose of copyright tracing cannot be achieved.

[0034] Based on this, an embodiment of the present application proposes a watermark information processing method based on deep learning. In various embodiments of the present application, the geometric part and texture part of the three-dimensional model are comprehensively considered in the watermark extraction stage. In this way, the target watermark information consistent with the embedded watermark information can be extracted based on the geometric part and / or texture part of the attacked three-dimensional model with embedded watermark. It can be seen that even if any part of the texture part or the geometric part of the three-dimensional model is missing or deleted artificially, the complete target watermark information can be extracted from the corresponding complete part, thereby improving the success rate of extracting the target watermark information, thereby effectively realizing the copyright traceability of the three-dimensional model.

[0035] The present invention provides a method for processing watermark information based on deep learning. The method is applied to a watermark information processing device based on deep learning. Figure 1 The process diagram of the watermark information processing method based on deep learning in the embodiment of the present application is as follows Figure 1 ;like Figure 1 As shown, the watermark information processing method based on deep learning includes:

[0036] Step 101: Obtain the original 3D model and embedded watermark information.

[0037] In the embodiment of the present application, the original three-dimensional model can be understood as a three-dimensional model to be embedded with a watermark. The original three-dimensional model here is a model for a three-dimensional digital human, and the original three-dimensional model includes data of two parts: a geometric part and a texture part.

[0038] Here, the geometric part can be understood as the 3D mesh of the original 3D model, and the 3D mesh is a collection of multiple small triangles, and of course it can also be a collection of multiple small quadrilaterals or small pentagons. Taking triangles as an example, for the 3D object surface in practical applications, a large number of small triangles are used in the computer to approximate it, that is, the collection of these small triangles constitutes a 3D mesh. Specifically, the 3D mesh can be understood as a collection of a series of 3D mesh vertices and the edges (connection relationships) between them. The texture part can be understood as the texture map of the original 3D model.

[0039] In practical applications, a watermark information processing device based on deep learning can determine the original three-dimensional model from a three-dimensional data file.

[0040] Based on this, in one embodiment, the obtaining of the original three-dimensional model includes: obtaining a three-dimensional data file; and preprocessing the three-dimensional data file to obtain the original three-dimensional model.

[0041] Here, the three-dimensional data file can be an OBJ file or a Python library (PYL) file, etc., which is not limited in the present embodiment. It should be noted that the OBJ file is a three-dimensional model file format, that is, a text file, which can be directly opened with a writing board for viewing and editing.

[0042] In one embodiment, the preprocessing of the three-dimensional data file to obtain the original three-dimensional model includes: determining a three-dimensional grid from the three-dimensional data file, and preprocessing the three-dimensional grid to obtain the original three-dimensional model.

[0043] Here, after acquiring the three-dimensional data file, the watermark information processing device based on deep learning can first read the three-dimensional mesh from the three-dimensional data file, that is, the three-dimensional mesh for the three-dimensional digital human, wherein the number of the three-dimensional meshes can be one or more; then, all non-triangular facets in the three-dimensional meshes are triangulated to obtain a three-dimensional mesh whose facets only include triangles; finally, the orientation and scale of the three-dimensional mesh whose facets only include triangles are normalized to obtain the original three-dimensional model.

[0044] It should be noted that the orientation and scale of the three-dimensional mesh whose face patch only includes triangles are normalized. Specifically, the orientation and scale of the three-dimensional mesh whose face patch only includes triangles are normalized to a unified initial state, wherein the unified initial state corresponding to the orientation and scale of the three-dimensional mesh can be pre-set according to actual needs.

[0045] In the embodiment of the present application, the embedded watermark information can be understood as a digital watermark that needs to be embedded in the original three-dimensional model, wherein the data type of the embedded watermark information is a binary data format, that is, the embedded watermark information is actually a binary watermark sequence, for example, the embedded watermark information can be represented as 0100100. It should be noted that the length of the embedded watermark information can be set according to actual needs.

[0046] Step 102: Input the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system to obtain a three-dimensional model embedded with a watermark.

[0047] In the embodiment of the present application, the watermark-embedded three-dimensional model represents a three-dimensional model in which both the geometric part and the texture part of the original three-dimensional model are embedded with the watermark-embedded information.

[0048] The watermark embedding process of the present application is described in detail below in conjunction with the embodiments.

[0049] Figure 2 FIG. 1 is a flow chart of a process of generating a three-dimensional model embedded with a watermark according to an embodiment of the present application. Figure 2 As shown, the process of generating the three-dimensional model embedded with the watermark includes the following steps:

[0050] Step 1021: Input the geometric part of the original three-dimensional model into the first feature extraction network of the watermark embedding module of the watermark processing system to obtain the features of the geometric part.

[0051] In an embodiment of the present application, the pre-trained watermark processing system is an end-to-end neural network framework, and the pre-trained watermark processing system includes three main modules: a watermark embedding module, a watermark extraction module, and a simulated attack module, wherein the watermark embedding module is used to embed the watermark into the original three-dimensional model to obtain a three-dimensional model embedded with a watermark; the simulated attack module is used to simulate various attacks on the three-dimensional model during actual use and transmission, and perform attack processing on the three-dimensional model embedded with a watermark to obtain a three-dimensional model embedded with a watermark after an attack; the watermark extraction module is used to decode the target watermark information from the three-dimensional model embedded with a watermark after an attack.

[0052] Here, the first feature extraction network of the watermark embedding module refers to the geometric feature extraction network of the watermark embedding module, which is used to extract geometric part features from the original three-dimensional model in the watermark embedding stage. In practical application, the geometric feature extraction network of the watermark embedding module can be implemented by a geometric feature encoder. Specifically, the geometric part of the original three-dimensional model, that is, the three-dimensional mesh of the original three-dimensional model is input into the first feature extraction network of the watermark embedding module of the pre-trained watermark processing system, and the geometric part of the original three-dimensional model is extracted by the first feature extraction network of the watermark embedding module to obtain the corresponding geometric part features.

[0053] Step 1022: Input the texture part of the original three-dimensional model into the second feature extraction network of the watermark embedding module of the watermark processing system to obtain the texture part feature.

[0054] Here, the second feature extraction network of the watermark embedding module refers to the texture feature extraction network of the watermark embedding module, which is used to extract the texture part features from the original three-dimensional model in the watermark embedding stage. In practical application, the texture feature extraction network of the watermark embedding module can be implemented by a texture feature encoder. Specifically, the texture part of the original three-dimensional model, that is, the texture map of the original three-dimensional model is input into the second feature extraction network of the watermark embedding module of the pre-trained watermark processing system, and the texture part of the original three-dimensional model is extracted by the second feature extraction network of the watermark embedding module to obtain the corresponding texture part features.

[0055] Step 1023: Generate a three-dimensional model embedded with a watermark based on the embedded watermark information, the geometric part features and the texture part features.

[0056] In actual application, after obtaining the embedded watermark information, the watermark information processing device based on deep learning may not further process the embedded watermark information, but directly input the embedded watermark information into the reconstruction network.

[0057] Based on this, in one embodiment, the three-dimensional model of the embedded watermark is generated based on the embedded watermark information, the geometric part characteristics and the texture part characteristics, including: inputting the embedded watermark information and the geometric part characteristics into the first reconstruction network of the watermark embedding module to obtain the geometric part of the embedded watermark; inputting the embedded watermark information and the texture part characteristics into the second reconstruction network of the watermark embedding module to obtain the texture part of the embedded watermark; based on the geometric part of the embedded watermark and the texture part of the embedded watermark, the three-dimensional model of the embedded watermark is generated.

[0058] Here, the first reconstruction network of the watermark embedding module refers to the geometric reconstruction network of the watermark embedding module, which is used to reconstruct the geometric part of the embedded watermark from the geometric part features in the watermark embedding stage. In practical applications, the geometric reconstruction network of the watermark embedding module can be implemented by a geometric decoder. Specifically, the geometric part features (i.e., three-dimensional grid features) and the unprocessed embedded watermark information are input into the first reconstruction network of the watermark embedding module, and the geometric part features are reconstructed by the first reconstruction network of the watermark embedding module to obtain the corresponding geometric part of the embedded watermark.

[0059] Here, the second reconstruction network of the watermark embedding module refers to the texture reconstruction network of the watermark embedding module, which is used to reconstruct the texture part of the embedded watermark from the texture part features during the watermark embedding stage. In practical applications, the texture reconstruction network of the watermark embedding module can be implemented by a texture decoder. Specifically, the texture part features (i.e., texture map features) and the unprocessed embedded watermark information are input into the second reconstruction network of the watermark embedding module, and the texture part features are reconstructed by the second reconstruction network of the watermark embedding module to obtain the corresponding texture part of the embedded watermark.

[0060] In actual application, after obtaining the geometric part of the embedded watermark and the texture part of the embedded watermark, the watermark information processing device based on deep learning combines the geometric part of the embedded watermark and the texture part of the embedded watermark to generate a three-dimensional model embedded with the watermark. It can be seen that the embodiment of the present application comprehensively considers the geometric part and the texture part of the three-dimensional model in the watermark embedding stage, that is, the watermark information is embedded in both the geometric part and the texture part of the original three-dimensional model to obtain the three-dimensional model embedded with the watermark, which can improve the protection of the digital watermark for the original three-dimensional model, so that the digital watermark information embedded in the entire original three-dimensional model has better robustness when the model is attacked.

[0061] In actual application, after acquiring the embedded watermark information, the watermark information processing device based on deep learning can further process the embedded watermark information to obtain the implicit features of the watermark, and then generate a three-dimensional model embedded with the watermark based on the implicit features of the watermark, the features of the geometric part and the features of the texture part. In this way, while providing comprehensive and reliable protection for the geometric part and the texture part of the original three-dimensional model, it can also ensure that the watermark embedded in the original three-dimensional model is invisible, thereby improving the security of the embedded watermark information.

[0062] Based on this, in another embodiment, the three-dimensional model of the embedded watermark is generated based on the embedded watermark information, the geometric part features and the texture part features, including: inputting the embedded watermark information into the third feature extraction network of the watermark embedding module to obtain the watermark implicit features; inputting the watermark implicit features and the geometric part features into the first reconstruction network of the watermark embedding module to obtain the geometric part of the embedded watermark; inputting the watermark implicit features and the texture part features into the second reconstruction network of the watermark embedding module to obtain the texture part of the embedded watermark; based on the geometric part of the embedded watermark and the texture part of the embedded watermark, the three-dimensional model of the embedded watermark is generated.

[0063] Here, the third feature extraction network of the watermark embedding module refers to a watermark feature extraction network, which is used to extract watermark implicit features from the input embedded watermark information in the watermark embedding stage. In practical applications, the third feature extraction network of the watermark embedding module can be implemented by a watermark encoder. Specifically, the embedded watermark information is input into the third feature extraction network of the watermark embedding module in the pre-trained watermark processing system, and the embedded watermark information is feature extracted by the third feature extraction network to obtain the corresponding watermark implicit features.

[0064] It should be noted that, in actual applications, since modifying the facets (topology) of the three-dimensional mesh will cause serious defects or distortions to the shape of the three-dimensional mesh, the watermark embedding module in the embodiment of the present application is specifically used to embed the watermark (i.e., embedded watermark information) into the vertex elements (also referred to as vertex information) in the geometric part (i.e., the three-dimensional mesh) of the original three-dimensional model.

[0065] Based on this, in one embodiment, the watermark implicit features and the geometric part features are input into the first reconstruction network of the watermark embedding module to obtain the geometric part embedded in the watermark, including: obtaining the vertex information of the original input; inputting the vertex information, the watermark implicit features and the geometric part features into the first reconstruction network of the watermark embedding module to obtain the geometric part embedded in the watermark; the geometric part embedded in the watermark represents the vertex information after the watermark is embedded.

[0066] Here, it should be noted that before the watermark implicit features are input into the first reconstruction network of the watermark embedding module, the watermark implicit features need to be copied and expanded to ensure that the dimensions of the expanded watermark implicit features are consistent with the input vertex information and geometric part features, and then the expanded watermark implicit features are stacked with the input vertex information and geometric part features in the channel dimension and input into the first reconstruction network of the watermark embedding module together.

[0067] In one embodiment, the step of inputting the watermark implicit features and the texture part features into the second reconstruction network of the watermark embedding module to obtain the texture part embedded in the watermark comprises: acquiring the texture part of the original input; inputting the texture part of the original input, the watermark implicit features and the texture part features into the second reconstruction network of the watermark embedding module to obtain the texture part embedded in the watermark.

[0068] It should be noted that the size of the texture part embedded in the watermark is consistent with the size of the original input texture part. In addition, before the watermark implicit feature is input into the second reconstruction network of the watermark embedding module, the watermark implicit feature needs to be copied and expanded to ensure that the dimensions of the expanded watermark implicit feature and the input original texture part and texture part feature are consistent, and then the expanded watermark implicit feature is stacked with the input original texture part and texture part feature in the channel dimension and input into the second reconstruction network of the watermark embedding module.

[0069] Step 103: extracting target watermark information from the attacked watermark-embedded three-dimensional model through the watermark processing system.

[0070] In an embodiment of the present application, the target watermark information is obtained based on the geometric part and / or texture part of the attacked three-dimensional model with embedded watermark, and is consistent with the embedded watermark information. That is to say, in the watermark extraction stage, in addition to being able to extract the target watermark information from the geometric part and texture part of the attacked three-dimensional model with embedded watermark, the target watermark information can also be extracted from any part of the geometric part or texture part of the attacked three-dimensional model with embedded watermark. In this way, even if the texture part is artificially deleted, the target watermark information can be extracted from the geometric part alone; similarly, even if the geometric part is missing, the target watermark information can be extracted only from the texture part. And for the same three-dimensional model, the target watermark information extracted using the trained watermark processing system should be consistent with the watermark originally embedded in the three-dimensional model (i.e., the embedded watermark information).

[0071] It can be seen that in the embodiments of the present application, even if any part of the texture part or the geometric part of the three-dimensional model is missing or deleted manually, the complete target watermark information can be extracted from the corresponding complete part, thereby improving the success rate of extracting the target watermark information, thereby effectively realizing the copyright traceability of the three-dimensional model.

[0072] The watermark extraction process of the present application is described in detail below in conjunction with the embodiments.

[0073] Figure 3 FIG. 1 is a flow chart of a process for generating target watermark information according to an embodiment of the present application. Figure 3As shown, the process of generating the target watermark information includes the following steps:

[0074] Step 1031: During the training process, one or more attack processes are performed on the watermark-embedded three-dimensional model through the simulated attack module of the watermark processing system to obtain an attacked watermark-embedded three-dimensional model.

[0075] The traditional watermarking method in the related art is to manually design the watermark embedding and extraction algorithm, which can resist a certain form of attack, but it is difficult to have universal robustness against all kinds of attack forms. To solve this technical problem, the embodiment of the present application proposes an end-to-end neural network framework (corresponding to the watermark processing system) including a watermark embedding module, a simulated attack module, and a watermark extraction module. During the training process of the watermark processing system, the simulated attack module is used to simulate various attacks that the three-dimensional model is subjected to during actual use and transmission, so as to perform attack processing on the three-dimensional model embedded with the watermark, and obtain the three-dimensional model embedded with the watermark after the attack, which effectively improves the universal robustness of the watermark against various common attacks.

[0076] Here, one or more attack processing is performed on the watermarked 3D model, including but not limited to the following operations: transformation and scaling, rotation, noise addition, cropping, smoothing, etc. Among them, transformation and scaling can be understood as scaling the watermarked 3D model, for example, scaling the watermarked 3D model with a ratio of s; rotation can be random rotation or fixed rotation at a set angle, for example, rotation at an angle α with (x, y, z) as the center; noise addition can be understood as adding noise to the geometric part and / or texture part of the watermarked 3D model, for example, adding random Gaussian noise to the 3D grid and / or texture map of the watermarked 3D model; cropping can be understood as randomly cropping a certain area in the geometric part of the watermarked 3D model; smoothing can be understood as smoothing the geometric part of the watermarked 3D model.

[0077] Step 1032: Input the attacked watermarked three-dimensional model into the watermark extraction module of the watermark processing system to obtain target watermark information.

[0078] In actual application, in the watermark extraction stage, the watermark information processing device based on deep learning can extract the target watermark information consistent with the embedded watermark information based on the geometric part and / or texture part of the attacked three-dimensional model with embedded watermark.

[0079] Based on this, in one embodiment, the inputting of the attacked watermarked three-dimensional model into the watermark extraction module of the watermark processing system to obtain target watermark information includes: inputting the geometric part of the attacked watermarked three-dimensional model into the first feature extraction network of the watermark extraction module to obtain the attacked geometric part feature; inputting the texture part of the attacked watermarked three-dimensional model into the second feature extraction network of the watermark extraction module to obtain the attacked texture part feature; generating the target watermark information based on the attacked geometric part feature and / or the attacked texture part feature.

[0080] Here, the first feature extraction network of the watermark extraction module refers to the geometric feature extraction network of the watermark extraction module, which is used to extract the attacked geometric part features from the attacked geometric part of the three-dimensional model embedded with the watermark in the watermark extraction stage. In actual application, the geometric feature extraction network of the watermark extraction module can be implemented by a geometric feature encoder.

[0081] Here, the second feature extraction network of the watermark extraction module refers to the texture feature extraction network of the watermark extraction module, which is used to extract the attacked texture part features from the attacked texture part of the watermark-embedded three-dimensional model in the watermark extraction stage. In actual application, the texture feature extraction network of the watermark extraction module can be implemented by a texture feature encoder.

[0082] In actual application, after obtaining the attacked geometric part features and / or the attacked texture part features, the watermark information processing device based on deep learning can obtain the corresponding watermark implicit features through adaptive global average pooling of the above features.

[0083] Based on this, in one embodiment, the target watermark information is generated based on the attacked geometric part features and / or the attacked texture part features, including: performing global average pooling processing on the attacked geometric part features to obtain target geometric watermark implicit features; performing global average pooling processing on the attacked texture part features to obtain target texture watermark implicit features; generating the target watermark information based on the target geometric watermark implicit features and / or the target texture watermark implicit features.

[0084] Here, the global average pooling processing is performed on the attacked geometric part features, which can be understood as the averaging operation of the attacked geometric part features in each pooling window to achieve the dimensionality reduction of the attacked geometric part features and the extraction of key information therein, that is, to obtain the target geometric watermark implicit features. Similarly, the global average pooling processing is performed on the attacked texture part features, which can be understood as the averaging operation of the attacked texture part features in each pooling window to achieve the dimensionality reduction of the attacked texture part features and the extraction of key information therein, that is, to obtain the target texture watermark implicit features.

[0085] Here, in one embodiment, the target watermark information is generated based on the target geometric watermark implicit features and / or the target texture watermark implicit features, including: inputting the target geometric watermark implicit features into the watermark recovery network of the watermark extraction module to obtain the target geometric watermark part; inputting the target texture watermark implicit features into the watermark recovery network of the watermark extraction module to obtain the target texture watermark part; generating the target watermark information based on the target geometric watermark part and / or the target texture watermark part.

[0086] Here, the watermark recovery network of the watermark extraction module is used to recover the target geometric watermark part from the target geometric watermark implicit features, and to recover the target texture watermark part from the target texture watermark implicit features. In practical applications, the watermark recovery network can be implemented by a watermark decoder.

[0087] Here, after obtaining the target watermark information, the watermark information processing device based on deep learning can binarize the target watermark information to obtain a binary watermark sequence corresponding to the target watermark information. Regardless of whether it is a binary watermark sequence extracted from the target geometric watermark part or a binary watermark sequence extracted from the target texture watermark part, it should be consistent with the binary watermark sequence of the embedded watermark information, thereby realizing the functions of three-dimensional model authentication and infringement tracing.

[0088] In actual application, the watermark information processing device based on deep learning needs to pre-train the untrained watermark processing system before inputting the original three-dimensional model and the embedded watermark information into the pre-trained watermark processing system to obtain a trained watermark processing system.

[0089] Based on this, in one embodiment, before inputting the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system, the method further includes: constructing a target loss function; training the watermark processing system based on the target loss function to obtain a trained watermark processing system.

[0090] Here, the trained watermark processing system can be understood as a trained watermark processing system or a pre-trained watermark processing system. The watermark processing system can be built using a deep learning model, such as a deep learning algorithm based on a convolutional neural network (CNN).

[0091] In the embodiment of the present application, the goal of training the watermark processing system is to make the embedded watermark invisible and robust against various attacks. To ensure the invisibility of the embedded watermark, it is necessary to constrain the geometric part, texture part and rendering effect of the three-dimensional model embedded with the watermark to be as consistent as possible with the original three-dimensional model, and to ensure the robustness against various attacks, it is necessary to constrain the target watermark information extracted from the attacked three-dimensional model embedded with the watermark to be consistent with the embedded watermark information. In other words, the target loss function in the embodiment of the present application is determined by the geometric loss, texture loss, rendering loss, and watermark recovery loss.

[0092] Based on this, in one embodiment, constructing the target loss function includes: constructing a first loss function, a second loss function, a third loss function and a fourth loss function; the first loss function represents a geometric loss function, the second loss function represents a texture loss function, the third loss function represents a rendering loss function, and the fourth loss function represents a watermark recovery loss function; based on the first loss function, the second loss function, the third loss function and the fourth loss function, construct the target loss function.

[0093] Here, the first loss function, the second loss function, the third loss function and the fourth loss function can all be calculated by mean square error (MSE). Of course, other loss functions that can calculate the difference between the original input and the watermarked input can also be used. The embodiment of the present application does not limit the construction method of the first loss function, the second loss function, the third loss function and the fourth loss function.

[0094] In one embodiment, the constructing of the first loss function includes: obtaining a geometric loss function based on the difference between the vertex coordinates of the geometric part of the original three-dimensional model and the corresponding vertex coordinates of the geometric part of the three-dimensional model embedded with the watermark. The constructing of the second loss function includes: obtaining a texture loss function based on the difference between the texture part of the original three-dimensional model and the texture part of the three-dimensional model embedded with the watermark. The constructing of the third loss function includes: obtaining a rendering loss function based on the difference between the rendered image of the original three-dimensional model and the rendered image of the three-dimensional model embedded with the watermark. The constructing of the fourth loss function includes: obtaining a watermark recovery loss function based on the difference between the embedded watermark information and the extracted target geometric watermark part and / or target texture watermark part.

[0095] In actual application, in one embodiment, the target loss function is constructed based on the first loss function, the second loss function, the third loss function and the fourth loss function, including: respectively determining the weight parameters corresponding to the normalized first loss function, the second loss function, the third loss function and the fourth loss function; determining the target loss function based on the weighted sum of the first loss function, the second loss function, the third loss function and the fourth loss function, and their respective corresponding weight parameters.

[0096] Here, the value ranges of the weight parameters corresponding to the first loss function, the second loss function, the third loss function and the fourth loss function are respectively 0 to 1, and the sum of the four weight parameters is 1. It should be noted that the weight parameters corresponding to the above loss functions can be set manually.

[0097] In practical applications, when constructing the target loss function, in addition to considering geometric loss, texture loss, rendering loss, and watermark recovery loss, the network regularization loss can also be considered to improve the accuracy of the target loss function.

[0098] Based on this, in one embodiment, the constructing of the target loss function includes: constructing a first loss function, a second loss function, a third loss function, a fourth loss function and a fifth loss function; the first loss function represents a geometric loss function, the second loss function represents a texture loss function, the third loss function represents a rendering loss function, the fourth loss function represents a watermark recovery loss function, and the fifth loss function represents a network regularization loss function; based on the first loss function, the second loss function, the third loss function, the fourth loss function and the fifth loss function, the target loss function is constructed.

[0099] The embodiment of the present application also provides another watermark information processing method based on deep learning, which is applied to a watermark information processing device based on deep learning. Figure 4 The process diagram of the watermark information processing method based on deep learning in the embodiment of the present application is as follows Figure 2 ;like Figure 4 As shown, the watermark information processing method based on deep learning includes:

[0100] Step 401: Acquire the original 3D model and embed watermark information.

[0101] In one embodiment, the obtaining of the original three-dimensional model includes: obtaining a three-dimensional data file; and preprocessing the three-dimensional data file to obtain the original three-dimensional model.

[0102] Step 402: Input the geometric part of the original three-dimensional model into the first feature extraction network of the watermark embedding module of the trained watermark processing system to obtain the geometric part feature, and input the texture part of the original three-dimensional model into the second feature extraction network of the watermark embedding module to obtain the texture part feature.

[0103] In one embodiment, before inputting the geometric part and the texture part of the original three-dimensional model into the trained watermark processing system, the method may also include: constructing a target loss function; training the watermark processing system based on the target loss function to obtain the trained watermark processing system.

[0104] In one embodiment, constructing the target loss function includes: constructing a first loss function, a second loss function, a third loss function and a fourth loss function; the first loss function represents a geometric loss function, the second loss function represents a texture loss function, the third loss function represents a rendering loss function, and the fourth loss function represents a watermark recovery loss function; based on the first loss function, the second loss function, the third loss function and the fourth loss function, constructing the target loss function.

[0105] Step 403: Generate a three-dimensional model embedded with a watermark based on the embedded watermark information, the geometric part features and the texture part features.

[0106] In the embodiment of the present application, the watermark-embedded three-dimensional model represents a three-dimensional model in which both the geometric part and the texture part of the original three-dimensional model are embedded with the watermark-embedded information.

[0107] In one embodiment, the method of generating a three-dimensional model embedded with a watermark based on the embedded watermark information, the geometric part features and the texture part features includes: inputting the embedded watermark information and the geometric part features into a first reconstruction network of the watermark embedding module to obtain the geometric part of the embedded watermark; inputting the embedded watermark information and the texture part features into a second reconstruction network of the watermark embedding module to obtain the texture part of the embedded watermark; and generating the three-dimensional model embedded with the watermark based on the geometric part of the embedded watermark and the texture part of the embedded watermark.

[0108] Step 404: During the training process, one or more attack processes are performed on the watermark-embedded three-dimensional model through the simulated attack module of the watermark processing system to obtain an attacked watermark-embedded three-dimensional model.

[0109] Step 405: Input the attacked geometric part of the watermarked three-dimensional model into the first feature extraction network of the watermark extraction module of the watermark processing system to obtain the attacked geometric part features, and input the attacked texture part of the watermarked three-dimensional model into the second feature extraction network of the watermark extraction module to obtain the attacked texture part features.

[0110] Step 406: Generate target watermark information based on the attacked geometric part features and / or the attacked texture part features.

[0111] In the embodiment of the present application, the target watermark information is obtained based on the geometric part and / or texture part of the attacked three-dimensional model with embedded watermark, and is consistent with the embedded watermark information.

[0112] In one embodiment, the target watermark information is generated based on the attacked geometric part features and / or the attacked texture part features, including: performing global average pooling processing on the attacked geometric part features to obtain target geometric watermark implicit features; performing global average pooling processing on the attacked texture part features to obtain target texture watermark implicit features; generating the target watermark information based on the target geometric watermark implicit features and / or the target texture watermark implicit features.

[0113] In one embodiment, the target watermark information is generated based on the target geometric watermark implicit features and / or the target texture watermark implicit features, including: inputting the target geometric watermark implicit features into the watermark recovery network of the watermark extraction module to obtain the target geometric watermark part; inputting the target texture watermark implicit features into the watermark recovery network of the watermark extraction module to obtain the target texture watermark part; generating the target watermark information based on the target geometric watermark part and / or the target texture watermark part.

[0114] By adopting the technical solution of the embodiment of the present application, the geometric part and texture part of the three-dimensional model are comprehensively considered in the watermark extraction stage. In this way, the target watermark information consistent with the embedded watermark information can be extracted based on the geometric part and / or texture part of the attacked three-dimensional model with embedded watermark. It can be seen that even if any part of the texture part or the geometric part of the three-dimensional model is missing or deleted artificially, the complete target watermark information can be extracted from the corresponding complete part, thereby improving the success rate of extracting the target watermark information, thereby effectively realizing the copyright traceability of the three-dimensional model.

[0115] The present application will be further described in detail below in conjunction with application examples.

[0116] The watermark processing solutions in related technologies have the following technical problems: 1) Traditional methods can resist certain forms of attacks by manually designing watermark embedding and extraction algorithms, but it is difficult to have universal robustness against all forms of attacks. 2) Most traditional and deep learning methods only focus on embedding watermarks in the geometric part of the three-dimensional mesh, and do not consider the texture map of the model, which cannot meet the copyright protection requirements of digital human textures in practical applications. 3) Although the existing geometric-texture joint watermarking technology comprehensively considers the geometric and texture parts of the model in the watermark embedding stage, if the texture of the model is directly deleted or completely replaced, the watermark extraction will fail.

[0117] In response to the above technical problems, the present application proposes a three-dimensional model watermark embedding and extraction method for digital human asset protection. Compared with the methods in the related art, the present application proposes an end-to-end neural network framework (corresponding to the aforementioned deep learning-based watermark processing system) including a watermark embedding module, a simulation attack module, and a watermark extraction module, which effectively improves the universal robustness of the watermark against common attacks such as transformation, rotation, noise, and cropping; watermarks are embedded in the geometric part and texture part of the three-dimensional model respectively, and the target watermark is extracted from the geometric part and / or texture part of the attacked three-dimensional model with embedded watermarks. In this way, even if any part of the texture part or the geometric part of the three-dimensional model is missing or deleted manually, the complete target watermark can be extracted through the other part.

[0118] Specifically, firstly, the geometric part (i.e., 3D mesh) and texture part (i.e., texture map) of the original 3D model are watermarked at the same time, and geometric loss, texture loss, and rendering loss are introduced to ensure the invisibility of the watermark in the 3D model after the watermark is embedded, and the use value of the digital human asset is not affected; then, the 3D model with embedded watermark is subjected to common attacks such as transformation, rotation, noise addition, and cropping through the simulation attack module to improve the universal robustness of the watermark; and in the watermark extraction stage, two feature extraction networks are also used to extract the attacked geometric part features and / or texture part features from the geometric part and / or texture part of the 3D model, and after global average pooling processing, they are respectively sent to the watermark recovery network to obtain two watermark vectors, and it is ensured that the two extracted watermark vectors are consistent with the embedded watermark vector. This means that even if any part of the geometry or texture of the 3D model is artificially deleted, the complete target watermark can be extracted through the other part to achieve the purpose of infringement tracing.

[0119] Figure 5 Schematic diagram of the architecture of the watermark processing system based on deep learning in the embodiment of the present application. Figure 5 As shown, the watermark processing system based on deep learning mainly includes three modules: watermark embedding module, simulated attack module, and watermark extraction module. The following is a detailed description of the watermark information processing method based on deep learning of the present application in conjunction with the architecture diagram.

[0120] 1. Input data of watermark processing system

[0121] Read the 3D mesh of the digital human from the 3D data file (such as OBJ, PYL, etc.), triangulate all non-triangular facets in the mesh, and normalize the orientation and scale of the 3D mesh that only includes triangles to a unified initial state to obtain a 3D model of the digital human (corresponding to the original 3D model mentioned above). Therefore, the 3D model of the digital human can be expressed as follows: in =(V in ,F in ,T in ).

[0122] Among them, the three-dimensional mesh vertex set is N v Represents the number of vertices of the three-dimensional mesh, C v Represents the number of vertex coordinates in a 3D mesh. A 3D mesh vertex set may include the spatial position and normal information of the vertex. N f Indicates the number of patches, C f Indicates the number of vertices that make up each patch, where C f The value can be 3, which can convert the vertex connection relationship represented by it into the adjacency matrix form A that can be processed by the neural network; texture map is a three-channel color image, H represents the length of the texture map, W represents the width of the texture map, C t Indicates the number of channels of the texture map, C t The possible value is 3. The pixels of the texture map have a one-to-one mapping relationship with the vertices on the 3D mesh surface, and its pixel value represents the color information of the 3D mesh surface.

[0123] For each digital human 3D model, a length of N can be assigned w Binary watermark sequence (corresponding to the aforementioned embedded watermark information), used to identify the copyright information of the digital human three-dimensional model.

[0124] 2. Network Structure of Watermark Processing System

[0125] The network structure of the watermark processing system proposed in this application is roughly composed of three parts: a watermark embedding module, a simulated attack module, and a watermark extraction module. Specifically, the original three-dimensional model and the embedded watermark information are input into the watermark embedding module to obtain a three-dimensional model with a watermark (corresponding to the aforementioned three-dimensional model with a watermark embedded). After passing through the simulated attack module, it is input into the watermark extraction module to decode the watermark information (corresponding to the aforementioned target watermark information). The simulated attack module is used to simulate various attacks on the three-dimensional model during actual use and transmission, and to apply various data interferences to the three-dimensional model with a watermark added. Among them, the attack operation may include but is not limited to the following operations: adding random Gaussian noise to the three-dimensional grid and texture map; rotating at an angle α with (x, y, z) as the center; scaling transformation with a ratio of s; smoothing the three-dimensional grid; and randomly cropping a certain area of ​​the three-dimensional grid.

[0126] Here, the learnable sub-networks in the watermark embedding module and the watermark extraction module include:

[0127] 1) Watermark feature extraction network E w and watermark recovery network D w

[0128] E w In the watermark embedding stage, the input binary watermark sequence w in (corresponding to the aforementioned embedded watermark information) is mapped to the watermark implicit feature z = E w (w in ), N z Indicates the length of the implicit feature of the watermark; D w In the watermark extraction stage, the target watermark implicit features are extracted from the geometric and texture parts of the attacked watermarked 3D model: (i.e., implicit features of the target geometric watermark), (i.e., the implicit features of the target texture watermark), and decode the two implicit features of the target watermark accordingly as (i.e. the target geometric watermark part), (ie the target texture watermark part).

[0129] 2) Geometric feature extraction network E for embedding watermarks into 3D meshes g and geometric reconstruction network D g

[0130] Since modifying the facets (topology) of a 3D mesh will cause serious defects or distortions in the shape of the 3D mesh, the watermark information is embedded into the vertex elements of the 3D model. g Extract geometric part features f from the original 3D mesh g =E g (V in ,A), then D g Combine the extracted geometric features (C v ' means E g The output channel dimension of E g The network outputs a length C for each three-dimensional vertex v ', so E g The collective feature encoding of the network output is a dimension of N v ×C v 'The matrix f g ), input vertex information V in , implicit feature z of the watermark to be embedded, synthesize vertex information after adding watermark (corresponding to the geometric part of the embedded watermark mentioned above). It should be noted that z is input to D g Previously, it would replicate and expand to N v ×N z , and f g and V in Concatenate in the channel dimension.

[0131] 3) Texture feature extraction network E for embedding watermarks into texture maps t and texture reconstruction network D t

[0132] With E g and D g Similar to E. t (composed of CNN) learns the texture feature f from the texture map of the original 3D model t =E t (T in ), then D tCombined with the extracted texture features (H' represents the height of the texture feature, W' represents the width of the texture feature, C t ' represents the number of channels of the texture feature (the value can be 3), f t represents a matrix composed of a series of feature maps learned by CNN), the input original texture map T in , implicit feature z of the watermark to be embedded, synthesize the texture map after adding the watermark (corresponding to the aforementioned texture parts of the embedded watermark), its size is the same as T in Keep consistent. It should be noted that z is input to D t Previously, it would be copied and expanded to H'×W'×N z , and f t and T in Stacked in the channel dimension.

[0133] 4) Geometric feature extraction network E used to extract watermarks from 3D meshes in the watermark extraction module g ', used to learn the attacked grid data The feature (corresponding to the aforementioned geometric feature after the attack) is obtained after the adaptive global average pooling (Adaptive Global Average Pooling) to obtain the target geometric watermark implicit feature

[0134] 5) Texture feature extraction network E in the watermark extraction module for extracting watermarks from texture maps t ', used to learn the attacked texture map The feature (corresponding to the aforementioned attacked texture feature) is obtained after adaptive global average pooling to obtain the target texture watermark implicit feature

[0135] Here, the geometric part (3D mesh) of the 3D model is processed using E g , D g and E g ', in order to adapt the algorithm to three-dimensional grids of different sizes (number of vertices) and different network topologies, these three sub-networks (E g , D g and E g ') can be built using a graph attention neural network (GAT). The GAT network is a convolution operation applied to a graph structure that can weightedly aggregate the features of the neighboring points of each vertex. The local spatial information obtained is independent of the vertex order and can intuitively and flexibly learn the implicit features of the 3D mesh.

[0136] For vertex v i , assuming that the feature input to the lth layer of the GAT network is f i l , then the output feature f of the GAT network i l +1 It can be expressed by the following formula (1):

[0137]

[0138] Among them, N(v i ) represents the vertex v i The connected vertex set, W0 represents the update v i The weight of its own characteristics, Represents vertex v j The features input to the lth layer of the GAT network. a ij Represents vertex v i Aggregate its adjacent vertices v j The attention weights of the features are learned by a multi-layer perceptron network (MLP), whose input is the vertex v i and v j The spatial coordinate offset of Among them, a ij It can be expressed by the following formula (2):

[0139] a ij =softmax(σ(Wo ij ))(2)

[0140] Where W and σ are the learnable weights and activation functions of MLP respectively. Each GAT layer is followed by a batch normalization layer (BN) and a ReLU activation function layer to form a network block. g , D g and E g 'That is, it is composed of multiple such network blocks. In order to improve the expressiveness and performance of the network block, residual learning can be introduced. In addition to the first network block, subsequent network blocks can add skip connections. g An additional GAT layer is connected to the end of the output vertex coordinates after embedding the watermark

[0141] The texture of the 3D model is processed using E t and D t 、E t ', mainly based on standard CNN construction, in practical applications, E t and D tThe U-Net structure can be used, wherein the watermark feature extraction network and the watermark recovery network can be constructed using MLP.

[0142] 3. Training and Reasoning Process of Watermark Processing System

[0143] The training goal of the watermark processing system is to make the embedded watermark invisible and robust against various attacks. In order to ensure the invisibility of the embedded watermark, it is necessary to constrain the geometry, texture and rendering effect of the 3D model embedded with the watermark to be as consistent as possible with the original 3D model.

[0144] Geometric loss: Calculate the mean square error between the vertex coordinates of the original 3D mesh and the corresponding vertex coordinates of the 3D mesh embedded with the watermark, which can be expressed by the following formula (3):

[0145]

[0146] Among them, L g represents the geometric loss, v i represents the vertex coordinates of the original 3D mesh, The corresponding vertex coordinates of the three-dimensional grid in which the watermark is embedded are represented. The same parameters as those mentioned above can be understood by referring to the meanings of the same parameters mentioned above.

[0147] Texture loss: Calculate the original texture map T in and watermarked texture maps The error between them can be expressed by the following formula (4):

[0148]

[0149] Among them, L t Indicates texture loss. For the same parameters as mentioned above, the meanings of the same parameters as mentioned above can be understood by referring to the meanings of the same parameters mentioned above.

[0150] Rendering loss: Calculate the rendered color image of the original 3D model (3D mesh with texture map) and a rendered image of a 3D model with a watermark embedded in it The error between them can be expressed by the following formula (5):

[0151]

[0152] Among them, L r represents the rendering loss, H r Represents the height of the texture map of the 3D mesh with texture map in the original 3D model, W r It represents the width of the texture map of the three-dimensional mesh with texture map in the original three-dimensional model. The same parameters as the above ones can be understood by referring to the meanings of the above same parameters.

[0153] In order to ensure the robustness against various attacks, it is necessary to constrain the target watermark information extracted from the attacked watermarked 3D model to be consistent with the input embedded watermark information.

[0154] Watermark recovery loss: The difference between the embedded watermark and the extracted geometric / texture watermark is calculated using the mean square error, which can be expressed by the following formula (6):

[0155]

[0156] Among them, L w represents the watermark recovery loss, N w Represents the embedded binary watermark sequence w in The length of the same parameters as above can be understood by referring to the meaning of the same parameters as above.

[0157] The final training loss (corresponding to the aforementioned target loss function) can be expressed by the following formula (7):

[0158] L=λ g L g +λ t L t +λ r L r +λ w L w +λ reg L reg (7)

[0159] Among them, L reg represents the network regularization loss, that is, the L2 sum of all network weights; λ g , t , r , w , reg They respectively represent the weight coefficients corresponding to geometric loss, texture loss, rendering loss, watermark recovery loss, and network regularization loss.

[0160] In the algorithm reasoning process, the watermark information is extracted from the geometric and texture parts of the digital human 3D model. (i.e. the target geometric watermark part), (i.e., the target texture watermark part) can be binarized by the following formula (8):

[0161]

[0162] Among them, sign refers to the sign function, and the clamp(.,0,1) function limits the input value to between 0 and 1. include and wout Express Through training, the target binary watermark sequence extracted from the geometric part or the texture part of the 3D model (corresponding to the aforementioned target watermark information) should be consistent with the embedded w in (corresponding to the aforementioned embedded watermark information) is consistent, thereby realizing the functions of digital human three-dimensional model authentication and infringement tracing.

[0163] Compared with the methods of related technologies, this application has the following beneficial effects:

[0164] 1) This application proposes to embed watermarks in both the geometry and texture parts of a digital human three-dimensional model, which can provide comprehensive and reliable protection for the geometry and texture of the digital human three-dimensional model, and extract the target watermark based on the geometry and / or texture parts of the attacked three-dimensional model with embedded watermarks. In this way, even if any part of the geometry or texture of the three-dimensional model is artificially deleted or missing, the complete target watermark can be extracted through the other part, thereby achieving the purpose of infringement tracing and avoiding the risk of failure in target watermark extraction due to malicious deletion or replacement of texture or geometry parts.

[0165] 2) This application proposes a deep learning watermarking algorithm that can achieve end-to-end training optimization. Through carefully designed training losses, the watermark embedding process does not destroy the geometric shape, texture details and rendering effects of the digital human three-dimensional model, does not affect the use value of the digital human assets, and can resist a variety of common attacks. It is efficient and has strong universality.

[0166] 3) The digital human three-dimensional model watermarking technology proposed in this application is aimed at various immersive businesses such as the metaverse and digital twins. It is used for copyright protection and infringement tracing of digital human assets, promotes the protection of the rights and interests of producers, improves the security of the use of digital human assets, and promotes the further development of the digital content industry.

[0167] In order to implement the watermark information processing method based on deep learning in the embodiment of the present application, the embodiment of the present application also provides a watermark information processing device based on deep learning, which is arranged on a watermark information processing device based on deep learning. Figure 6 Schematic diagram of the composition structure of the watermark information processing device based on deep learning in the embodiment of the present application. Figure 6 As shown, the watermark information processing device based on deep learning includes:

[0168] An acquisition unit 61, used to acquire the original three-dimensional model and the embedded watermark information;

[0169] A first processing unit 62 is used to input the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system to obtain a three-dimensional model embedded with a watermark; the three-dimensional model embedded with a watermark represents a three-dimensional model in which both the geometric part and the texture part of the original three-dimensional model are embedded with the embedded watermark information;

[0170] The second processing unit 63 is used to extract target watermark information from the attacked three-dimensional model with embedded watermark through the watermark processing system; the target watermark information is obtained based on the geometric part and / or texture part in the attacked three-dimensional model with embedded watermark, and is consistent with the embedded watermark information.

[0171] In one embodiment, the acquisition unit 61 is specifically configured to:

[0172] Acquire a three-dimensional data file; and preprocess the three-dimensional data file to obtain the original three-dimensional model.

[0173] In one embodiment, the first processing unit 62 includes a first extraction unit, a second extraction unit and a first generation unit; wherein,

[0174] The first extraction unit is used to input the geometric part of the original three-dimensional model into the first feature extraction network of the watermark embedding module of the watermark processing system to obtain the geometric part feature;

[0175] The second extraction unit is used to input the texture part of the original three-dimensional model into the second feature extraction network of the watermark embedding module of the watermark processing system to obtain the texture part feature;

[0176] The first generating unit is used to generate the watermark-embedded three-dimensional model based on the embedded watermark information, the geometric part features and the texture part features.

[0177] In one embodiment, the first generating unit is specifically configured to:

[0178] The embedded watermark information and the geometric part features are input into the first reconstruction network of the watermark embedding module to obtain the geometric part of the embedded watermark; the embedded watermark information and the texture part features are input into the second reconstruction network of the watermark embedding module to obtain the texture part of the embedded watermark; based on the embedded watermark geometric part and the embedded watermark texture part, a three-dimensional model of the embedded watermark is generated.

[0179] In one embodiment, the second processing unit 63 includes a first processing subunit and a second processing subunit; wherein,

[0180] The first processing subunit is used to perform one or more attack processing on the watermarked three-dimensional model through the simulated attack module of the watermark processing system during the training process to obtain the attacked watermarked three-dimensional model;

[0181] The second processing subunit is used to input the attacked watermarked three-dimensional model into the watermark extraction module of the watermark processing system to obtain the target watermark information.

[0182] In one embodiment, the second processing subunit includes a third extraction unit, a fourth extraction unit and a second generation unit; wherein,

[0183] The third extraction unit is used to input the attacked geometric part of the watermarked three-dimensional model into the first feature extraction network of the watermark extraction module to obtain the attacked geometric part feature;

[0184] The fourth extraction unit is used to input the texture part of the attacked watermarked three-dimensional model into the second feature extraction network of the watermark extraction module to obtain the features of the attacked texture part;

[0185] The second generating unit is used to generate the target watermark information based on the attacked geometric part features and / or the attacked texture part features.

[0186] In one embodiment, the second generating unit is specifically configured to:

[0187] Performing global average pooling processing on the attacked geometric part features to obtain target geometric watermark implicit features; performing global average pooling processing on the attacked texture part features to obtain target texture watermark implicit features; generating the target watermark information based on the target geometric watermark implicit features and / or the target texture watermark implicit features.

[0188] In another embodiment, the second generating unit is further specifically configured to:

[0189] The target geometric watermark implicit features are input into the watermark recovery network of the watermark extraction module to obtain the target geometric watermark part; the target texture watermark implicit features are input into the watermark recovery network of the watermark extraction module to obtain the target texture watermark part; based on the target geometric watermark part and / or the target texture watermark part, the target watermark information is generated.

[0190] In one embodiment, the watermark information processing device based on deep learning may further include: a construction unit and a training unit; wherein,

[0191] The construction unit is used to construct a target loss function;

[0192] The training unit is used to train the watermark processing system based on the target loss function to obtain a trained watermark processing system.

[0193] In one embodiment, the construction unit is specifically used for:

[0194] Construct a first loss function, a second loss function, a third loss function and a fourth loss function; the first loss function represents a geometric loss function, the second loss function represents a texture loss function, the third loss function represents a rendering loss function, and the fourth loss function represents a watermark recovery loss function; based on the first loss function, the second loss function, the third loss function and the fourth loss function, construct the target loss function.

[0195] In actual application, the acquisition unit 61 can be implemented by a communication interface in a watermark information processing device based on deep learning, and the first processing unit 62 and the second processing unit 63 can be implemented by a processor in the watermark information processing device based on deep learning.

[0196] It should be noted that: the watermark information processing device based on deep learning provided in the above embodiment only uses the division of the above program modules as an example when performing watermark information processing. In practical applications, the above processing can be assigned to different program modules as needed, that is, the internal structure of the device is divided into different program modules to complete all or part of the processing described above. In addition, the watermark information processing device based on deep learning provided in the above embodiment and the watermark information processing method based on deep learning belong to the same concept. The specific implementation process is detailed in the watermark information processing method based on deep learning, which will not be repeated here.

[0197] Based on the hardware implementation of the above program modules, and in order to implement the watermark information processing method based on deep learning in the embodiment of the present application, the embodiment of the present application also provides a watermark information processing device based on deep learning, Figure 7 Schematic diagram of the hardware structure of the watermark information processing device based on deep learning in the embodiment of the present application. Figure 7 As shown, the watermark information processing device 70 based on deep learning includes:

[0198] Communication interface 71, capable of exchanging information with other devices;

[0199] The processor 72 is connected to the communication interface 71 to realize information interaction with other devices, and is used to execute the watermark information processing method based on deep learning provided by one or more of the above technical solutions when running a computer program, and the computer program is stored in the memory 73.

[0200] Specifically, the communication interface 71 is used to obtain the original three-dimensional model and embed watermark information;

[0201] The processor 72 is used to input the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system to obtain a three-dimensional model embedded with a watermark; the three-dimensional model embedded with a watermark represents a three-dimensional model in which both the geometric part and the texture part of the original three-dimensional model are embedded with the embedded watermark information; the target watermark information is extracted from the attacked three-dimensional model embedded with a watermark through the watermark processing system; the target watermark information is obtained based on the geometric part and / or texture part in the attacked three-dimensional model embedded with a watermark, and is consistent with the embedded watermark information.

[0202] In one embodiment, the communication interface 71 is specifically used for:

[0203] Acquire a three-dimensional data file; and preprocess the three-dimensional data file to obtain the original three-dimensional model.

[0204] In one embodiment, the processor 72 is specifically configured to:

[0205] The geometric part of the original three-dimensional model is input into the first feature extraction network of the watermark embedding module of the watermark processing system to obtain the geometric part feature; the texture part of the original three-dimensional model is input into the second feature extraction network of the watermark embedding module of the watermark processing system to obtain the texture part feature; based on the embedded watermark information, the geometric part feature and the texture part feature, the watermarked three-dimensional model is generated.

[0206] In one embodiment, the processor 72 is specifically configured to:

[0207] The embedded watermark information and the geometric part features are input into the first reconstruction network of the watermark embedding module to obtain the geometric part of the embedded watermark; the embedded watermark information and the texture part features are input into the second reconstruction network of the watermark embedding module to obtain the texture part of the embedded watermark; based on the embedded watermark geometric part and the embedded watermark texture part, a three-dimensional model of the embedded watermark is generated.

[0208] In one embodiment, the processor 72 is specifically configured to:

[0209] During the training process, one or more attack processes are performed on the three-dimensional model embedded with the watermark through the simulated attack module of the watermark processing system to obtain the attacked three-dimensional model embedded with the watermark; the attacked three-dimensional model embedded with the watermark is input into the watermark extraction module of the watermark processing system to obtain the target watermark information.

[0210] In one embodiment, the processor 72 is specifically configured to:

[0211] The geometric part of the attacked watermarked three-dimensional model is input into the first feature extraction network of the watermark extraction module to obtain the attacked geometric part features; the texture part of the attacked watermarked three-dimensional model is input into the second feature extraction network of the watermark extraction module to obtain the attacked texture part features; based on the attacked geometric part features and / or the attacked texture part features, the target watermark information is generated.

[0212] In one embodiment, the processor 72 is specifically configured to:

[0213] Performing global average pooling processing on the attacked geometric part features to obtain target geometric watermark implicit features; performing global average pooling processing on the attacked texture part features to obtain target texture watermark implicit features; generating the target watermark information based on the target geometric watermark implicit features and / or the target texture watermark implicit features.

[0214] In one embodiment, the processor 72 is specifically configured to:

[0215] The target geometric watermark implicit features are input into the watermark recovery network of the watermark extraction module to obtain the target geometric watermark part; the target texture watermark implicit features are input into the watermark recovery network of the watermark extraction module to obtain the target texture watermark part; based on the target geometric watermark part and / or the target texture watermark part, the target watermark information is generated.

[0216] In one embodiment, the processor 72 is further configured to:

[0217] Constructing a target loss function; and training the watermark processing system based on the target loss function to obtain a trained watermark processing system.

[0218] In one embodiment, the processor 72 is specifically configured to:

[0219] Construct a first loss function, a second loss function, a third loss function and a fourth loss function; the first loss function represents a geometric loss function, the second loss function represents a texture loss function, the third loss function represents a rendering loss function, and the fourth loss function represents a watermark recovery loss function; based on the first loss function, the second loss function, the third loss function and the fourth loss function, construct the target loss function.

[0220] It should be noted that the specific processing process of the communication interface 71 and the processor 72 can be understood by referring to the above-mentioned watermark information processing method based on deep learning.

[0221] Of course, in actual application, the various components in the watermark information processing device 70 based on deep learning are coupled together through the bus system 74. It can be understood that the bus system 74 is used to realize the connection and communication between these components. In addition to the data bus, the bus system 74 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, Figure 7 Various buses are labeled as bus system 74 .

[0222] The memory 73 in the embodiment of the present application is used to store various types of data to support the operation of the deep learning-based watermark information processing device 70. Examples of such data include: any computer program used to operate on the deep learning-based watermark information processing device 70.

[0223] The watermark information processing method based on deep learning disclosed in the above embodiment of the present application can be applied to the processor 72, or implemented by the processor 72. The processor 72 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the watermark information processing method based on deep learning can be completed by the hardware integrated logic circuit or software instructions in the processor 72. The above processor 72 may be a general processor, a digital signal processor (DSP, Digital Signal Processor), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The processor 72 can implement or execute each watermark information processing method based on deep learning disclosed in the embodiment of the present application, steps and logic block diagrams. A general processor may be a microprocessor or any conventional processor, etc. In combination with the steps of the watermark information processing method based on deep learning disclosed in the embodiment of the present application, it can be directly embodied as a hardware decoding processor to execute, or it can be executed by a combination of hardware and software modules in the decoding processor. The software module may be located in a storage medium, which is located in the memory 73. The processor 72 reads the information in the memory 73 and completes the steps of the aforementioned watermark information processing method based on deep learning in combination with its hardware.

[0224] In an exemplary embodiment, the deep learning-based watermark information processing device 70 can be implemented by one or more application-specific integrated circuits (ASIC), DSP, programmable logic device (PLD), complex programmable logic device (CPLD), field programmable gate array (FPGA), general-purpose processor, controller, microcontroller (MCU), microprocessor, or other electronic components to execute the aforementioned deep learning-based watermark information processing method.

[0225] It can be understood that the memory 73 of the embodiment of the present application can be a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk memory or a tape memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM, SyncLink Dynamic Random Access Memory), and direct RAMbus random access memory (DRRAM, Direct Rambus Random Access Memory).The memory 73 described in the embodiments of the present application is intended to include but is not limited to these and any other suitable types of memory.

[0226] In an exemplary embodiment, the present application also provides a storage medium, namely a computer storage medium, specifically a computer-readable storage medium, for example, a memory 73 storing a computer program, and the above-mentioned computer program can be executed by a processor 72 in a watermark information processing device 70 based on deep learning to complete the steps of the watermark information processing method based on deep learning described in the above-mentioned embodiment of the present application. Among them, the computer-readable storage medium can be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, Flash Memory, magnetic surface storage, optical disk, or CD-ROM.

[0227] In an exemplary embodiment, the present application also provides a computer program product, including a computer program, which can be executed by a processor 72 in a deep learning-based watermark information processing device 70 to complete the steps of the deep learning-based watermark information processing method described in the aforementioned embodiment of the present application.

[0228] It should be noted that: "first", "second", "third", etc. are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0229] In addition, the technical solutions described in the embodiments of the present application can be combined arbitrarily without conflict.

[0230] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A watermark information processing method based on deep learning, characterized in that: The method comprises: Obtain the original 3D model and embedded watermark information; Inputting the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system to obtain a three-dimensional model embedded with a watermark; the three-dimensional model embedded with the watermark represents a three-dimensional model in which both the geometric part and the texture part of the original three-dimensional model are embedded with the embedded watermark information; The target watermark information is extracted from the attacked three-dimensional model with embedded watermark by the watermark processing system; the target watermark information is obtained based on the geometric part and / or texture part of the attacked three-dimensional model with embedded watermark and is consistent with the embedded watermark information.

2. The method according to claim 1, characterized in that The obtaining of the original three-dimensional model includes: Obtain three-dimensional data files; The three-dimensional data file is preprocessed to obtain the original three-dimensional model.

3. The method according to claim 1, characterized in that The step of inputting the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system to obtain the watermarked three-dimensional model comprises: Inputting the geometric part of the original three-dimensional model into the first feature extraction network of the watermark embedding module of the watermark processing system to obtain the geometric part feature; Inputting the texture portion of the original three-dimensional model into the second feature extraction network of the watermark embedding module of the watermark processing system to obtain texture portion features; The watermark-embedded three-dimensional model is generated based on the watermark-embedded information, the geometric part features and the texture part features.

4. The method according to claim 3, characterized in that The step of generating the watermarked three-dimensional model based on the watermarked information, the geometric features, and the texture features includes: Inputting the embedded watermark information and the geometric part features into a first reconstruction network of the watermark embedding module to obtain the geometric part of the embedded watermark; Inputting the embedded watermark information and the texture part features into the second reconstruction network of the watermark embedding module to obtain the texture part embedded with the watermark; A three-dimensional model of the embedded watermark is generated based on the geometric part of the embedded watermark and the texture part of the embedded watermark.

5. The method according to claim 1, wherein The method of extracting target watermark information from the attacked watermarked three-dimensional model by the watermark processing system includes: During the training process, one or more attack processes are performed on the watermarked three-dimensional model through the simulated attack module of the watermark processing system to obtain the attacked watermarked three-dimensional model; The attacked watermarked three-dimensional model is input into the watermark extraction module of the watermark processing system to obtain the target watermark information.

6. The method according to claim 5, characterized in that The step of inputting the attacked watermarked three-dimensional model into the watermark extraction module of the watermark processing system to obtain the target watermark information comprises: Inputting the attacked geometric part of the watermarked three-dimensional model into the first feature extraction network of the watermark extraction module to obtain features of the attacked geometric part; Inputting the texture portion of the attacked watermarked three-dimensional model into the second feature extraction network of the watermark extraction module to obtain features of the attacked texture portion; The target watermark information is generated based on the attacked geometric part features and / or the attacked texture part features.

7. The method according to claim 6, characterized in that The generating the target watermark information based on the attacked geometric part feature and / or the attacked texture part feature includes: Performing global average pooling processing on the attacked geometric part features to obtain the target geometric watermark implicit features; Performing global average pooling processing on the attacked texture part features to obtain the target texture watermark implicit features; The target watermark information is generated based on the target geometric watermark implicit features and / or the target texture watermark implicit features.

8. The method according to claim 7, characterized in that The generating the target watermark information based on the target geometric watermark implicit feature and / or the target texture watermark implicit feature includes: Inputting the target geometric watermark implicit features into the watermark recovery network of the watermark extraction module to obtain the target geometric watermark part; Inputting the target texture watermark implicit features into the watermark recovery network of the watermark extraction module to obtain the target texture watermark part; The target watermark information is generated based on the target geometric watermark portion and / or the target texture watermark portion.

9. The method according to claim 1, characterized in that The method further comprises: Construct the target loss function; The watermark processing system is trained based on the target loss function to obtain a trained watermark processing system.

10. The method according to claim 9, characterized in that The constructing of the target loss function includes: Constructing a first loss function, a second loss function, a third loss function and a fourth loss function; the first loss function represents a geometric loss function, the second loss function represents a texture loss function, the third loss function represents a rendering loss function, and the fourth loss function represents a watermark recovery loss function; The target loss function is constructed based on the first loss function, the second loss function, the third loss function and the fourth loss function.

11. A watermark information processing device based on deep learning, characterized in that: The device comprises: An acquisition unit, used for acquiring the original three-dimensional model and embedding watermark information; a first processing unit, configured to input the original three-dimensional model and the embedded watermark information into a pre-trained watermark processing system to obtain a watermarked three-dimensional model; the watermarked three-dimensional model represents a three-dimensional model in which both the geometric part and the texture part of the original three-dimensional model are embedded with the embedded watermark information; The second processing unit is used to extract target watermark information from the attacked three-dimensional model with embedded watermark through the watermark processing system; the target watermark information is obtained based on the geometric part and / or texture part of the attacked three-dimensional model with embedded watermark, and is consistent with the embedded watermark information.

12. A watermark information processing device based on deep learning, characterized in that: include: a processor and a memory for storing a computer program capable of running on said processor; Wherein, when the processor is used to run the computer program, it executes the steps of the method according to any one of claims 1 to 10.

13. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

14. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 10.

Citation Information

Cited By

  • Digital human anti-counterfeiting method and device and electronic equipment

    CN119646779A