CN-122023101-A - Reversible watermark generation method and device for Gaussian splash model

CN122023101ACN 122023101 ACN122023101 ACN 122023101ACN-122023101-A

Abstract

The application provides a reversible watermark generation method and device of a Gaussian splash model, and relates to the technical field of watermark algorithms; selecting a reference Gaussian primitive from a Gaussian splash model to be protected, determining a reference position based on covariance of the reference Gaussian primitive, selecting a plurality of nearest neighbor Gaussian primitives in a neighborhood by taking the reference position as a center to construct a local primitive set, performing feature decoupling on the local primitive set to obtain a plurality of geometric appearance features, adding nonlinear disturbance to the geometric appearance features to obtain disturbance geometric appearance features, performing transformation processing on a target seal model by utilizing the disturbance geometric appearance features, and embedding the transformed target seal model into the Gaussian splash model to be protected to obtain the Gaussian splash model carrying the dominant watermark seal. By adopting the method and the device for generating the reversible watermark of the Gaussian splatter model, the problems that the accuracy of the 3DGS model cannot be recovered after the watermark is removed and the complexity of watermark generation is high are solved.

Inventors

LI CHENCHEN
ZHAO KAIYONG
WANG XIANGCHUN

Assignees

深圳市其域创新科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. A gaussian splatter model reversible watermark generation method, comprising: constructing at least one watermark seal model, and selecting a target seal model from the at least one watermark seal model; Selecting a reference Gaussian primitive from a Gaussian splatter model to be protected, and determining a reference position based on covariance of the reference Gaussian primitive; Selecting a plurality of nearest neighbor Gaussian primitives in the neighborhood by taking the reference position as a center to construct a local primitive set, and performing feature decoupling on the local primitive set to obtain a plurality of geometric appearance features; Adding nonlinear disturbance to the geometric appearance features to obtain disturbance geometric appearance features; and carrying out transformation processing on the target seal model by utilizing the disturbance geometric appearance characteristics, and embedding the transformed target seal model into the Gaussian splatter model to be protected so as to obtain the Gaussian splatter model carrying the dominant watermark seal.
2. The method of claim 1, wherein the plurality of geometric appearance features includes a normal vector, a reference scale, and a reference hue, and wherein the step of feature decoupling the local primitive set to obtain the plurality of geometric appearance features comprises: performing principal component analysis on each center point coordinate in the local primitive set, determining a plurality of eigenvectors of a covariance matrix, and selecting a normal vector from the eigenvectors; Determining the statistic value of the scaling parameters of all nearest neighbor Gaussian primitives in the local primitive set as a reference scale; And extracting spherical harmonic coefficients of all nearest neighbor Gaussian primitives in the local primitive set, and determining a reference tone based on the spherical harmonic coefficients.
3. The method of claim 1, further comprising, prior to transforming the target stamp model with the perturbed geometric appearance feature: Geometrically encoding the target seal model to obtain a binary watermark seal sequence; carrying out neighborhood screening on each original Gaussian primitive in the Gaussian splatter model to be protected to obtain a neighborhood set of each original Gaussian primitive; generating auxiliary Gaussian primitives for each neighborhood set in sequence according to the sequence of the binary watermark seal sequence; All auxiliary Gaussian primitives are formed into a hidden auxiliary Gaussian layer, and the hidden auxiliary Gaussian layer is combined with the original Gaussian primitives to obtain a Gaussian splatter model carrying a hidden watermark seal.
4. A method according to claim 3, wherein the step of generating auxiliary gaussian points for each set of neighborhoods in turn in the order of the sequence of binary watermark seals comprises: For each original Gaussian primitive, determining the generation position of an auxiliary Gaussian primitive corresponding to the original Gaussian primitive through offset processing based on the distribution condition of a neighborhood set corresponding to the original Gaussian primitive; mapping each value in the binary watermark seal sequence into covariance of auxiliary Gaussian primitives at corresponding generation positions according to a preset coding rule; The opacity of each auxiliary gaussian cell is set so that all auxiliary gaussian cells have implicit characteristics.
5. The method of claim 1, wherein the step of selecting the reference gaussian primitives from the gaussian splatter model to be protected comprises: And identifying a target area meeting the requirement of a preset area in the Gaussian splash model to be protected by utilizing a semantic segmentation model, and selecting a reference Gaussian primitive from the target area.
6. The method of claim 1, wherein the target stamp model is transformed with the perturbed geometric appearance features by: and carrying out translation transformation, rotation alignment, scaling and spherical harmonic fusion and opacity conversion processing on the target seal model by utilizing the disturbance geometric appearance characteristics.
7. The method of claim 1, wherein the step of constructing at least one watermark stamp model comprises: Respectively carrying out three-dimensional Gaussian splatter reconstruction on each preset watermark mark to obtain an initial watermark seal model; and carrying out normalization processing on each initial watermark seal model to construct at least one watermark seal model.
8. The method of claim 7, wherein the step of normalizing each initial watermark stamp model to construct a watermark stamp model comprises: Moving the geometric center of each initial watermark seal model to a coordinate origin, and scaling the whole size of each initial watermark seal model to be within a preset unit sphere range; and unifying the orientations of all the initial watermark seal models to obtain the watermark seal models.
9. The method of claim 1, wherein the at least one watermark seal model comprises a plurality of watermark seal models, and further comprising, after embedding the transformed target seal model in the gaussian splatter model to be protected: And returning to the step of selecting the target seal model from the at least one watermark seal model so as to embed the re-selected target watermark model into the Gaussian splatter model to be protected.
10. A gaussian splash model reversible watermark generation apparatus, comprising: The target model selecting module is used for constructing at least one watermark seal model and selecting a target seal model from the at least one watermark seal model; the reference position determining module is used for selecting a reference Gaussian primitive from the Gaussian splatter model to be protected and determining a reference position based on covariance of the reference Gaussian primitive; The feature acquisition module is used for selecting a plurality of nearest neighbor Gaussian primitives in the neighborhood by taking the reference position as a center to construct a local primitive set, and performing feature decoupling on the local primitive set to obtain a plurality of geometric appearance features; The disturbance adding module is used for adding nonlinear disturbance to the geometric appearance characteristics to obtain disturbance geometric appearance characteristics; and the watermark embedding module is used for carrying out transformation processing on the target seal model by utilizing the disturbance geometric appearance characteristics, and embedding the transformed target seal model into the Gaussian splatter model to be protected so as to obtain the Gaussian splatter model carrying the dominant watermark seal.

Description

Reversible watermark generation method and device for Gaussian splash model Technical Field The application relates to the technical field of watermarking algorithms, in particular to a reversible watermark generation method and device of a Gaussian splash model. Background Three-dimensional gaussian splatter (3D Gaussian Splatting,3DGS) is a breakthrough technology in the field of three-dimensional reconstruction and real-time rendering in recent years, and adopts an explicit representation method based on 3D gaussian distribution means, covariance, opacity and spherical harmonics. But as the commercial value of 3DGS model assets becomes increasingly prominent, copyright problems such as model leakage, illegal tampering, unauthorized transmission and the like are also accompanied. Since the core data of the 3DGS is a series of parameter files stored in a point cloud-like form, an attacker can easily copy, redistribute and even make commercial use without authorization. Therefore, developing an efficient watermarking technology for 3DGS models has extremely critical legal and commercial significance for rights and infringement tracking and prevention of unauthorized businesses. Existing 3DGS model watermarking techniques mainly embed watermark information by fine tuning parameters of original gaussian points or generate watermarks based on neural network models. However, if information is embedded only by fine tuning the parameters of the original gaussian points, irreversible shift of scene geometry or colors often occurs, and the accuracy of the original 3DGS model cannot be recovered after watermark removal. If the watermark generation method based on the neural network model needs a complex generation process, so that the 3DGS watermark itself becomes an operation with high calculation cost. Disclosure of Invention Accordingly, the present application is directed to a method and apparatus for generating a reversible watermark of a gaussian splash model, which overcomes at least one of the above-mentioned drawbacks. In a first aspect, an embodiment of the present application provides a method for generating a reversible watermark of a gaussian splash model, including: constructing at least one watermark seal model, and selecting a target seal model from the at least one watermark seal model; selecting a reference Gaussian primitive from the Gaussian splash model to be protected, and determining a reference position based on covariance of the reference Gaussian primitive; selecting a plurality of nearest neighbor Gaussian primitives in the neighborhood by taking the reference position as the center to construct a local primitive set, and performing feature decoupling on the local primitive set to obtain a plurality of geometric appearance features; adding nonlinear disturbance to the geometric appearance features to obtain disturbance geometric appearance features; and carrying out transformation processing on the target seal model by using the disturbance geometric appearance characteristics, and embedding the transformed target seal model into the Gaussian splatter model to be protected so as to obtain the Gaussian splatter model carrying the dominant watermark seal. In an alternative embodiment, the step of performing feature decoupling on the local primitive set to obtain the plurality of geometric appearance features comprises performing principal component analysis on each center point coordinate in the local primitive set, determining a plurality of feature vectors of a covariance matrix, selecting the normal vector from the plurality of feature vectors, determining statistic values of scaling parameters of all nearest neighbor Gaussian primitives in the local primitive set as a reference scale, extracting spherical harmonic coefficients of all nearest neighbor Gaussian primitives in the local primitive set, and determining the reference tone based on the spherical harmonic coefficients. In an alternative embodiment, before the target seal model is transformed by using disturbance geometric appearance characteristics, the method further comprises the steps of carrying out geometric encoding on the target seal model to obtain a binary watermark seal sequence, carrying out neighborhood screening on each original Gaussian primitive in the Gaussian splatter model to be protected to obtain a neighborhood set of each original Gaussian primitive, generating auxiliary Gaussian primitives for each neighborhood set in sequence according to the sequence of the binary watermark seal sequence, forming all the auxiliary Gaussian primitives into a hidden auxiliary Gaussian layer, and combining the hidden auxiliary Gaussian layer with the original Gaussian primitives to obtain the Gaussian splatter model carrying the hidden watermark seal. In an alternative embodiment, the step of generating auxiliary Gaussian points for each neighborhood set in turn according to the sequence of the binary watermark seal sequence