Model watermark embedding method, apparatus, device, and computer storage medium

Abstract

The present application relates to the technical field of images, and discloses a model watermark embedding method, an apparatus, a device, and a computer storage medium. The method comprises: acquiring a density prediction value and a color prediction value of a target object in a three-dimensional space by means of a neural radiance field network; embedding watermark information into the color prediction value to obtain a watermarked color feature value; then performing volume rendering on the density prediction value and the color feature value to obtain color values of the target object in the three-dimensional space; and obtaining a two-dimensional image of the target object at any viewing angle on the basis of the color values, i.e., obtaining a three-dimensional model corresponding to the target object. In this way, the watermark information can be embedded into the three-dimensional model of the target object, achieving copyright protection for the three-dimensional model.

Description

Model watermark embedding method, device, setting and computer storage medium Technical Field

[0001] The present application relates to the field of image technology, and in particular to a model watermark embedding method, device, setting and computer storage medium. Background Art

[0002] At present, with the rapid development of applied image technology, the application of 3D model data in various fields is becoming more and more common. Users can publish their own 3D models, research and download useful 3D model works through the Internet.

[0003] A large number of 3D models are circulated on the Internet. However, since these 3D models can be copied at will and spread quickly, it is difficult to protect the copyright of the 3D models.

[0004] Summary of the Invention

[0005] The present application embodiment provides a model watermark embedding method. The technical solution is as follows:

[0006] According to one aspect of the present application, a model watermark embedding method is provided, the method comprising:

[0007] Acquire an image dataset and watermark information, wherein the image dataset includes image information of at least one two-dimensional picture of a target object;

[0008] Processing the image data set through a neural radiation field network to obtain a density prediction value and a color prediction value of the target object in three-dimensional space;

[0009] Using steganography to fuse the watermark information and the color prediction value to obtain a watermarked color feature value;

[0010] Performing volume rendering processing on the density prediction value and the color feature value to obtain a color value of the target object in a three-dimensional space;

[0011] A three-dimensional model corresponding to the target object is obtained according to the color value.

[0012] Optionally, the watermark information and the color prediction value are fused by using steganography to obtain a watermarked color feature value, including:

[0013] The watermark information and the color prediction value are fused through a convolutional neural network to obtain the watermarked color feature value;

[0014] Alternatively, the watermark information and the color prediction value are fused through a generative adversarial network to obtain the watermarked color feature value.

[0015] Optionally, performing volume rendering processing on the density prediction value and the color prediction value to obtain a color value of the target object in a three-dimensional space includes:

[0016] The color value is obtained by a volume rendering formula based on the camera's viewing angle, and the volume rendering formula includes:

[0017] Wherein, C(x, d) is the color observed from the position x of the camera along the direction d, that is, the color value, σ(p) is the density prediction value of position p, c(p) is the watermarked color feature value of position p, T(t) is the transmittance from the position of the camera to position p, tfar is the far boundary of the preset three-dimensional scene, tnear is the near boundary of the preset three-dimensional scene, the near boundary is located on the side of the far boundary close to the position of the camera, and the preset three-dimensional scene includes the target object.

[0018] Optionally, the watermark information S(x), the color prediction value M(p) and the watermarked color feature value c(p) satisfy an embedding formula, and the embedding formula includes: c(p)=M(p)+α*S(x);

[0019] Among them, α is the preset embedding strength parameter.

[0020] Optionally, the image dataset includes image information of at least two two-dimensional pictures of a target object, and the image information of the two-dimensional pictures includes three-dimensional coordinates, posture information, and a target feature map;

[0021] The acquiring of the image data set comprises:

[0022] Acquire N target object images, where the N target object images are two-dimensional images obtained by capturing images of the same target object at different angles, where N is an integer greater than or equal to 2;

[0023] Based on the N target object images, obtaining three-dimensional coordinates, posture information, and N target feature maps corresponding to the target objects, where the target feature maps are images obtained by segmenting the target object images;

[0024] Based on the three-dimensional coordinates, the pose information, and the target feature map, image information of the two-dimensional image is obtained to acquire an image dataset.

[0025] Optionally, the acquiring, based on the N target object images, the three-dimensional coordinates, posture information, and target feature map corresponding to the target object includes:

[0026] Extracting multiple feature points from each of the N target object images through a feature detection network;

[0027] Obtaining a feature descriptor for each feature point;

[0028] Matching the feature points of the N target object images based on the feature descriptors to obtain geometric relationships between the matched feature points in the N target object images;

[0029] According to the geometric relationship, obtaining the posture information;

[0030] Acquire three-dimensional coordinates of a plurality of the feature points according to the geometric relationship and the posture information;

[0031] Acquiring the three-dimensional coordinates of the target object based on the three-dimensional coordinates of the multiple feature points;

[0032] The N target object images are input into a convolutional neural network for image segmentation processing to segment the target feature map.

[0033] Optionally, acquiring the image dataset includes:

[0034] Acquire a target object picture, wherein the target object picture is a picture containing an image of the target object;

[0035] Based on the target object image, obtaining N converted images, where the N converted images are two-dimensional images of different perspectives generated after processing the target object image, where N is an integer greater than or equal to 2;

[0036] Based on the N converted images, obtaining three-dimensional coordinates, posture information, and N target feature maps corresponding to the target object, where the target feature maps are images obtained by segmenting the converted images;

[0037] Based on the three-dimensional coordinates, the pose information, and the target feature map, image information of the two-dimensional image is obtained to acquire an image dataset.

[0038] Optionally, obtaining N converted images based on the target object image includes:

[0039] Performing depth estimation on the target object image to obtain a depth map corresponding to the target object image;

[0040] Extracting key feature points from the target object image through a feature detection network;

[0041] Based on the depth map and the key feature points, N converted images corresponding to the target object image are obtained.

[0042] Optionally, obtaining a target object image includes:

[0043] Obtaining text information describing the target object;

[0044] The text information is converted into the target object image through a deep learning network.

[0045] Optionally, obtaining the watermark information includes:

[0046] Get the user's unique identity information;

[0047] Converting the identity information into a binary sequence to form initial watermark information in a binary format;

[0048] The initial watermark information is subjected to spread spectrum processing to obtain the encoded watermark information.

[0049] Optionally, after obtaining the three-dimensional model corresponding to the target object, the method further includes:

[0050] Converting the three-dimensional model from the time domain to the frequency domain by fast Fourier transform to obtain a target three-dimensional model;

[0051] Performing frequency detection on the target three-dimensional model to determine whether the target three-dimensional model has a frequency corresponding to the watermark information, wherein the frequency of the watermark information is different from the frequency of the three-dimensional model;

[0052] In response to detecting a frequency corresponding to the watermark information in the target three-dimensional model, the watermark information is extracted from the target three-dimensional model.

[0053] Optionally, extracting the watermark information from the target three-dimensional model includes:

[0054] separating target watermark information from the target three-dimensional model according to the frequency corresponding to the watermark information;

[0055] The target watermark information is converted from the frequency domain to the time domain by inverse Fourier transform to obtain the watermark information.

[0056] Optionally, in response to detecting the frequency corresponding to the watermark information in the target three-dimensional model, extracting the watermark information from the target three-dimensional model, the method further includes:

[0057] performing inverse encoding on the watermark information to restore the original watermark information;

[0058] The initial watermark information is converted into a text format to obtain the identity information corresponding to the watermark information.

[0059] Optionally, acquiring a three-dimensional model corresponding to the target object according to the color value includes:

[0060] Obtaining an initial three-dimensional model according to the color value;

[0061] Based on the initial three-dimensional model, extracting a grid feature map corresponding to the initial three-dimensional model;

[0062] The grid feature map is colored according to the color value to obtain a texture grid, wherein the texture grid includes two-dimensional images of the target object under multiple viewing angles, and the texture grid is a three-dimensional model corresponding to the target object.

[0063] Optionally, extracting a mesh feature map corresponding to the initial three-dimensional model based on the initial three-dimensional model includes:

[0064] Extracting an initial grid feature map based on the initial three-dimensional model;

[0065] The initial grid feature map is iteratively optimized a preset number of times through a neural radiation field network to obtain the grid feature map.

[0066] According to another aspect of the present application, a model watermark embedding device is provided, the device comprising:

[0067] an acquisition module, configured to acquire an image data set and watermark information, wherein the image data set includes image information of at least one two-dimensional picture of a target object;

[0068] a radiation module, configured to process the image data set through a neural radiation field network to obtain a density prediction value and a color prediction value of the target object in three-dimensional space;

[0069] A fusion module, configured to fuse the watermark information and the color prediction value using steganography to obtain a watermarked color feature value;

[0070] a rendering module, configured to perform volume rendering processing on the density prediction value and the color feature value to obtain a color value of the target object in a three-dimensional space;

[0071] The integration module is used to obtain a three-dimensional model corresponding to the target object according to the color value.

[0072] Optionally, the fusion module includes:

[0073] A convolution unit, configured to fuse the watermark information and the color prediction value through a convolutional neural network to obtain the watermarked color feature value;

[0074] Alternatively, the adversarial unit is configured to fuse the watermark information and the color prediction value through a generative adversarial network to obtain the watermarked color feature value.

[0075] According to another aspect of the present application, a model watermark embedding device is provided, which includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, the at least one program, the code set or instruction set is loaded and executed by the processor to implement the model watermark embedding method as described above.

[0076] According to another aspect of the present application, a computer storage medium is provided, in which at least one instruction, at least one program, code set or instruction set is stored. The at least one instruction, the at least one program, the code set or instruction set is loaded and executed by a processor to implement the model watermark embedding method as described above.

[0077] The beneficial effects of the technical solutions provided in the embodiments of the present application include at least:

[0078] A model watermark embedding method is provided. Using a neural radiance field network, the density and color predictions of a target object in three-dimensional space are obtained. Watermark information is then embedded into the color predictions to generate watermarked color eigenvalues. Volume rendering is then performed on these density and color predictions to obtain the target object's color values ​​in three-dimensional space. Based on these color values, a two-dimensional image of the target object at any viewing angle is generated, thereby obtaining the corresponding three-dimensional model of the target object. This method can embed watermark information into the target object's three-dimensional model, resolving the difficulty of copyright protection for 3D models in related technologies and effectively protecting the copyright of the generated 3D model.

[0079] In addition, since the density prediction value is data related to the geometric features of the three-dimensional model of the target object, if the watermark information is embedded in the density prediction value, it is more likely to have a greater impact on the subsequently generated three-dimensional model. This application can reduce the impact of the watermark information on the three-dimensional model by only adding watermark information to the color prediction value. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0081] FIG1 is a flow chart of a model watermark embedding method provided by an embodiment of the present application;

[0082] FIG2 is a flow chart of another model watermark embedding method provided by an embodiment of the present application;

[0083] FIG3 is a flow chart of obtaining an image data set according to an embodiment of the present application;

[0084] FIG4 is a flowchart of another embodiment of the present application for obtaining an image data set;

[0085] FIG5 is a flowchart of obtaining a depth map provided by an embodiment of the present application;

[0086] FIG6 is a flow chart of obtaining watermark information provided by an embodiment of the present application;

[0087] FIG7 is a flowchart of a neural radiation field network provided in an embodiment of the present application;

[0088] FIG8 is a flow chart of another model watermark embedding method provided by the present application;

[0089] FIG9 is a structural block diagram of a model watermark embedding device provided in an embodiment of the present application.

[0090] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0091] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0092] First, the application scenarios involved in the embodiments of the present application are introduced.

[0093] 3D model reconstruction technology refers to an image processing technique that reconstructs 3D information from single-view or multi-view images to obtain a complete 3D model. For example, in the medical industry, 3D model reconstruction technology can be used to create accurate models of human organs. In the mechanical engineering industry, 3D model reconstruction technology can be used to demonstrate new equipment and vehicles. 3D model reconstruction technology can also include 3D model reconstruction technology based on machine learning.

[0094] Machine learning is a branch of artificial intelligence. It uses computers as tools to learn from big data representations of real-world phenomena that can be directly applied to computer calculations. In the field of imaging technology, machine learning can be applied to object detection, image generation, and image segmentation. Artificial intelligence is a technology that studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0095] Digital watermarking technology uses algorithms to embed identifying information into images, videos, audio, or 3D models without affecting the original image, video, audio, or 3D model. Watermarks embedded in this manner are often imperceptible to users and are resilient to certain attacks. Digital watermarking can embed author information, company logos, file transfer information, and other information into original images, videos, audio, or 3D models for copyright protection and content authentication.

[0096] It should be noted that the application scenarios described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Ordinary technicians in this field can know that the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0097] The implementation environment may include a camera, a server, and a display terminal. The camera may include a camera that can be used to capture real photos into digitized images. The server includes a processor that can establish a wired or wireless connection with the camera to generate a three-dimensional model based on the images captured by the camera and display the three-dimensional model on the display terminal.

[0098] FIG1 is a flow chart of a model watermark embedding method provided by an embodiment of the present application. The method can be applied to the server in the above implementation environment. The method may include the following steps:

[0099] Step 101: Obtain image data set and watermark information.

[0100] The image data set includes image information of at least one two-dimensional picture of a target object.

[0101] Step 102: Process the image data set through a neural radiation field network to obtain a density prediction value and a color prediction value of the target object in three-dimensional space.

[0102] Step 103: Using steganography to fuse the watermark information and the color prediction value to obtain the watermarked color feature value.

[0103] Step 104: Perform volume rendering processing on the density prediction value and the color feature value to obtain the color value of the target object in three-dimensional space.

[0104] Step 105: Obtain a three-dimensional model corresponding to the target object according to the color value.

[0105] In summary, the embodiments of the present application provide a model watermark embedding method. This method uses a neural radiation field network to obtain density prediction values ​​and color prediction values ​​of a target object in three-dimensional space. Watermark information is then embedded into the color prediction values ​​to obtain watermarked color feature values. Volume rendering is then performed on the density prediction values ​​and color feature values ​​to obtain the color values ​​of the target object in three-dimensional space. Based on the color values, a two-dimensional image of the target object at any viewing angle is obtained, thereby obtaining a corresponding three-dimensional model of the target object. In this way, watermark information can be embedded into the three-dimensional model of the target object, resolving the difficulty of copyright protection for three-dimensional models in related technologies and achieving the copyright protection of the generated three-dimensional model.

[0106] In addition, since the density prediction value is data related to the geometric features of the three-dimensional model of the target object, if the watermark information is embedded in the density prediction value, it is more likely to have a greater impact on the subsequently generated three-dimensional model. This application can reduce the impact of the watermark information on the three-dimensional model by only adding watermark information to the color prediction value.

[0107] FIG2 is a flow chart of another model watermark embedding method provided by an embodiment of the present application. The method can be applied to the server in the above implementation environment. The method may include the following steps:

[0108] Step 201: Acquire an image dataset.

[0109] The image data set includes image information of at least one two-dimensional picture of a target object.

[0110] At least one two-dimensional picture with a target object may be referred to as a target object picture, which may include an image of the target object and its surrounding scene. The target object may be an object such as a building, a machine, or a vehicle.

[0111] In one possible implementation, the target object image can be a two-dimensional image obtained by taking a photo with a camera of the terminal device; or, the target object image can also be a two-dimensional image obtained from inside the terminal device. For example, the target object image can be a two-dimensional image stored in the photo album of the terminal device, or a two-dimensional image obtained by the terminal device from the cloud; or, the target object image can be a two-dimensional image generated by a simple text description.

[0112] In an optional embodiment, the image dataset may include image information of at least two two-dimensional pictures of the target object, and the image information of the two-dimensional pictures includes three-dimensional coordinates, posture information, and a target feature map.

[0113] As shown in Figure 3, obtaining an image dataset may include the following three sub-steps:

[0114] Sub-step 2011: Obtain N target object images.

[0115] The N target object images are two-dimensional images obtained by capturing images of the same target object at different angles, where N is an integer greater than or equal to 2.

[0116] The user can use a camera or a mobile terminal with a camera function to shoot the target object from multiple angles to obtain N target object images, where the target object in the N target object images has different postures and different states. For example, 49 target object images can be obtained. The N target object images can provide basic data for subsequent three-dimensional model reconstruction. The user can use augmented reality technology (English: Augmented Reality; abbreviated: AR) technology to shoot the target object to reduce the difficulty of shooting the target object.

[0117] Sub-step 2012: Based on the N target object images, obtain the three-dimensional coordinates, posture information and N target feature maps corresponding to the target object.

[0118] Among them, the target feature map is an image obtained by segmenting the target object image.

[0119] Based on N target object images, a pose estimation algorithm can be used to obtain the precise position and orientation of the camera in 3D space for each target object image. This is done by first detecting and extracting feature points from all target object images, then matching the extracted feature points. The camera's pose information and the coordinates of the feature points in 3D space are then obtained based on the matched feature points.

[0120] Based on N target object images, obtaining the corresponding 3D coordinates, pose information, and N target feature maps of the target object can include the following steps:

[0121] (1) Extract multiple feature points from each of N target object images through a feature detection network.

[0122] Points with significant features can be detected from each target object image as multiple feature points of each target object image. Feature points can have differences and repeatability. Difference means that feature points can present obvious features that are different from non-feature points, so that feature points can be detected; repeatability means that the same feature point can be repeatedly detected in different target object images. Exemplarily, a feature detection network can be used to extract multiple feature points from each of N target object images. Exemplarily, the feature detection network can include a scale-invariant feature transformation algorithm (English: Scale Invariant Feature Transform; abbreviated: SIFT) and a fast feature point extraction and description algorithm (English: Oriented FAST and Rotated BRIEF; abbreviated: ORB).

[0123] (2) Obtain the feature descriptor of each feature point.

[0124] Feature points and feature descriptors can have a one-to-one correspondence. Feature descriptors are used to describe the characteristics of the corresponding feature points within a certain area. Both feature points and feature descriptors are unique and invariant.

[0125] (3) Based on the feature descriptors, the feature points of the N target object images are matched to obtain the geometric relationship between the matched feature points in the N target object images.

[0126] Since feature descriptors are unique, feature points with the same or similar feature descriptors in N target object images can be matched based on the feature descriptors. According to the matching of feature points, the camera motion between the N target object images can be obtained, that is, the geometric relationship between the matched multiple feature points can be obtained.

[0127] (4) Obtain posture information based on geometric relationships.

[0128] The geometric relationship obtained by feature point matching is used to obtain the relative pose of N target object images taken by the camera, so as to obtain the camera's intrinsic parameters and the camera's extrinsic parameters, and the camera's intrinsic parameters and extrinsic parameters are used as pose information.

[0129] (5) According to the geometric relationship and posture information, the three-dimensional coordinates of multiple feature points are obtained, and based on the three-dimensional coordinates of the multiple feature points, the three-dimensional coordinates of the target object are obtained.

[0130] Based on the geometric relationship between the camera's pose information and the matched feature points, the feature points in the N target object images are mapped from the camera coordinate system to a three-dimensional coordinate system to obtain the three-dimensional coordinates of the multiple feature points. Since the multiple feature points represent the salient features of the target object, the coordinates of the entire target object in the three-dimensional coordinate system can be obtained based on the three-dimensional coordinates of the multiple feature points. For example, the coordinates of the multiple feature points in three-dimensional space can be solved by triangulation.

[0131] (6) Input N target object images into the convolutional neural network for image segmentation processing to segment the target feature map.

[0132] Image segmentation technology (English: Image Segmentation) refers to the technology of dividing an image into multiple image subsets to facilitate the analysis of the target object in the image. In the embodiment of the present application, image segmentation technology can be used to extract the image of the target object from each target object image, thereby reducing background interference and improving the quality of the subsequently generated 3D model. The extracted image of the target object is the target feature map.

[0133] A convolutional neural network can be used to perform image segmentation on each target object image. For example, the convolutional neural network can include a U-Net network or a Mask R-CNN network. The convolutional neural network can identify and segment specific objects or areas in the image through deep learning.

[0134] In an optional embodiment, the following four methods can also be used to segment the target object image. The threshold segmentation method can set a preset threshold and segment the target object image into foreground and background according to the preset threshold. Exemplarily, the target object image is processed into a grayscale image, and all areas in the grayscale image with pixel values ​​higher than the preset threshold can be divided into foreground. The color or texture-based segmentation method is that the color information or texture information in the target object image can be used to distinguish different objects or areas. The edge detection segmentation method determines the boundary of the object by finding the area with sudden changes in color or brightness in the target object image. The region growing segmentation method can first select at least one target pixel in the target object image, and add pixels with similar characteristics adjacent to the target pixel to the same area.

[0135] (7) Optimize the pose information and three-dimensional coordinates of the target object through the bundle adjustment method (English: Bundle Adjustment; abbreviated: BA).

[0136] The bundle adjustment method can be used to optimize the camera's pose information and the coordinates of the target object in three-dimensional space to ensure that the target objects photographed from different perspectives can be aligned in the same coordinate system, which can improve the accuracy of the subsequently generated three-dimensional model.

[0137] Sub-step 2013: Based on the three-dimensional coordinates, the posture information, and the target feature map, obtain the image information of the two-dimensional image to obtain an image dataset.

[0138] This image dataset can be used as input information for the Neural Radiance Fields (NeRF) network. That is, the target feature map and the corresponding three-dimensional coordinates and pose information are input into the NeRF network to synthesize a two-dimensional image of the target object from any perspective through the NeRF network.

[0139] In an optional embodiment, as shown in FIG4 , obtaining an image dataset may include the following four sub-steps:

[0140] Sub-step 2014: Obtain a target object image.

[0141] The number of the target object picture can be one, and the target object picture is a picture containing an image of the target object. The target object picture can be obtained by photographing the target object, or the target object picture can be a two-dimensional picture stored in the terminal device.

[0142] Optionally, the target object image can be a two-dimensional image generated from a simple text description. Textual information describing the target object can be obtained first. For example, a user can upload the textual information of the target object to be converted into an image to a server via a terminal. The uploaded text information can be in at least one of the following formats: txt, pdf, and docx. The server can receive the text information and pre-process it. This pre-processing can include at least one of text cleaning, formatting, and encoding conversion.

[0143] The text information is then converted into an image of the target object through a deep learning network. The deep learning network can be trained in advance so that it can understand the text information and convert it into an image representation related to the text information.

[0144] The conversion process of converting text information into a target object image through a deep learning network can be regarded as a conversion process of a conditional generation task, where the text information is a condition and the goal is to generate an image corresponding to the text information. For example, the semantics of the text information can be understood by a language model (English: ChatGLM-6B), and the image can be generated by a generative model (English: Stable diffusion). The generative model can be fine-tuned by a low-rank adaptation of a large language model (English: Low-Rank Adaptation of Large Language Models; abbreviated: LoRA) and a control network (English: ControlNet) to improve the quality of the generated target object image.

[0145] The conversion function that converts the text information describing the target object into the target object image can be: I=G(T|θ G )

[0146] Among them, I is the generated target object image, T is the input text information, G is a text image generation model, θ G are the parameters of the generative model.

[0147] Sub-step 2015: Obtain N converted images based on the target object image.

[0148] The N converted images are two-dimensional images at different perspectives generated after processing the target object image, where N is an integer greater than or equal to 2.

[0149] Obtaining N converted images can include the following three steps:

[0150] (1) Perform depth estimation on the target object image to obtain a depth map corresponding to the target object image.

[0151] Depth estimation is a computer vision technique that can generate a depth image from a two-dimensional image. The depth image contains information about the distance from the viewpoint (which can be the camera that captured the image) to the objects in the image. A deep learning network can be used to estimate the depth of each pixel in the image of the target object. This deep learning network can be a monocular depth estimation network (MonoDepth).

[0152] The depth value can be obtained through the depth formula, which can be: z = D(I, p); where D(I, p) represents the depth function, I represents the input target object image, p represents a pixel in the target object image, and z is the depth value of pixel p returned by the depth function.

[0153] Please refer to Figure 5, which is a flowchart of obtaining a depth map provided by an embodiment of the present application. In an optional implementation, a left-eye view (which can be a target object image) can be obtained first, and the left-eye view can be processed to obtain a right-eye view. The obtained right-eye view is then optimized to obtain an optimized right-eye view, and a binocular view is obtained based on the left-eye view and the optimized right-eye view. Disparity estimation is performed based on the binocular view, and the depth value of each pixel in the target object image can be obtained based on the result of the disparity estimation.

[0154] There are two methods for processing the left-eye view to obtain the right-eye view. As shown in Solution 1 in Figure 5, the first method is to perform monocular depth estimation on the left-eye view to obtain a depth map for the left-eye view. This depth map is then projected and structured to obtain the right-eye view. As shown in Solution 2 in Figure 5, the left-eye view is input into a deep network, which uses the input left-eye view to estimate the view from another viewpoint to obtain the right-eye view.

[0155] (2) Extract key feature points from the target object image through the feature detection network.

[0156] A feature detection network may be used to extract key feature points in a two-dimensional image. The key feature points may include feature points of a target object image in the target object image, but not include feature points of a background image in the target object image.

[0157] (3) Based on the depth map and key feature points, obtain N converted images corresponding to the target object image.

[0158] The depth map can be combined with key feature points to generate N converted images with different perspectives.

[0159] Optionally, before extracting key feature points from the target object image using the feature detection network, saliency detection can also be performed on the target object image to identify and extract the target object and its region in the target object image, thereby preventing omission of the target object image during subsequent image processing. The formula for saliency detection can be s = S(I, p); where S(I, p) is a saliency function that returns the saliency value of pixel p, with the saliency value ranging from [0 to 1].

[0160] Sub-step 2016: Based on the N converted images, obtain the three-dimensional coordinates, posture information and N target feature maps corresponding to the target object.

[0161] Among them, the target feature map is an image obtained by segmenting the converted image.

[0162] Sub-step 2017: Based on the three-dimensional coordinates, the pose information, and the target feature map, obtain the image information of the two-dimensional image to obtain an image dataset.

[0163] The implementation process of sub-steps 2016 and 2017 can refer to sub-steps 2012 and 2013, and this embodiment of the application will not be repeated here.

[0164] Step 202: Obtain watermark information.

[0165] Watermark information can be used to record copyright owner information for use in copyright disputes. By associating watermark information with the user's identity information, the original creator of the file can be traced and the ultimate ownership of the file can be proved.

[0166] As shown in Figure 6, obtaining watermark information may include the following three sub-steps:

[0167] Sub-step 2021: Obtain the user's unique identity information.

[0168] A user's unique identification information may include a username, user ID, or email address. When a user uses a terminal to capture an image of an object, or uploads text information about an object, they can log in or authenticate their account on the mobile device to obtain their unique identification information. Authentication can include a username and password combination, or more advanced biometric technologies such as facial recognition or fingerprint scanning.

[0169] Sub-step 2022: Convert the identity information into a binary sequence to form initial watermark information in binary format.

[0170] By converting identity information into a binary sequence to form initial watermark information in binary format, a unique binary token can be generated for each user, making the initial watermark information personalized and difficult to imitate. This initial watermark information can be called a secret label.

[0171] The corresponding relationship between each user's watermark information and his / her identity information can be stored in a database.

[0172] Sub-step 2023: Perform spread spectrum processing on the initial watermark information to obtain encoded watermark information.

[0173] The initial watermark information is subjected to spread spectrum processing to facilitate subsequent detection of the encoded watermark information, and the ability of the watermark information to resist attacks and interference can be improved.

[0174] In an optional implementation, the watermark information can be generated by a hash function. The formula of the hash function is:u =H(U|θ H ); where W u is the user's watermark information, U is the user's unique identity information, H is the hash function, θ H are the parameters of the hash function.

[0175] Step 203: Process the image data set through the neural radiation field network to obtain the density prediction value and color prediction value of the target object in three-dimensional space.

[0176] Neural Radiance Fields (NeRF) is a deep learning framework for synthesizing 3D scenes from sparse or irregular 2D viewports. NeRF uses a neural network to represent the 3D radiance field of a scene, enabling high-quality 3D reconstruction of complex scenes.

[0177] The Neural Radiance Field Network (NRF) process consists of two main stages: 3D reconstruction and volume rendering. During the 3D reconstruction stage, the NRF uses a neural network to map the image dataset into pixel-wise color and density predictions. This neural network is a multilayer perceptron (MLP), where each layer contains multiple hidden units.

[0178] The density prediction value indicates the probability that a ray will terminate at that pixel, while the color prediction value indicates the color (RGB value) corresponding to that pixel. In other words, the density prediction value reflects the density of particles in the 3D model at any given ray, that is, the density of particles at a specific 3D coordinate, while the color prediction value reflects the color reflected by the particles at that specific 3D coordinate, as viewed from the direction of that ray.

[0179] Step 204: Using steganography to fuse the watermark information and the color prediction value to obtain the watermarked color feature value.

[0180] Steganography can be used to embed watermark information into color prediction values. For example, frequency domain steganography technology can be used to embed watermark information into color prediction values. Frequency domain steganography technology refers to transforming the color prediction value into the frequency domain (also called the wavelet domain) through transform domain technology (such as Fourier transform, discrete cosine transform or wavelet transform), and then adding a watermark to the color prediction value in the frequency domain, and converting the color prediction value back to the spatial domain through inverse transformation to obtain the watermarked color feature value.

[0181] Alternatively, a characteristic of human vision, that is, the perception of changes below a certain threshold is not obvious, can be used to fuse the watermark information and the color prediction value through the least significant bit (English: Least Significant Bit; abbreviated: LSB) technology.

[0182] In an optional implementation, the watermark information and the color prediction value may be fused through a convolutional neural network (CNN) to obtain a watermarked color feature value.

[0183] Alternatively, the watermark information and the color prediction value are fused through a generative adversarial network (English: Generative Adversarial Networks; abbreviated: GAN) to obtain the watermarked color feature value.

[0184] Step 205: Perform volume rendering processing on the density prediction value and the color feature value to obtain the color value of the target object in three-dimensional space.

[0185] The volume rendering stage of the neural radiance field network is implemented using a differentiable volume rendering equation. This equation calculates the color from the camera's position to the pixel based on the scene's 3D coordinates and the view direction. Specifically, each pixel is sampled using multiple random sample points and corresponding weights. The colors of all the sample points are accumulated to produce a color value, which represents the pixel color.

[0186] Optionally, a color value is obtained using a volume rendering formula based on the camera's viewing angle. The volume rendering formula includes:

[0187] Among them, C(x, d) is the color observed from the camera position x along the direction d, that is, the color value, σ(p) is the density prediction value of position p, c(p) is the watermarked color feature value of position p, T(t) is the transmittance from the camera position to position p, tfar is the far boundary of the preset three-dimensional scene, tnear is the near boundary of the preset three-dimensional scene, the near boundary is located on the side of the far boundary close to the camera position, and the preset three-dimensional scene includes the target object.

[0188] Optionally, the watermark information S(x), the color prediction value M(p) and the watermarked color feature value c(p) satisfy an embedding formula, and the embedding formula includes: c(p)=M(p)+α*S(x);

[0189] Where α is the preset embedding strength parameter. The watermark’s visibility can be adjusted by adjusting the embedding strength parameter.

[0190] Step 206: Obtain a three-dimensional model corresponding to the target object according to the color value.

[0191] By embedding watermark information into the color prediction value, a watermarked color feature value is obtained, and then the density prediction value and the color feature value are volume rendered to obtain the color value of the target object in three-dimensional space, and based on the color value, a two-dimensional graphic of the target object at any viewing angle is obtained, that is, a three-dimensional model corresponding to the target object is obtained. In this way, the watermark information can be embedded into the three-dimensional model of the target object, and the copyright of the generated three-dimensional model can be protected. In addition, since the density prediction value is data related to the geometric features of the three-dimensional model of the target object, if the watermark information is embedded in the density prediction value, it is easy to cause a greater impact on the subsequently generated three-dimensional model. The present application can reduce the impact of the watermark information on the three-dimensional model by adding watermark information only to the color prediction value. In other words, it is possible to add watermark information to the three-dimensional model without loss.

[0192] Please refer to Figure 7, which is a flowchart of a neural radiation field network provided in an embodiment of the present application. After completing the reconstruction step of the three-dimensional model with the watermark information added, the three-dimensional model can be displayed in two ways. The first is a fast display method, which can directly generate an MP4 format video showing the three-dimensional model. This method is suitable for users who need to quickly view the results of three-dimensional reconstruction. The second is a high-precision display method, which can export the mesh feature map for secondary editing to obtain a textured mesh. The mesh feature map can be obtained through the neural radiation field network to facilitate the user to render the three-dimensional model on the terminal.

[0193] Optionally, after generating the three-dimensional model, the three-dimensional model can be rendered from multiple perspectives. The formula for multi-perspective rendering is V = R(M, θ); where R(M, θ) is the rendering function, M is the three-dimensional model, θ is the observation angle, and V is the three-dimensional view returned by the rendering function of the three-dimensional model M observed from the angle θ.

[0194] Optionally, according to the color value, obtaining the three-dimensional model corresponding to the target object may include the following steps: first, according to the color value, obtaining the initial three-dimensional model, that is, according to the color value, obtaining the initial two-dimensional graphics of the target object at any viewing angle, that is, obtaining the initial three-dimensional model corresponding to the target object. Based on the initial three-dimensional model, extract the grid feature map corresponding to the initial three-dimensional model. Since the appearance accuracy of the initial three-dimensional model obtained by neural radiation field network rendering is low, there is impurity image information such as noise and artifacts in the initial three-dimensional model, resulting in low purity of the initial three-dimensional model. Since these impurity image information have a greater impact on the color information in the initial three-dimensional model and a smaller impact on the geometric information in the initial three-dimensional model, the grid feature map may include the geometric information of the initial three-dimensional model. Therefore, the grid feature map can be extracted and the network feature map can be colored to form an optimized three-dimensional model to reduce the impact of the impurity image information in the three-dimensional model on the three-dimensional model.

[0195] Next, the grid feature map is colored according to its color value to obtain a textured mesh. This textured mesh includes two-dimensional images of the target object from multiple perspectives. In other words, the textured mesh represents the three-dimensional model corresponding to the target object. Because the color value of each pixel in the initial three-dimensional model can be obtained through a point-by-point query, the shape of the extracted network feature map can be identical to that of the initial three-dimensional model. Multiple locations in the initial three-dimensional model can correspond one-to-one with multiple locations in the network feature map. Therefore, the color value of a location in the initial three-dimensional model can be identical to the color value of the corresponding location in the network feature map. Each pixel in the grid feature map can be colored according to the color value obtained by the neural radiation field network to obtain a colored textured mesh. This textured mesh can include two-dimensional images of the target object from multiple perspectives. For example, this textured mesh can include two-dimensional images of the target object from any perspective, and it contains less background information. In other words, the textured mesh represents the optimized three-dimensional model corresponding to the target object. Compared to the image impurities in the initial three-dimensional model before optimization, the optimized three-dimensional model contains less impurity information, thereby improving the quality of the three-dimensional model.

[0196] Optionally, extracting a mesh feature map corresponding to the initial three-dimensional model based on the initial three-dimensional model may include: extracting an initial mesh feature map based on the initial three-dimensional model. The mesh feature map is a polygonal mesh, a data structure used in computer graphics to model various irregular objects. The mesh feature map comprises vertices, edges, and faces. The geometric shape of the initial mesh feature map differs significantly from the geometric shape of the actual three-dimensional model.

[0197] The neural radiation field network may be used to iteratively optimize the initial mesh feature map a preset number of times to obtain the mesh feature map. The preset number of times may be greater than or equal to 3. The vertices, edges, and faces in the initial mesh feature map may be optimized so that the geometry of the initial mesh feature approximates the geometry of the three-dimensional model.

[0198] Step 207: Convert the three-dimensional model from the time domain to the frequency domain through fast Fourier transform to obtain a target three-dimensional model.

[0199] When a user suspects a 3D model has been illegally copied or distributed, they can download or purchase the model. Because the frequency of the watermark information differs from the frequency of the 3D model, the acquired 3D model can first be subjected to a Fast Fourier Transform (FFT) to convert from the time domain to the frequency domain, yielding the target 3D model.

[0200] Step 208: Perform frequency detection on the target three-dimensional model to determine whether the target three-dimensional model has a frequency corresponding to the watermark information.

[0201] The frequency of the watermark information can be determined first, and the frequency detection of the target three-dimensional model can be performed based on the frequency of the watermark information. If the frequency of the watermark information does not exist in the target three-dimensional model, it means that the target three-dimensional model is not a three-dimensional model with watermark information added, and there is no need to implement subsequent steps.

[0202] Step 209: In response to detecting a frequency corresponding to watermark information in the target three-dimensional model, extract the watermark information from the target three-dimensional model.

[0203] If the frequency of the watermark information exists in the target 3D model, it means that the target 3D model is a 3D model with watermark information added. The watermark information can be extracted from the target 3D model to further determine the content of the watermark information.

[0204] In an optional implementation, extracting watermark information from a target three-dimensional model may include the following two steps: first, separating the target watermark information from the target three-dimensional model based on the frequency corresponding to the watermark information; second, converting the target watermark information from the frequency domain to the time domain through an inverse Fourier transform to obtain the watermark information.

[0205] When separating the target watermark information from the target three-dimensional model, the target three-dimensional model may be compared with a three-dimensional model stored by the user in the terminal to extract the watermark signal from the target three-dimensional model.

[0206] Exemplarily, the watermark extraction formula can be W e =E(V); where the extraction function is E(V), V is the target 3D model, and W eThe watermark signal extracted from the target 3D model is returned by the extraction function.

[0207] Step 210: Decode the watermark information to restore the original watermark information.

[0208] Since the watermark information extracted from the target three-dimensional model is encoded information, the watermark information can be inversely encoded to restore the original watermark information, and the initial watermark information can be binary watermark information.

[0209] Step 211: Convert the initial watermark information into a text format to obtain the identity information corresponding to the watermark information.

[0210] The identity information corresponding to the watermark information may include a digital ID, a character string, or other forms of user identifiers.

[0211] For example, the conversion formula may be U'=D dec (W e );D dec (W e ) is the conversion function, and U' is the identity information returned by the conversion function.

[0212] Optionally, after obtaining the identity information corresponding to the watermark, the watermark can be verified and traced. Because each user's watermark is unique, the one-to-one correspondence between the watermark and the identity information can be used to track the creator of the 3D model, ensuring copyright protection for the 3D model and providing strong evidence for legal proceedings and rights protection.

[0213] After obtaining the identity information of the creator of the 3D model, the creator can be informed that the 3D model has been illegally copied or distributed through notification methods such as email or text message.

[0214] Optionally, the initial watermark information is converted into a text format, and the identity information in the text format corresponding to the watermark information can be searched in a preset database. The database stores the corresponding relationship between each user's watermark information and his / her identity information.

[0215] In summary, the embodiments of the present application provide a model watermark embedding method. This method uses a neural radiation field network to obtain density prediction values ​​and color prediction values ​​of a target object in three-dimensional space. Watermark information is then embedded into the color prediction values ​​to obtain watermarked color feature values. Volume rendering is then performed on the density prediction values ​​and color feature values ​​to obtain the color values ​​of the target object in three-dimensional space. Based on the color values, a two-dimensional image of the target object at any viewing angle is obtained, thereby obtaining a corresponding three-dimensional model of the target object. In this way, watermark information can be embedded into the three-dimensional model of the target object, resolving the difficulty of copyright protection for three-dimensional models in related technologies and achieving the copyright protection of the generated three-dimensional model.

[0216] In addition, since the density prediction value is data related to the geometric features of the three-dimensional model of the target object, if the watermark information is embedded in the density prediction value, it is more likely to have a greater impact on the subsequently generated three-dimensional model. This application can reduce the impact of the watermark information on the three-dimensional model by only adding watermark information to the color prediction value.

[0217] As shown in Figure 8, Figure 8 is a flow chart of another model watermark embedding method provided by the present application, which can add watermark information during the three-dimensional model reconstruction process, and the technology of adding watermark information can also be applied to the process of model conversion and large model Wensheng drawing. In the process of three-dimensional reconstruction, multiple target object images can be first obtained through AR assisted shooting, and the watermark information corresponding to user 1 is obtained in the database according to the user information, and the color prediction value is obtained through the NeRF network. The watermark information and the color prediction value are then fused to obtain a watermarked color feature value, and the watermarked color feature value and the predicted density value are used as the image data after color channel encryption, and a watermarked multi-perspective two-dimensional image is generated based on the color channel encrypted image data. Exemplarily, the watermarked multi-perspective two-dimensional image can include a watermarked first-perspective two-dimensional image, a watermarked second-perspective two-dimensional image, and a watermarked third-perspective two-dimensional image.

[0218] During the text generation and conversion model process, a two-dimensional image of the target object can be generated from the text content entered by the user, or the target object image can be directly obtained and the watermark information corresponding to user 1 can be embedded into the pixel values ​​of the target object image to obtain a watermarked target object image. This watermarked target object image can then be used to generate a watermarked multi-view image through the conversion model. Both the watermarked multi-view two-dimensional image and the watermarked multi-view image can be displayed on the user's terminal device, thus achieving an end-to-end watermarking method.

[0219] FIG9 is a structural block diagram of a model watermark embedding device provided in an embodiment of the present application. The model watermark embedding device 900 includes:

[0220] The acquisition module 901 is configured to acquire an image dataset and watermark information, where the image dataset includes image information of at least one two-dimensional picture of a target object.

[0221] The radiation module 902 is used to process the image data set through a neural radiation field network to obtain a density prediction value and a color prediction value of the target object in three-dimensional space.

[0222] The fusion module 903 is used to fuse the watermark information and the color prediction value using steganography to obtain the watermarked color feature value.

[0223] The rendering module 904 is used to perform volume rendering processing on the density prediction value and the color feature value to obtain the color value of the target object in the three-dimensional space.

[0224] The integration module 905 is used to obtain a three-dimensional model corresponding to the target object according to the color value.

[0225] Optionally, the fusion module includes:

[0226] The convolution unit is used to fuse the watermark information and the color prediction value through a convolutional neural network to obtain the watermarked color feature value.

[0227] Alternatively, the adversarial unit is used to fuse the watermark information and the color prediction value through a generative adversarial network to obtain a watermarked color feature value.

[0228] In addition, an embodiment of the present application also provides a model watermark embedding device, which includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the model watermark embedding method as in any of the above embodiments.

[0229] In addition, an embodiment of the present application also provides a computer storage medium, which stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by a processor to implement the model watermark embedding method as in any of the above embodiments.

[0230] In addition, embodiments of the present application further provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the model watermark embedding method of any of the above embodiments.

[0231] In this application, the terms "first", "second" and "third" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. The term "plurality" refers to two or more than two, unless expressly limited otherwise.

[0232] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0233] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0234] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0235] The above description is merely an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A model watermark embedding method, characterized in that: The method comprises: Acquire an image dataset and watermark information, wherein the image dataset includes image information of at least one two-dimensional picture of a target object; Processing the image data set through a neural radiation field network to obtain a density prediction value and a color prediction value of the target object in three-dimensional space; Using steganography to fuse the watermark information and the color prediction value to obtain a watermarked color feature value; Performing volume rendering processing on the density prediction value and the color feature value to obtain a color value of the target object in a three-dimensional space; A three-dimensional model corresponding to the target object is obtained according to the color value.

2. The method according to claim 1, characterized in that The method of fusing the watermark information and the color prediction value by using steganography to obtain a watermarked color feature value includes: The watermark information and the color prediction value are fused through a convolutional neural network to obtain the watermarked color feature value; Alternatively, the watermark information and the color prediction value are fused through a generative adversarial network to obtain the watermarked color feature value.

3. The method according to claim 1, characterized in that Performing volume rendering processing on the density prediction value and the color prediction value to obtain a color value of the target object in a three-dimensional space includes: The color value is obtained by a volume rendering formula based on the camera's viewing angle, and the volume rendering formula includes: Wherein, C(x, d) is the color observed from the position x of the camera along the direction d, that is, the color value, σ(p) is the density prediction value of position p, c(p) is the watermarked color feature value of position p, T(t) is the transmittance from the position of the camera to position p, tfar is the far boundary of the preset three-dimensional scene, tnear is the near boundary of the preset three-dimensional scene, the near boundary is located on the side of the far boundary close to the position of the camera, and the preset three-dimensional scene includes the target object.

4. The method according to claim 1, wherein The watermark information S(x), the color prediction value M(p) and the watermarked color feature value c(p) satisfy an embedding formula, and the embedding formula includes: c(p)=M(p)+α*S(x); Among them, α is the preset embedding strength parameter.

5. The method according to claim 1, wherein The image dataset includes image information of at least two two-dimensional images of a target object, wherein the image information of the two-dimensional images includes three-dimensional coordinates, pose information, and a target feature map; The acquiring of the image data set comprises: Acquire N target object images, where the N target object images are two-dimensional images obtained by capturing images of the same target object at different angles, where N is an integer greater than or equal to 2; Based on the N target object images, obtaining three-dimensional coordinates, posture information, and N target feature maps corresponding to the target objects, where the target feature maps are images obtained by segmenting the target object images; Based on the three-dimensional coordinates, the pose information, and the target feature map, image information of the two-dimensional image is obtained to acquire an image dataset.

6. The method according to claim 5, characterized in that The acquiring of the three-dimensional coordinates, posture information, and target feature map corresponding to the target object based on the N target object images includes: Extracting multiple feature points from each of the N target object images through a feature detection network; Obtaining a feature descriptor for each feature point; Matching the feature points of the N target object images based on the feature descriptors to obtain geometric relationships between the matched feature points in the N target object images; According to the geometric relationship, obtaining the posture information; Acquire three-dimensional coordinates of a plurality of the feature points according to the geometric relationship and the posture information; Acquiring the three-dimensional coordinates of the target object based on the three-dimensional coordinates of the multiple feature points; The N target object images are input into a convolutional neural network for image segmentation processing to segment the target feature map.

7. The method according to claim 1, characterized in that The acquiring of the image data set comprises: Acquire a target object picture, wherein the target object picture is a picture containing an image of the target object; Based on the target object image, obtaining N converted images, where the N converted images are two-dimensional images of different perspectives generated after processing the target object image, where N is an integer greater than or equal to 2; Based on the N converted images, obtaining three-dimensional coordinates, posture information, and N target feature maps corresponding to the target object, where the target feature maps are images obtained by segmenting the converted images; Based on the three-dimensional coordinates, the pose information, and the target feature map, image information of the two-dimensional image is obtained to acquire an image dataset.

8. The method according to claim 7, characterized in that The step of obtaining N converted images based on the target object image includes: Performing depth estimation on the target object image to obtain a depth map corresponding to the target object image; Extracting key feature points from the target object image through a feature detection network; Based on the depth map and the key feature points, N converted images corresponding to the target object image are obtained.

9. The method according to claim 7, characterized in that The step of obtaining a target object image includes: Obtaining text information describing the target object; The text information is converted into the target object image through a deep learning network.

10. The method according to claim 1, characterized in that Obtaining the watermark information includes: Get the user's unique identity information; Converting the identity information into a binary sequence to form initial watermark information in a binary format; The initial watermark information is subjected to spread spectrum processing to obtain the encoded watermark information.

11. The method according to claim 1, wherein After obtaining the three-dimensional model corresponding to the target object, the method further includes: Converting the three-dimensional model from the time domain to the frequency domain by fast Fourier transform to obtain a target three-dimensional model; Performing frequency detection on the target three-dimensional model to determine whether the target three-dimensional model has a frequency corresponding to the watermark information, wherein the frequency of the watermark information is different from the frequency of the three-dimensional model; In response to detecting a frequency corresponding to the watermark information in the target three-dimensional model, the watermark information is extracted from the target three-dimensional model.

12. The method according to claim 11, characterized in that The extracting the watermark information from the target three-dimensional model includes: separating target watermark information from the target three-dimensional model according to the frequency corresponding to the watermark information; The target watermark information is converted from the frequency domain to the time domain by inverse Fourier transform to obtain the watermark information.

13. The method according to claim 11, characterized in that In response to detecting the frequency corresponding to the watermark information in the target three-dimensional model, extracting the watermark information from the target three-dimensional model, the method further includes: performing inverse encoding on the watermark information to restore the original watermark information; The initial watermark information is converted into a text format to obtain the identity information corresponding to the watermark information.

14. The method according to claim 1, wherein The step of obtaining a three-dimensional model corresponding to the target object according to the color value includes: Obtaining an initial three-dimensional model according to the color value; Based on the initial three-dimensional model, extracting a grid feature map corresponding to the initial three-dimensional model; The grid feature map is colored according to the color value to obtain a texture grid, wherein the texture grid includes two-dimensional images of the target object under multiple viewing angles, and the texture grid is a three-dimensional model corresponding to the target object.

15. The method according to claim 14, characterized in that The extracting, based on the initial three-dimensional model, a grid feature map corresponding to the three-dimensional model includes: Extracting an initial grid feature map based on the initial three-dimensional model; The initial grid feature map is iteratively optimized a preset number of times through a neural radiation field network to obtain the grid feature map.

16. A model watermark embedding device, characterized in that: The model watermark embedding device comprises: an acquisition module, configured to acquire an image data set and watermark information, wherein the image data set includes image information of at least one two-dimensional picture of a target object; a radiation module, configured to process the image data set through a neural radiation field network to obtain a density prediction value and a color prediction value of the target object in three-dimensional space; A fusion module, configured to fuse the watermark information and the color prediction value using steganography to obtain a watermarked color feature value; a rendering module, configured to perform volume rendering processing on the density prediction value and the color feature value to obtain a color value of the target object in a three-dimensional space; The integration module is used to obtain a three-dimensional model corresponding to the target object according to the color value.

17. The model watermark embedding device according to claim 16, characterized in that: The fusion module includes: A convolution unit, configured to fuse the watermark information and the color prediction value through a convolutional neural network to obtain the watermarked color feature value; Alternatively, the adversarial unit is configured to fuse the watermark information and the color prediction value through a generative adversarial network to obtain the watermarked color feature value.

18. A model watermark embedding device, characterized in that: The model watermark embedding device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, the at least one program, the code set or instruction set is loaded and executed by the processor to implement the model watermark embedding method according to any one of claims 1 to 13.

19. A computer storage medium, characterized in that The computer storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the model watermark embedding method according to any one of claims 1 to 13.

Images