Panoramic watermark processing method, device and equipment of three-dimensional Gaussian scene and medium

Abstract

The present application relates to the technical field of three-dimensional Gaussian scene watermarking, and particularly relates to a panoramic watermark processing method, device and equipment for a three-dimensional Gaussian scene and a medium, the method comprising: determining parameter sensitivity of each first Gaussian unit according to a difference degree value between a two-dimensional original image and a first rendering image under each first view angle corresponding to each first Gaussian unit; selecting a second Gaussian unit corresponding to parameter sensitivity greater than a preset parameter sensitivity threshold from all first Gaussian units; and embedding a copyright message in the second Gaussian unit in the three-dimensional Gaussian scene, thereby completing embedding operation of panoramic watermarking. The present application estimates parameter sensitivity of the first Gaussian unit through two-dimensional original images and first rendering images under different view angles, so as to consider geometric structural characteristics of the three-dimensional scene when embedding the watermark, and avoid watermark information distortion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional Gaussian scene watermarking, in particular to a panoramic watermark processing method and device for a three-dimensional Gaussian scene, equipment and a medium. BACKGROUND

[0002] As a breakthrough technology in the field of three-dimensional computer vision, the three-dimensional Gaussian (3D Gaussian Splatting, 3DGS) scene constructs an explicit geometry and appearance model of the scene by discretizing a set of three-dimensional Gaussian units, realizing the innovation of traditional three-dimensional scene representation and rendering methods. However, the copyright protection of three-dimensional Gaussian scene assets faces severe challenges: due to the discrete and easy-to-edit nature of three-dimensional Gaussian parameters, related digital assets are extremely vulnerable to unauthorized tampering, and the tampered assets are difficult to trace the original ownership.

[0003] The existing two-dimensional watermarking technology does not take into account the geometric structure characteristics of three-dimensional scenes, and in the application process, once the view angle is switched or the scene is locally edited, the watermark information is easily distorted, and the adaptability of the three-dimensional Gaussian scene assets is severely limited. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a panoramic watermark processing method and device for a three-dimensional Gaussian scene, which can embed watermarks in a three-dimensional Gaussian scene based on different viewing angles to avoid watermark information distortion.

[0005] In a first aspect, the present application provides a panoramic watermark processing method for a three-dimensional Gaussian scene, the method comprising:

[0006] obtaining a three-dimensional Gaussian scene to be processed, a two-dimensional original image corresponding to the three-dimensional Gaussian scene at each first viewing angle, and a first rendered image corresponding to the three-dimensional Gaussian scene at each first viewing angle; each first rendered image is a two-dimensional image obtained by image rendering of a first Gaussian unit in the three-dimensional Gaussian scene at the corresponding first viewing angle;

[0007] determining the parameter sensitivity of each first Gaussian unit according to the difference degree value between the two-dimensional original image and the first rendered image at each first viewing angle in each first Gaussian unit; the parameter sensitivity is used to quantify the influence of the parameter disturbance introduced by embedding watermarks at the corresponding first Gaussian unit on the geometric structure integrity and rendering visual consistency of the three-dimensional Gaussian scene;

[0008] from all first Gaussian units, selecting a second Gaussian unit corresponding to a parameter sensitivity greater than a preset parameter sensitivity threshold;

[0009] Embedding the copyright message in the second Gaussian unit in the three-dimensional Gaussian scene, so as to complete the embedding operation of the panorama watermark.

[0010] In a possible implementation, the parameter sensitivity of any first Gaussian unit is calculated according to the following steps, comprising:

[0011] For each first view angle in which the first Gaussian unit is located, a parameter sensitivity matrix of the first Gaussian unit in the first view angle is calculated according to a difference degree value between the two-dimensional original image and the first rendered image in the first view angle.

[0012] The parameter sensitivity matrices of the first Gaussian unit in all corresponding first view angles are fused to obtain the parameter sensitivity of the first Gaussian unit.

[0013] In a possible implementation, the parameter sensitivity matrix of the first Gaussian unit in the first view angle is calculated according to the difference degree value between the two-dimensional original image and the first rendered image in the first view angle, comprising:

[0014] The difference degree value between the two-dimensional original image and the first rendered image in the first view angle is substituted into the following formula to obtain the parameter sensitivity matrix of the first Gaussian unit in the first view angle.

[0015] ;

[0016] wherein, is the parameter sensitivity matrix of the first Gaussian unit in the first view angle; is the kth first view angle; is the parameter set of the ith first Gaussian unit; is the difference degree value between the two-dimensional original image and the first rendered image in the first view angle . is the difference degree value .is the first-order gradient vector of the parameter set of the ith first Gaussian unit . is the transpose.

[0017] In a possible implementation, the parameter sensitivity of the first Gaussian unit is obtained by fusing the parameter sensitivity matrices of the first Gaussian unit in all corresponding first view angles, comprising:

[0018] The parameter sensitivity matrices of the first Gaussian unit in all corresponding first view angles are substituted into the following formula to obtain the parameter sensitivity of the first Gaussian unit.

[0019] ;

[0020] wherein, is a parameter sensitivity of the i-th first Gaussian cell, is a number of first view angles in which the i-th first Gaussian cell is located, is a trace of a matrix, is a parameter sensitivity matrix of the first Gaussian cell under a first view angle, is the k-th first view angle, is a parameter set of the i-th first Gaussian cell.

[0021] In a possible implementation, after embedding the copyright message in the second Gaussian cell in the three-dimensional Gaussian scene, the method further comprises:

[0022] inputting each third Gaussian cell parameter in a target three-dimensional Gaussian scene into a three-dimensional watermark extraction model to obtain a first copyright extraction message; the target three-dimensional Gaussian scene is the three-dimensional Gaussian scene after the copyright message is embedded;

[0023] and / or, inputting all second rendered images corresponding to the target three-dimensional Gaussian scene into a two-dimensional watermark extraction model to obtain a second copyright extraction message; each second rendered image is a two-dimensional image obtained by image rendering on a fourth Gaussian cell under a corresponding second view angle in the target three-dimensional Gaussian scene;

[0024] determining a watermark verification result corresponding to the target three-dimensional Gaussian scene according to a matching degree between the first copyright extraction message and the embedded copyright message, and / or a matching degree between the second copyright extraction message and the embedded copyright message.

[0025] In a possible implementation, the inputting each third Gaussian cell parameter in a target three-dimensional Gaussian scene into a three-dimensional watermark extraction model to obtain a first copyright extraction message comprises:

[0026] inputting each third Gaussian cell parameter into a feature extraction module in the three-dimensional watermark extraction model respectively to obtain a local feature vector corresponding to each third Gaussian cell; the local feature vector corresponding to each third Gaussian cell is used to represent an association relationship between parameters inside the third Gaussian cell, and / or an association relationship between the parameter of the third Gaussian cell and the parameter of a third Gaussian cell adjacent to the third Gaussian cell;

[0027] inputting all local feature vectors corresponding to all third Gaussian cells into a feature aggregation module in the three-dimensional watermark extraction model to obtain a first global feature vector corresponding to the target three-dimensional Gaussian scene;

[0028] input the first global feature vector into a first copyright message decoding module in the three-dimensional watermark extraction model to obtain a first copyright message.

[0029] In a possible implementation, the inputting the second global feature vector into a second copyright message decoding module in the two-dimensional watermark extraction model to obtain a second copyright message comprises:

[0030] input each second rendered image into a multi-scale feature extraction layer in the two-dimensional watermark extraction model to obtain a first image feature map of each second rendered image at each preset scale;

[0031] for each second rendered image, input the first image feature map of the second rendered image at all preset scales into a first feature fusion layer in the two-dimensional watermark extraction model to obtain a second image feature vector corresponding to the second rendered image;

[0032] input the second image feature vectors corresponding to all second rendered images into a second feature fusion layer in the two-dimensional watermark extraction model to obtain a second global feature vector corresponding to the target three-dimensional Gaussian scene;

[0033] input the second global feature vector into a second copyright message decoding module in the two-dimensional watermark extraction model to obtain a second copyright message.

[0034] In a second aspect, the embodiments of the present application further provide a panoramic watermark processing device for a three-dimensional Gaussian scene, and the device comprises:

[0035] a obtaining module, configured to obtain a three-dimensional Gaussian scene to be processed, a two-dimensional original image corresponding to the three-dimensional Gaussian scene at each first view angle, and a first rendered image corresponding to the three-dimensional Gaussian scene at each first view angle; each first rendered image is a two-dimensional image obtained by image rendering on a first Gaussian unit corresponding to the first view angle in the three-dimensional Gaussian scene;

[0036] a sensitivity determining module, configured to determine a parameter sensitivity of each first Gaussian unit according to a difference degree value between the two-dimensional original image and the first rendered image corresponding to each first view angle at each first Gaussian unit; the parameter sensitivity is used to quantify an influence degree of a parameter disturbance introduced by embedding a watermark at the corresponding first Gaussian unit on the geometric structure integrity and the rendering visual consistency of the three-dimensional Gaussian scene;

[0037] a Gaussian unit screening module, configured to screen a second Gaussian unit corresponding to a parameter sensitivity greater than a preset parameter sensitivity threshold from all first Gaussian units;

[0038] The watermark embedding module is configured to embed a copyright message in the second Gaussian unit in the three-dimensional Gaussian scene, so as to complete the embedding operation of the panorama watermark.

[0039] In a third aspect, the embodiments of the present application further provide an electronic device, comprising a processor, a storage medium and a bus, the storage medium stores machine readable instructions executable by the processor, when the electronic device is running, the processor communicates with the storage medium through the bus, and the processor executes the machine readable instructions to perform the steps of the panorama watermark processing method of the three-dimensional Gaussian scene according to any one of the first aspect.

[0040] In a fourth aspect, the embodiments of the present application further provide a computer readable storage medium, the computer readable storage medium stores a computer program, when the computer program is run by a processor, the steps of the panorama watermark processing method of the three-dimensional Gaussian scene according to any one of the first aspect are executed.

[0041] The embodiments of the present application provide a panorama watermark processing method, device and equipment of a three-dimensional Gaussian scene and a medium, the method comprising: acquiring a three-dimensional Gaussian scene to be processed, a two-dimensional original image at each first view angle for constructing a three-dimensional Gaussian scene, and a first rendering image at each first view angle corresponding to the three-dimensional Gaussian scene; determining the parameter sensitivity of each first Gaussian unit according to the difference degree value between the two-dimensional original image and the first rendering image at each first view angle corresponding to each first Gaussian unit; screening a second Gaussian unit corresponding to a parameter sensitivity greater than a preset parameter sensitivity threshold from all first Gaussian units; and embedding a copyright message in the second Gaussian unit in the three-dimensional Gaussian scene, thereby completing the embedding operation of the panorama watermark. The embodiments of the present application estimate the parameter sensitivity of the first Gaussian unit through the two-dimensional original image and the first rendering image at different view angles, so as to consider the geometric structure characteristics of the three-dimensional scene when embedding the watermark, and avoid the distortion of the watermark information. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be considered as limiting the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0043] Figure 1 A flowchart of a panorama watermark processing method of a three-dimensional Gaussian scene provided by the embodiments of the present application is shown;

[0044] Figure 2 A flowchart of watermark extraction verification provided by the embodiments of the present application is shown;

[0045] Figure 3 Fig. 1 shows a structural schematic diagram of a panoramic watermark processing device of a three-dimensional Gaussian scene according to an embodiment of the present application.

[0046] Figure 4 Fig. 2 shows a structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of description and illustration, and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn according to the actual proportions. The flowchart shows the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowchart can not be implemented in sequence, and the steps without logical context relationship can be reversed in sequence or implemented simultaneously. In addition, one or more other operations can be added to the flowchart or removed from the flowchart under the guidance of the content of the present application by those skilled in the art.

[0048] In addition, the described embodiments are only some of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0049] In order to enable those skilled in the art to use the content of the present application, the following implementation is given in combination with the specific application scenario "three-dimensional Gaussian scene watermark technology field". For those skilled in the art, the general principles defined herein can be applied to other embodiments and application scenarios without departing from the spirit and scope of the present application. Although the present application is mainly described in relation to the "three-dimensional Gaussian scene watermark technology field", it should be understood that this is only an exemplary embodiment.

[0050] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0051] The watermark processing method provided by the embodiments of the present application will be described in detail below.

[0052] Reference Figure 1The diagram shown is a flowchart illustrating a panoramic watermarking method for a 3D Gaussian scene according to an embodiment of this application. The exemplary steps of this embodiment are described below:

[0053] S101. Obtain the three-dimensional Gaussian scene to be processed, the two-dimensional original images of the three-dimensional Gaussian scene in each first-view perspective, and the first rendered images of the three-dimensional Gaussian scene in each first-view perspective.

[0054] In this embodiment, images of the target scene are acquired from various original viewpoints (ensuring full viewpoint coverage (surround scene or multi-angle shooting)) to obtain two-dimensional original images. Using 3D Gaussian sputtering technology, a three-dimensional visualization scene is constructed based on the two-dimensional original images from each viewpoint, resulting in the three-dimensional Gaussian scene to be processed (3D Gaussian scene). Randomly selected from all original viewpoints. A representative perspective is used to obtain the first-person perspective, denoted as... For each first-person perspective The first rendered image is generated by Gaussian splash rendering. ; where the color of each pixel in the first rendered image It is obtained by weighted superposition of the colors of Gaussian units after depth sorting; that is, each first rendered image is a two-dimensional image obtained by rendering the first Gaussian unit in the corresponding first view in the three-dimensional Gaussian scene.

[0055] Among them, the viewpoint refers to the camera shooting angle corresponding to the original multi-view image used to reconstruct the 3D Gaussian scene, which is one of the core components of the input data during the 3DGS training process. This viewpoint contains two parts of information: (1) camera pose: that is, the position (translation) and orientation (rotation) of the camera in the world coordinate system when the image is captured; (2) camera parameters: such as focal length, distortion coefficient, etc. (which determine the projection relationship of the image).

[0056] Here, a 3D Gaussian scene refers to a scene representation model composed of multiple independent three-dimensional Gaussian units (3D Gaussian elements). Each 3D Gaussian element is described by a complete set of parameters that define its geometric and optical properties. .in, Represents the 3D center coordinates of the Gaussian element, defining its specific position in the world coordinate system; It is an orthogonal rotation matrix used to control the orientation of the Gaussian element and ensure that the rotation operation does not change the size of the Gaussian element; The diagonal scaling matrix controls the scale of the Gaussian element along the x, y, and z axes, respectively. The spherical harmonic (SH) coefficient set is usually a 3-order SH function, which contains 16 coefficients, and is used to accurately represent the color response of the Gaussian unit under different lighting conditions, to ensure the color fidelity of the rendered image. The transparency parameter controls the superposition weight of the Gaussian unit in the rendering process. The lower the transparency, the greater the contribution of the Gaussian unit to the final rendering color. The 3D Gaussian covariance matrix is a key indicator for describing the spatial distribution characteristics of the 3D Gaussian unit. The covariance matrix of each 3D Gaussian unit is calculated by a rotation matrix and a scaling matrix. The matrix is a positive definite matrix, which ensures the rationality of the Gaussian distribution, and the eigenvalues directly reflect the distribution range of the Gaussian unit in different directions, providing an important basis for Gaussian densification.

[0057] In addition, Gaussian Splatting Rendering is a core real-time rendering technology for 3D Gaussian scenes (3DGS), which is specially used to quickly convert 3D Gaussian primitives into 2D images, and is a key method to realize "million-level Gaussian units + photo-level details + real-time frame rate". The core principle is that Gaussian Splatting Rendering realizes efficient rendering through rasterization acceleration + Gaussian primitive projection superposition: (1) Gaussian primitive projection: project each 3D Gaussian unit (ellipsoid with position, shape, color, and transparency) onto the 2D pixel plane of the current camera, and calculate its coverage range on the pixel; (2) depth sorting: sort the projected Gaussian units by depth (distance from the camera), to ensure that the near Gaussians can correctly occlude the far Gaussians; (3) color superposition: superimpose the color (combined with transparency) of each Gaussian onto the corresponding pixel in the sorted order, to finally generate a 2D image.

[0058] S102, according to the difference degree value between the two-dimensional original image and the first rendering image under each first view angle corresponding to each first Gaussian unit, determine the parameter sensitivity of each first Gaussian unit.

[0059] In the embodiments of the present application, the first Gaussian unit refers to the Gaussian unit in the first view angle in the three-dimensional Gaussian scene. The parameter sensitivity is used to quantify the parameter disturbance introduced by embedding a watermark in the corresponding first Gaussian unit, and the influence degree on the geometric structure integrity and rendering visual consistency of the three-dimensional Gaussian scene. Specifically, the parameter sensitivity of any first Gaussian unit is calculated according to the following steps:

[0060] Step one, for each first view angle where the first Gaussian unit is located, according to the difference degree value between the two-dimensional original image and the first rendering image under the first view angle, calculate the parameter sensitivity matrix of the first Gaussian unit under the first view angle.

[0061] In this embodiment, sensitivity quantification is achieved through approximate calculation of the Hessian matrix (i.e., the parameter sensitivity matrix). The Hessian matrix, as a second-order derivative matrix, can more accurately reflect parameter sensitivity; however, directly calculating the second derivative is extremely complex (especially when the number of Gaussian units reaches millions). Approximating the Hessian matrix using the gradient outer product method can significantly improve computational efficiency while maintaining accuracy. Therefore, the parameter sensitivity matrix is... Defined as a loss function with respect to parameters The second-order partial derivative matrix. Therefore, by substituting the difference between the original two-dimensional image and the first rendered image from the first viewpoint into the following formula, the parameter sensitivity matrix of the first Gaussian unit under the first viewpoint is obtained;

[0062] ;

[0063] in, The parameter sensitivity matrix of the first Gaussian element under the first viewpoint; This is the k-th first-person perspective; Let be the parameter set of the i-th first Gaussian unit; For first-person perspective Below, the difference value between the original 2D image and the first rendered image; Values ​​representing the degree of difference The parameter set of the i-th first Gaussian unit The first-order gradient vector; This is a transpose.

[0064] Here, in the construction and training of the 3D Gaussian scene, the loss function affects the parameters. The loss function in the first-order gradient is primarily photometric loss, and in some scenarios, regularization loss is added to assist in optimization. Photometric loss This is the most important loss function during training, used to measure the pixel-level difference between the "image rendered by the 3D Gaussian model" and the "real input multi-view image". The above approximation method has high accuracy after the 3DGS parameter optimization converges, because at this time the parameters are close to the optimal value, and the second derivative of the loss function can be effectively approximated by the outer product of the first derivative. Therefore, the 3D Gaussian scene obtained in this embodiment has completed parameter optimization and convergence, that is, the rendering quality of the scene has reached the practical requirements (e.g., PSNR≥25dB, SSIM≥0.8), and the number of Gaussian units and parameter distribution of the scene have stabilized, so there is no need for large-scale parameter adjustment.

[0065] Wherein, PSNR (Peak Signal-to-Noise Ratio): an index for measuring the distortion degree of an image, the unit is decibel (dB), the higher the value is, the smaller the image distortion is, and the better the visual quality is. SSIM (Structural Similarity Index): an index for measuring the structural similarity degree of two images, the value range is [0, 1], the closer the value is to 1, the more consistent the structure, brightness and contrast of the two images are, and the smaller the visual difference is.

[0066] Step two, fuse the parameter sensitivity matrix of the first Gaussian unit under all corresponding first view angles to obtain the parameter sensitivity of the first Gaussian unit.

[0067] In the embodiment of the application, after obtaining the Hessian matrix of the parameter sensitivity of the first Gaussian unit under all corresponding first view angles, the global sensitivity of each first Gaussian unit needs to be further quantified. Specifically, the parameter sensitivity matrix of the first Gaussian unit under all corresponding first view angles is substituted into the following formula to obtain the parameter sensitivity of the first Gaussian unit.

[0068] ;

[0069] Wherein, is the parameter sensitivity of the i-th first Gaussian unit, is the number of the first view angle where the i-th first Gaussian unit is located, is the trace of the matrix (i.e. the sum of the diagonal elements of the matrix), which is used to convert the Hessian information in the form of a matrix into a scalar value, so as to facilitate subsequent threshold screening and comparison, is the parameter sensitivity matrix of the first Gaussian unit under the first view angle, is the k-th first view angle, is the parameter set of the i-th first Gaussian unit.

[0070] Here, the Hessian matrix trace of each first Gaussian unit under all first view angles is accumulated to obtain the global parameter sensitivity of the unit. The physical meaning of the matrix trace is that it can comprehensively reflect the overall size of the Hessian matrix, i.e. the second-order influence degree of the parameter on the loss function; the greater the trace value is, the higher the second-order sensitivity of the parameter is, and the stronger the tolerance to perturbation is.

[0071] S103, from all the first Gaussian units, screen the second Gaussian unit corresponding to the parameter sensitivity greater than the preset parameter sensitivity threshold.

[0072] In this embodiment of the application, sensitivity to all first Gaussian element parameters is considered. The sorting process typically selects the first Gaussian unit, which has the highest parameter sensitivity (5% to 10%), as the preset parameter sensitivity threshold for filtering high-tolerance units. Filter out those that meet the requirements The first Gaussian element is used as the second Gaussian element. These Gaussian units are mainly distributed in the background areas, transition areas between objects, and smooth surface areas of a 3D Gaussian scene. The Gaussian units in these areas have a significantly higher tolerance for parameter perturbations than in other areas. Even if their parameters are moderately modified or additional denser units are added, there will be no perceptible impact on the geometry and rendering quality of the 3D Gaussian scene, making them ideal areas for watermark embedding.

[0073] Here, parameter sensitivity is the core indicator characterizing the sensitivity of 3D Gaussian parameters to the rendering result, denoted as . . The magnitude of this value directly reflects the parameter's tolerance to external disturbances: A higher value indicates a larger posterior variance for the parameter during scene optimization, a stronger tolerance to external disturbances, and a smaller impact on rendering quality after watermark embedding; conversely, a lower value indicates a lower posterior variance. The lower the value, the more sensitive the parameter is to perturbations; even minor modifications can lead to significant distortion in the rendered image. Quantifying parameter sensitivity is the key foundation for achieving watermark stealth embedding in this invention. Its calculation method is based on the Laplacian approximation in Bayesian learning and the sensitivity estimation approach of the Hessian matrix.

[0074] S104. Embed the copyright message in the second Gaussian unit in the 3D Gaussian scene to complete the panoramic watermark embedding operation.

[0075] In this embodiment of the application, the copyright message M is encoded into the distribution characteristics of the dense Gaussian unit by generating a dense Gaussian unit. Then, the dense Gaussian unit is fused with the second Gaussian unit in the three-dimensional Gaussian scene to obtain the target Gaussian scene after embedding the panoramic watermark.

[0076] Specifically, the ownership identification information of the 3D Gaussian scene to be embedded is encoded using a binary encoding format to obtain the copyright message, denoted as M. The length of the copyright message is... The maximum supported length is 32 bits. Copyright messages typically contain core information such as the asset owner's identifier, asset creation timestamp, and authorized usage permissions, ensuring integrity during message transmission.

[0077] Furthermore, based on the aforementioned watermark embedding mechanism, this application embodiment also provides a watermark extraction and verification mechanism. (Refer to...)Figure 2 Fig. 2 shows a flowchart of a watermark extraction and verification method provided by an embodiment of the present application. Specifically, after embedding the copyright message in the second Gaussian cell in the three-dimensional Gaussian scene, the method further includes:

[0078] S201, inputting parameters of each third Gaussian cell in the target three-dimensional Gaussian scene into a three-dimensional watermark extraction model to obtain a first copyright message; the target three-dimensional Gaussian scene is the three-dimensional Gaussian scene after embedding the copyright message.

[0079] In the embodiment of the present application, the three-dimensional watermark extraction model directly extracts the copyright message from the target three-dimensional Gaussian scene with watermark. Since the number of Gaussian cells in the target three-dimensional Gaussian scene is usually large (several hundred thousand to several million), if all Gaussian cells are directly processed, the extraction efficiency will be extremely low, therefore a hierarchical sampling strategy is needed to randomly sample 10k-50k Gaussian cells from the target three-dimensional Gaussian scene to obtain the third Gaussian cell. The parameter of the third Gaussian cell is the parameter of the third Gaussian cell.

[0080] Here, the network structure design of the three-dimensional watermark extraction model adopts a scheme of processing 3D Gaussian cell parameters using a convolutional neural network architecture, and constructs a three-dimensional watermark extraction model . This architecture is specially designed for discrete 3D data such as point clouds, can effectively extract global features, and is very suitable for the parameter characteristics of 3D Gaussian cells. The three-dimensional watermark extraction model is divided into three stages: a feature extraction module, a feature aggregation module, and a first copyright message decoding module.

[0081] In addition, in order to improve the robustness of the copyright message extraction, a 3D anti-interference training is added before the input of the noise simulation data into the 3D watermark extraction model, common 3D asset tampering attacks are simulated, and it is ensured that the 3D watermark extraction model can still accurately extract the copyright message in the case of attack. The 3D anti-interference training mainly simulates four types of attacks: Gaussian noise attack, adding Gaussian noise with a standard deviation σ = 0.1 to the center position of the first sampling Gaussian unit, simulating noise interference in the parameter transmission process; rotation attack, applying random rotation around any axis with an angle range of ± π / 6 to all first sampling Gaussian units, simulating the scenario of 3D asset being rotated and tampered; translation attack, adding random translation with a range of [0, 1000] to the center position of all first sampling Gaussian units, simulating the scenario of 3D asset being moved as a whole; clipping attack, randomly removing 10% of the first sampling Gaussian units (clipping rate cr = 0.1), simulating the scenario of 3D asset being partially deleted. The output of the anti-interference layer is used as the input of the 3D watermark extraction model, and by incorporating these attack scenarios in the training process, the 3D watermark extraction model learns features with anti-interference ability, ensuring that even if the 3D asset is tampered with, the copyright message can still be accurately extracted in actual application. That is, the above-mentioned 3D watermark extraction model is trained in advance by using the 3D Gaussian scene samples after embedding the copyright message and the corresponding true copyright message. Among them, the first sampling Gaussian unit in the 3D Gaussian scene sample for inputting the corresponding parameters into the 3D watermark extraction model is a Gaussian unit after being attacked by at least one of the Gaussian noise attack, the rotation attack, the translation attack, and the clipping attack.

[0082] Specifically, according to the following steps, the parameters of each third Gaussian unit in the target 3D Gaussian scene are input into the 3D watermark extraction model to obtain the first copyright extraction message:

[0083] Step one, input the parameters of each third Gaussian unit into the feature extraction module in the 3D watermark extraction model respectively to obtain the local feature vector corresponding to each third Gaussian unit.

[0084] In the embodiments of the present application, the local feature vector corresponding to each third Gaussian unit is used to represent the correlation between the parameters inside the third Gaussian unit, and / or the correlation between the parameters of the third Gaussian unit and the parameters of its adjacent third Gaussian unit. For example, the correlation between the orthogonal rotation matrix and the diagonal scaling matrix inside the third Gaussian unit, whether the colors of adjacent third Gaussian units are continuous, etc.

[0085] Here, the feature extraction module is composed of three fully connected layers, and each third Gaussian cell parameter sequentially passes through the three fully connected layers, and the feature vector dimensions output by each fully connected layer are 64, 128, and 128, respectively, to extract local parameter correlation. The output of the last fully connected layer is determined as the final local feature vector.

[0086] Step two, input all the local feature vectors corresponding to the third Gaussian cells into the feature aggregation module in the three-dimensional watermark extraction model to obtain a first global feature vector corresponding to the target three-dimensional Gaussian scene.

[0087] In the embodiment of the present application, the local feature vectors corresponding to all the third Gaussian cells are aggregated by max-pooling to obtain a 128-dimensional first global feature vector.

[0088] Step three, input the first global feature vector into the first watermark decoding module in the three-dimensional watermark extraction model to obtain a first copyright removal message.

[0089] In the embodiment of the present application, the first global feature vector is input into a first fully connected layer to output a 96-dimensional first feature vector; then the 96-dimensional first feature vector is input into a second fully connected layer to be mapped to a first binary prediction vector with a length consistent with the embedded copyright message; finally, the first binary prediction vector is mapped to the interval [0, 1] by a Sigmoid activation function for output; and the output value greater than a preset threshold (0.5) is predicted as 1, otherwise as 0, to obtain a binary first copyright removal message.

[0090] S202, and / or, input all the second rendered images corresponding to the target three-dimensional Gaussian scene into a two-dimensional watermark extraction model to obtain a second copyright removal message; each second rendered image is a two-dimensional image obtained by image rendering on a fourth Gaussian cell in the target three-dimensional Gaussian scene under a corresponding second view angle.

[0091] In the embodiment of the present application, the two-dimensional watermark extraction model extracts the second copyright removal message from the second rendered image set of the target three-dimensional Gaussian scene ​The copyright message is extracted, and the design needs to fully consider the distortion characteristics of the two-dimensional image. Through image preprocessing, special network structure, anti-interference layer design and multi-view message fusion, the reliable extraction of the copyright message is realized. Among them, the second rendering image set refers to the target three-dimensional Gaussian scene, uniformly selecting multiple ring view angles (i.e. the second view angle covers 360° horizontal view angle and 90° vertical view angle) in the spherical coordinate system with the scene center as the origin, and generating a 2D rendering image set through Gaussian splashing technology. The second rendering image set includes the second rendering image. The resolution of each second rendering image can be set according to the actual application requirements, usually 1024x768 to 2048x1536, to ensure that the second rendering image can completely cover all key areas of the scene. The fourth Gaussian unit refers to the Gaussian unit in the target three-dimensional Gaussian scene.

[0092] In addition, in order to improve the robustness of the two-dimensional watermark extraction model, 2D anti-interference data is added during image preprocessing to simulate common 2D image tampering attacks, mainly including four types of attacks: JPEG compression attack, using quality factor of JPEG compression (JPEG compression is the most common distortion method in image transmission, and quality factor 50 corresponds to moderate compression degree, which will introduce certain block effect); scaling attack, scaling the image to 25% of the original size and then enlarging it to the original size through bilinear interpolation, simulating the scenario of image scaling tampering; Gaussian blur attack, using a 3x3 Gaussian blur kernel with a standard deviation of 0.1 to blur the image, simulating the blur distortion in the image transmission process; noise attack, adding Gaussian noise with a standard deviation of 0.05, simulating image sensor noise or transmission noise. By incorporating these attack scenarios during training, the decoder learns image features with anti-interference ability, ensuring that even if the 2D rendering image is tampered with, the copyright message can still be accurately extracted. That is, the above two-dimensional watermark extraction model is trained by the rendering sample image corresponding to the three-dimensional Gaussian scene sample after embedding the copyright message (the same as the acquisition method of the second rendering image) and the real copyright message. Among them, the rendering sample image is an image attacked by at least one of the JPEG compression attack, scaling attack, Gaussian blur attack and noise attack.

[0093] Specifically, according to the following steps, all second rendering images corresponding to the target three-dimensional Gaussian scene are input into the two-dimensional watermark extraction model to obtain the second copyright extraction message, including:

[0094] Step one, input each second rendering image into the multi-scale feature extraction layer in the two-dimensional watermark extraction model to obtain the first image feature map of each second rendering image at each preset scale.

[0095] In the embodiment of the present application, the multi-scale feature extraction layer includes 4 convolutional blocks, each of which is used to extract a first image feature map at a corresponding preset scale to capture multi-scale pixel information. The preset scales include 512x384x64, 256x192x128, 128x96x256, and 32x24x512.

[0096] Step two, for each second rendered image, input the first image feature map of the second rendered image at all preset scales into the first feature fusion layer in the two-dimensional watermark extraction model to obtain a second image feature vector corresponding to the second rendered image.

[0097] In the embodiment of the present application, the first image feature map of the second rendered image at all preset scales is subjected to average pooling and compressed into a 512-dimensional global feature of a single-view rendered image to obtain the second image feature vector. The second image feature vector can comprehensively reflect the overall features of the second rendered image, including the subtle feature differences formed by the watermark Gaussian in the rendered image.

[0098] Step three, input the second image feature map corresponding to all second rendered images into the second feature fusion layer in the two-dimensional watermark extraction model to obtain a second global feature vector corresponding to the target three-dimensional Gaussian scene.

[0099] Step four, input the second global feature vector into the second copyright message decoding module in the two-dimensional watermark extraction model to obtain a second copyright removal message.

[0100] In the embodiment of the present application, the second global feature vector is input into a first fully connected layer to output a 96-dimensional second feature vector; then the 96-dimensional second feature vector is input into a second fully connected layer to be mapped to a second binary prediction vector with a length consistent with the embedded copyright message; finally, the second binary prediction vector is mapped to the interval [0, 1] through a Sigmoid activation function for output; and the output value greater than a preset threshold (0.5) is predicted as 1, otherwise as 0, to obtain a binary second copyright removal message.

[0101] Here, the core goal of the watermark extraction model is to reliably extract the copyright message from the watermarked 3D Gaussian scene and its second rendered image set, supporting 3D and 2D dual-end verification to cope with different application scenarios - the three-dimensional watermark extraction model is suitable for the scene where the 3D asset has not been rendered, only the Gaussian parameters are retained (such as 3D asset trading, cloud storage verification), and the two-dimensional watermark extraction model is suitable for the scene where the 3D asset has been rendered into a 2D image and the 3D parameters cannot be obtained (such as social media dissemination, offline display verification). By designing a special three-dimensional watermark extraction model and a two-dimensional watermark extraction model, combined with the design of an anti-interference layer, it is ensured that the copyright message can still be accurately extracted in the presence of tampering attacks.

[0102] S203, according to the matching degree between the first copyright extraction message and the embedded copyright message, and / or the matching degree between the second copyright extraction message and the embedded copyright message, determine the watermark verification result corresponding to the target three-dimensional Gaussian scene.

[0103] In the embodiments of the present application, if the matching degree is greater than a preset matching threshold (such as 95%), the watermark verification result corresponding to the target three-dimensional Gaussian scene is verified, otherwise it is verified.

[0104] When determining the watermark verification result corresponding to the target three-dimensional Gaussian scene according to the matching degree between the first copyright extraction message and the embedded copyright message, and the matching degree between the second copyright extraction message and the embedded copyright message, the two matching degrees can be weighted and fused or averaged and fused to determine the target matching degree, and then compared with the preset matching threshold to obtain the watermark verification result corresponding to the target three-dimensional Gaussian scene; it can also be considered that the watermark verification result corresponding to the target three-dimensional Gaussian scene is verified when any one of the matching degrees is greater than the preset matching threshold, otherwise it is verified.

[0105] Here, the calculation expression of the matching degree is: , wherein is an indicator function, which takes the value 1 when the condition in the parentheses is true, otherwise it takes the value 0. is the first copyright extraction message or the second copyright extraction message is the value of the tthbit is the embedded copyright message is the value of the tthbit of the embedded copyright message. is the length of the embedded copyright message

[0106] ​Based on the same inventive concept, the embodiment of the present application also provides a panorama watermark processing device for a three-dimensional Gaussian scene, which corresponds to the panorama watermark processing method for the three-dimensional Gaussian scene. Since the principle of the device in the embodiment of the present application for solving the problem is similar to the above-mentioned panorama watermark processing method for the three-dimensional Gaussian scene, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described here.

[0107] Referring to Figure 3 FIG. 1 shows a schematic diagram of a panorama watermark processing device for a three-dimensional Gaussian scene provided by the embodiment of the present application, which comprises:

[0108] The acquisition module 301 is configured to acquire a three-dimensional Gaussian scene to be processed, a two-dimensional original image corresponding to the three-dimensional Gaussian scene at each first view angle, and a first rendered image corresponding to the three-dimensional Gaussian scene at each first view angle. Each first rendered image is a two-dimensional image obtained by image rendering of a first Gaussian unit at the corresponding first view angle in the three-dimensional Gaussian scene.

[0109] The sensitivity determination module 302 is configured to determine the parameter sensitivity of each first Gaussian unit according to the difference degree value between the two-dimensional original image and the first rendered image at each first view angle of each first Gaussian unit. The parameter sensitivity is used to quantify the influence degree of the parameter disturbance introduced by embedding a watermark at the corresponding first Gaussian unit on the geometric structure integrity and the rendering visual consistency of the three-dimensional Gaussian scene.

[0110] The Gaussian unit screening module 303 is configured to screen a second Gaussian unit corresponding to a parameter sensitivity greater than a preset parameter sensitivity threshold from all first Gaussian units.

[0111] The watermark embedding module 304 is configured to embed a copyright message in the second Gaussian unit in the three-dimensional Gaussian scene, thereby completing the embedding operation of the panorama watermark.

[0112] In a possible implementation, the sensitivity determination module 302 is specifically configured to calculate the parameter sensitivity of any first Gaussian unit according to the following steps: for each first view angle at which the first Gaussian unit is located, calculating a parameter sensitivity matrix of the first Gaussian unit at the first view angle according to the difference degree value between the two-dimensional original image and the first rendered image at the first view angle; and fusing the parameter sensitivity matrices of the first Gaussian unit at all corresponding first view angles to obtain the parameter sensitivity of the first Gaussian unit.

[0113] In a possible implementation, the sensitivity determining module 302 is specifically configured to substitute the difference degree value between the two-dimensional original image and the first rendered image at the first view angle into the following formula to obtain the parameter sensitivity matrix of the first Gaussian unit at the first view angle.

[0114] ;

[0115] wherein, is the parameter sensitivity matrix of the first Gaussian unit at the first view angle; is the kth first view angle; is the parameter set of the ith first Gaussian unit; is the difference degree value between the two-dimensional original image and the first rendered image at the first view angle; is the difference degree value between the two-dimensional original image and the first rendered image at the first view angle; is the first-order gradient vector of the parameter set of the ith first Gaussian unit; is the first-order gradient vector of the parameter set of the ith first Gaussian unit; is the transpose.

[0116] In a possible implementation, the sensitivity determining module 302 is specifically configured to substitute the parameter sensitivity matrix of the first Gaussian unit at all corresponding first view angles into the following formula to obtain the parameter sensitivity of the first Gaussian unit.

[0117] ;

[0118] wherein, is the parameter sensitivity of the ith first Gaussian unit, is the number of first view angles in which the ith first Gaussian unit is located, is the trace of a matrix, is the parameter sensitivity matrix of the first Gaussian unit at the first view angle, is the kth first view angle, is the parameter set of the ith first Gaussian unit.

[0119] ​In a possible implementation, the apparatus further comprises a watermark extraction verification module 305; after the copyright message is embedded in the second Gaussian unit in the three-dimensional Gaussian scene, the watermark extraction verification module 305 is configured to input parameters of each third Gaussian unit in a target three-dimensional Gaussian scene into a three-dimensional watermark extraction model to obtain a first copyright extraction message; the target three-dimensional Gaussian scene is the three-dimensional Gaussian scene after the copyright message is embedded; and / or input all second rendered images corresponding to the target three-dimensional Gaussian scene into a two-dimensional watermark extraction model to obtain a second copyright extraction message; each second rendered image is a two-dimensional image obtained by image rendering on a fourth Gaussian unit in the target three-dimensional Gaussian scene at a corresponding second view angle; and determine a watermark verification result corresponding to the target three-dimensional Gaussian scene according to a matching degree between the first copyright extraction message and the embedded copyright message, and / or a matching degree between the second copyright extraction message and the embedded copyright message.

[0120] In a possible implementation, the watermark extraction verification module 305 is specifically configured to input parameters of each third Gaussian unit into a feature extraction module in the three-dimensional watermark extraction model respectively to obtain a local feature vector corresponding to each third Gaussian unit; the local feature vector corresponding to each third Gaussian unit is used to represent a correlation between parameters in the third Gaussian unit, and / or a correlation between the parameters of the third Gaussian unit and parameters of a third Gaussian unit adjacent to the third Gaussian unit; input the local feature vectors corresponding to all third Gaussian units into a feature aggregation module in the three-dimensional watermark extraction model to obtain a first global feature vector corresponding to the target three-dimensional Gaussian scene; and input the first global feature vector into a first copyright message decoding module in the three-dimensional watermark extraction model to obtain the first copyright extraction message.

[0121] In a possible implementation, the watermark extraction verification module 305 is specifically configured to input each second rendered image into a multi-scale feature extraction layer in the two-dimensional watermark extraction model respectively to obtain a first image feature map of each second rendered image at each preset scale; for each second rendered image, input the first image feature maps of the second rendered image at all preset scales into a first feature fusion layer in the two-dimensional watermark extraction model to obtain a second image feature vector corresponding to the second rendered image; input the second image feature vectors corresponding to all second rendered images into a second feature fusion layer in the two-dimensional watermark extraction model to obtain a second global feature vector corresponding to the target three-dimensional Gaussian scene; and input the second global feature vector into a second copyright message decoding module in the two-dimensional watermark extraction model to obtain the second copyright extraction message.

[0122] Here, the embodiment of the application provides a panoramic watermark processing device of a three-dimensional Gaussian scene. The device estimates the parameter sensitivity of a first Gaussian unit through a two-dimensional original image and a first rendering image of different viewing angles, so as to consider the geometric structure characteristics of the three-dimensional scene when embedding the watermark, and avoid distortion of the watermark information.

[0123] As shown in Figure 4 The embodiment of the application provides an electronic device 400, which comprises a processor 401, a memory 402 and a bus. The memory 402 stores machine readable instructions executable by the processor 401. When the electronic device is running, the processor 401 and the memory 402 communicate through the bus. The processor 401 executes the machine readable instructions to perform the steps of the panoramic watermark processing method of the three-dimensional Gaussian scene.

[0124] Specifically, the memory 402 and the processor 401 can be general memory and processor, which are not limited here. When the processor 401 runs the computer program stored in the memory 402, the panoramic watermark processing method of the three-dimensional Gaussian scene can be executed.

[0125] Corresponding to the panoramic watermark processing method of the three-dimensional Gaussian scene, the embodiment of the application further provides a computer readable storage medium, and the computer readable storage medium stores a computer program. When the computer program is run by the processor, the steps of the panoramic watermark processing method of the three-dimensional Gaussian scene are executed.

[0126] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the above-mentioned system and device can refer to the corresponding process in the method embodiment, which will not be repeated in the application. In the several embodiments provided in the application, it should be understood that the disclosed system, device and method can be implemented by other ways. The above-mentioned device embodiments are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed mutual elements can be indirect coupling or communication connection through some communication interface, device or module, which can be electrical, mechanical or other forms.

[0127] The modules described as separate components can or can not be physically separated, and the components shown as modules can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiment of the application.

[0128] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit.

[0129] If the functions are realized in the form of software functional units and sold or used as independent products, they can be stored in a nonvolatile computer readable storage medium executable by a processor. Based on this understanding, the technical solutions of the present application or the parts of the technical solutions that essentially contribute to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the three-dimensional Gaussian scene panorama watermarking processing method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a ROM, a RAM, a magnetic disk or an optical disk, and various storage program codes.

[0130] 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 can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A panorama watermarking method of a three-dimensional Gaussian scene, characterized by, The method comprises: acquiring a three-dimensional Gaussian scene to be processed, a two-dimensional original image corresponding to each first view angle of the three-dimensional Gaussian scene, and a first rendered image corresponding to each first view angle of the three-dimensional Gaussian scene; each first rendered image is a two-dimensional image obtained by image rendering of a first Gaussian unit in the three-dimensional Gaussian scene at a corresponding first view angle; determining a parameter sensitivity of each first Gaussian unit according to a difference degree value between the two-dimensional original image and the first rendered image at each first view angle of each first Gaussian unit; the parameter sensitivity is used to quantify the influence degree of parameter disturbance introduced by embedding a watermark at a corresponding first Gaussian unit on the geometric structure integrity and the rendering visual consistency of the three-dimensional Gaussian scene; from all the first Gaussian units, filtering a second Gaussian unit corresponding to a parameter sensitivity greater than a preset parameter sensitivity threshold; embedding a copyright message in the second Gaussian unit in the three-dimensional Gaussian scene, thereby completing the embedding operation of the panoramic watermark.

2. The method of claim 1, wherein, The parameter sensitivity of any first Gaussian unit is calculated according to the following steps, comprising: for each first view angle in which the first Gaussian unit is located, calculating a parameter sensitivity matrix of the first Gaussian unit at the first view angle according to a difference degree value between the two-dimensional original image and the first rendered image at the first view angle; fusing the parameter sensitivity matrices of the first Gaussian unit at all corresponding first view angles to obtain the parameter sensitivity of the first Gaussian unit.

3. The method of claim 2, wherein, The parameter sensitivity matrix of the first Gaussian unit at the first view angle is calculated according to a difference degree value between the two-dimensional original image and the first rendered image at the first view angle, comprising: substituting the difference degree value between the two-dimensional original image and the first rendered image at the first view angle into the following formula to obtain the parameter sensitivity matrix of the first Gaussian unit at the first view angle; ； in, The parameter sensitivity matrix of the first Gaussian element under the first viewpoint; This is the k-th first-person perspective; Let be the parameter set of the i-th first Gaussian unit; For first-person perspective Below, the difference value between the original 2D image and the first rendered image; Values ​​representing the degree of difference The parameter set of the i-th first Gaussian unit The first-order gradient vector; This is a transpose.

4. The method of claim 2, wherein, The parameter sensitivity of the first Gaussian unit is fused by substituting the parameter sensitivity matrices of the first Gaussian unit at all corresponding first view angles into the following formula to obtain the parameter sensitivity of the first Gaussian unit. After the copyright message is embedded in the second Gaussian unit in the three-dimensional Gaussian scene, the method further comprises: ； wherein, is a parameter sensitivity of the i-th first Gaussian element, is a number of first views in which the i-th first Gaussian element is located, is a trace of a matrix, is a parameter sensitivity matrix of the first Gaussian element in the first view, is the k-th first view, is a parameter set of the i-th first Gaussian element.

5. The method of claim 1, wherein, inputting each third Gaussian unit parameter in a target three-dimensional Gaussian scene into a three-dimensional watermark extraction model to obtain a first copyright extraction message; the target three-dimensional Gaussian scene is the three-dimensional Gaussian scene after the copyright message is embedded; and / or, inputting all second rendered images corresponding to the target three-dimensional Gaussian scene into a two-dimensional watermark extraction model to obtain a second copyright extraction message; each second rendered image is a two-dimensional image obtained by image rendering of a fourth Gaussian unit in the target three-dimensional Gaussian scene at a corresponding second view angle; ​ According to the matching degree between the first copyright removal message and the embedded copyright message, and / or the matching degree between the second copyright removal message and the embedded copyright message, a watermark verification result corresponding to the target three-dimensional Gaussian scene is determined.

6. The method of claim 5, wherein, The inputting of each third Gaussian unit parameter in the target three-dimensional Gaussian scene into the three-dimensional watermark extraction model to obtain a first copyright removal message comprises: The inputting of each third Gaussian unit parameter into a feature extraction module in the three-dimensional watermark extraction model to obtain a local feature vector corresponding to each third Gaussian unit; the local feature vector corresponding to each third Gaussian unit is used to represent the correlation between parameters inside the third Gaussian unit, and / or the correlation between the parameters of the third Gaussian unit and the parameters of the adjacent third Gaussian unit; The inputting of the local feature vectors corresponding to all third Gaussian units into a feature aggregation module in the three-dimensional watermark extraction model to obtain a first global feature vector corresponding to the target three-dimensional Gaussian scene; The inputting of the first global feature vector into a first copyright message decoding module in the three-dimensional watermark extraction model to obtain a first copyright removal message.

7. The method of claim 5, wherein the method further comprises: The inputting of all second rendered images corresponding to the target three-dimensional Gaussian scene into a two-dimensional watermark extraction model to obtain a second copyright removal message comprises: The inputting of each second rendered image into a multi-scale feature extraction layer in the two-dimensional watermark extraction model to obtain a first image feature map of each second rendered image at each preset scale; For each second rendered image, the inputting of the first image feature maps of the second rendered image at all preset scales into a first feature fusion layer in the two-dimensional watermark extraction model to obtain a second image feature vector corresponding to the second rendered image; The inputting of the second image feature vectors corresponding to all second rendered images into a second feature fusion layer in the two-dimensional watermark extraction model to obtain a second global feature vector corresponding to the target three-dimensional Gaussian scene; The inputting of the second global feature vector into a second copyright message decoding module in the two-dimensional watermark extraction model to obtain a second copyright removal message.

8. An apparatus for panorama watermarking of a three-dimensional Gaussian scene, characterized by The device comprises: An acquisition module is configured to acquire a three-dimensional Gaussian scene to be processed, two-dimensional original images corresponding to the three-dimensional Gaussian scene at each first view angle, and first rendered images corresponding to the three-dimensional Gaussian scene at each first view angle; each first rendered image is a two-dimensional image obtained by image rendering on a first Gaussian unit in the three-dimensional Gaussian scene at a corresponding first view angle; A sensitivity determination module is configured to determine the parameter sensitivity of each first Gaussian unit according to the difference degree value between the two-dimensional original image and the first rendered image at each first view angle corresponding to each first Gaussian unit; the parameter sensitivity is used to quantify the influence degree of the parameter disturbance introduced by embedding a watermark at the corresponding first Gaussian unit on the geometric structure integrity and the rendering visual consistency of the three-dimensional Gaussian scene; A Gaussian unit screening module is configured to screen second Gaussian units with a parameter sensitivity greater than a preset parameter sensitivity threshold from all first Gaussian units. The watermark embedding module is configured to embed a copyright message in a second Gaussian cell in the three-dimensional Gaussian scene, thereby completing embedding of the panoramic watermark.

9. An electronic device, comprising: The method comprises the following steps: A processor, a storage medium and a bus, the storage medium stores machine readable instructions executable by the processor, when the electronic device is running, the processor and the storage medium communicate through the bus, the processor executes the machine readable instructions to execute the steps of the panoramic watermark processing method of the three-dimensional Gaussian scene as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, when the computer program is run by the processor, the steps of the panoramic watermark processing method of the three-dimensional Gaussian scene as claimed in any one of claims 1 to 7 are executed.

Images