A zero-watermark method for multi-feature clustering

Abstract

The present invention discloses a zero watermark method of multi-feature clustering, comprising the following steps: (1) reconstructing an object in three dimensions to obtain a three-dimensional grid model with texture color; (2) normalizing the three-dimensional grid model with texture color, calculating the spherical surface integral invariant, texture roughness, first, second, third order moments of color and color entropy in units of neighborhood spheres, and using them as the geometric, texture and color features of the three-dimensional grid model with texture color respectively; (3) generating a watermark: clustering the features of the three-dimensional grid model with texture color using a spectral clustering method to generate a zero watermark, and applying the zero watermark to the three-dimensional grid model. IPR Information database registration; (4) Watermark detection: The object is generated into a zero watermark according to the zero watermark method of multi-feature clustering mentioned above. IPR The zero watermark in the information database calculates the correlation coefficient, thereby determining whether the object has a zero watermark; the present invention enables the zero watermark to represent more model features and has robustness against more types of attacks.

Description

Technical Field

[0001] The invention relates to a zero watermark method for multi-feature clustering, and belongs to the field of information hiding. Background Art

[0002] Watermarks are patterns, portraits or words "embedded" in paper by changing the density of pulp fibers to make the paper more transparent during the papermaking process. When viewed against the light, the patterns, portraits or words with light and dark textures can be clearly seen. The earliest watermarks were created by papermaking craftsmen in the Tang Dynasty in my country in the late 8th and early 9th centuries. In the 13th century, Italian papermaking experts also began to print watermarks. Banknotes, certificates, securities, important bills, etc. all use watermarks to prevent counterfeiting.

[0003] Science and technology are changing with each passing day. With the development of digital technology, people have begun to study "digital watermarking", that is, embedding watermarks in digital images, videos, audio, text and other documents. A large number of fruitful research results have emerged in the field of digital watermarking at home and abroad, which has attracted great attention from academia and industry, and has been widely used in the fields of copyright protection, traceability, anti-counterfeiting, authentication, etc. However, with the promotion and deepening of digital watermarking applications, it is far from enough to embed watermarks only in digital carriers. Existing digital watermarking technology is difficult to apply to the anti-counterfeiting and traceability of real objects such as banknotes, ceramics, and artworks. Therefore, the existing digital watermark must return from the digital image carrier to the physical carrier and return to practical application.

[0004] Traditional watermarking methods need to ensure both the invisibility and robustness of the watermark. The less watermark data is embedded, the less impact it has on the host image, and the higher the watermark's invisibility. However, considering the robustness of the method, the more watermark data is embedded, the better. Therefore, it is difficult to meet both invisibility and robustness when designing a watermarking method. In addition, traditional watermarking methods will more or less modify the image data when embedding watermark information, and this "modification" causes certain damage to the image quality. In 2003, Wenquan et al. first proposed the concept of "zero watermark", which is to use the important robust features of the image to construct the watermark information, rather than modifying these features and embedding the watermark information. The "zero" here means that the original data is not modified. The high-order cumulant is used to construct the zero watermark, which can resist geometric attacks such as rotation, but the calculation is large. In 2007, Zhang Yanning et al. calculated the spherical coordinate mapping matrix of the three-dimensional mesh model and performed a two-dimensional wavelet transform to calculate the zero watermark. This scheme is highly robust to vertex reordering, translation, rotation and scaling attacks, but because it depends on the model center and a specific starting vertex, the displacement of the model center after shearing, smoothing and subdivision attacks may cause the constructed spherical mesh to be deformed. There is a big difference between the surface coordinate mapping matrix and the spherical coordinate mapping matrix of the original mesh, resulting in poor robustness to shearing, smoothing and subdivision attacks; in 2017, Zhan Yongzhao et al. used the extraction and tracking method of ordered statistical vertices to determine stable vertices, calculated the shape diameter function value and the biased attitude measurement of the vertex norm for the stable vertices, and combined the distribution characteristics of the number of stable vertices to generate a zero watermark. Due to the use of shape diameter function values ​​and vertex norms, it is not affected by similarity transformations or vertex reordering. However, the impact of noise and clipping attacks on global geometric features is more profound than that on local geometric features. In addition, changes in vertex curvature will also affect vertex selection, so the robustness to noise and clipping attacks is low.

[0005] As the saying goes, "Birds of a feather flock together." Clustering in the computer field is based on this idea. The purpose of clustering is to classify individuals with similar characteristics into one category, so that the similarity between objects of the same category is as large as possible, and the similarity between objects of different categories is as small as possible, so as to find the association and difference between individuals. The K-means method first determines how many categories it will be divided into, selects several points as the initial center points, and then divides them into several categories according to the distance from each point to the center point, iterates several times until the goal is met. The advantages are fast convergence speed, simple calculation and good effect, and the disadvantage is that it is very sensitive to initial values, noise and outliers; density clustering examines the connectivity between samples from the perspective of sample density, and continuously expands clusters based on connectable samples to obtain the final classification results. It can achieve irregular clustering and has better processing of noise data, but it is very sensitive to parameter settings; hierarchical clustering attempts to divide the data set at different levels to form a tree-like clustering structure, which is suitable for situations with large data volumes. The spectral clustering used in this patent first uses Laplace eigenmaps to reduce the dimension of the data, and then uses the K-means method to complete the clustering. The effect is usually better than the K-means method and the computational complexity is lower.

[0006] In 2017, Cui Chenchen et al. constructed a zero watermark through model normalization, grouping and spherical integral invariant. Since the local geometric features of the three-dimensional model are used to calculate the spherical integral invariant, it has good robustness against unified resampling attacks and simplification attacks, but cannot resist cropping attacks and the distortion of the surface texture and color of the three-dimensional mesh model after attacks; in 2012, Wang Kai used the local roughness of the mesh derived from Gaussian curvature to measure the visual quality of the three-dimensional mesh model. Applying this principle to watermarks will have a good visual masking effect and have a high invulnerability. Visibility, can better express the surface texture characteristics of the three-dimensional mesh model, and is resistant to the distortion of the surface texture of the three-dimensional mesh model caused by attacks; color feature is the most widely used visual feature in image retrieval, and it is also the most important perceptual feature for people to recognize images. Color feature has little dependence on the direction, size and viewing angle of the image and the object itself, and has high robustness. Color moment and color entropy are commonly used color retrieval methods, which can better express the essential characteristics of the color of the three-dimensional model, and have good resistance to the distortion of the surface color of the three-dimensional mesh model caused by attacks.

[0007] Nowadays, there is an urgent need for a digital watermarking method with strong robustness and high imperceptibility to embed watermarks in physical carriers. The zero watermarking method based on multi-feature clustering proposed in this patent is a watermarking method that meets such requirements. Summary of the invention

[0008] In order to solve the technical problems mentioned in the above background technology, the present invention proposes a zero-watermark method of multi-feature clustering.

[0009] In order to achieve the above technical objectives, the technical solution of the present invention is:

[0010] A zero watermark method for multi-feature clustering, characterized by comprising the following steps:

[0011] (1) Reconstruct the object in three dimensions to obtain a three-dimensional mesh model with texture and color;

[0012] (2) The three-dimensional mesh model with texture and color is normalized, and the spherical surface integral invariant, texture roughness, first, second, and third order moments of color, and color entropy are calculated with the neighborhood sphere as the unit, which are used as the geometric, texture, and color features of the three-dimensional mesh model with texture and color, respectively;

[0013] (3) Cluster the features of the 3D mesh model with texture color using the spectral clustering method, generate a zero watermark, and register it in the IPR information database;

[0014] (4) Generate a zero watermark for the object to be detected according to the method of steps (1) to (3) above, and calculate the correlation coefficient with the zero watermark in the IPR information database to determine whether the object has a zero watermark.

[0015] Preferably, the three-dimensional reconstruction method in step (1) is to use a three-dimensional scanner to scan the object, perform preprocessing, generate a three-dimensional mesh model, and simultaneously collect the object image, perform texture color mapping, and calculate the collinear equation to generate a three-dimensional mesh model with texture color.

[0016] Preferably, the specific process of step (2) is as follows: converting all vertex coordinates of the three-dimensional grid model with texture color in the three-dimensional rectangular coordinate system into vertex coordinates in the spherical coordinate system, emitting a series of rays uniformly in sequence from the center of the model, calculating the intersection points of the rays with the three-dimensional grid model with texture color, and then calculating the distance between every two intersection points to obtain the shortest distance, and constructing a neighborhood sphere with the intersection point as the sphere center and the shortest distance as the radius; using the attention coefficient Att to represent the attention-attracting ability of the neighborhood sphere, and the calculation method thereof is:

[0017]

[0018] Where L represents the number of neighborhood balls, and the rectangular coordinates of the intersection of the ray and the surface of the three-dimensional grid model with texture color are regarded as a 3×1 column vector X, X i The rectangular coordinate column vector representing the intersection of the i-th ray and the surface of the three-dimensional mesh model with texture color, X j The rectangular coordinate column vector representing the intersection of the jth ray and the surface of the three-dimensional mesh model with texture color, ||X i -X j ||2 means calculate Xi With X j The Euclidean norm of the difference, σ is used to control the local range of the Gaussian kernel function, F i 、F j Indicates the color saturation of the i-th and j-th neighborhood balls;

[0019] A threshold τ is set. When the attention coefficient Att>τ of the neighborhood ball, the neighborhood ball is considered to be a high-attention area; when the attention coefficient Att≤τ of the neighborhood ball, the neighborhood ball is considered to be a low-attention area. After classifying all neighborhood balls into two queues of high-attention area and low-attention area, a neighborhood ball is taken out from the high-attention area queue to generate a ball with the intersection of the ray and the surface of the three-dimensional grid model with texture color as the center and H as the radius. The rays of all neighborhood balls in the low-attention area queue within the ball are moved by an angle ξ toward the direction of the ray of the high-attention neighborhood ball, and the neighborhood ball is removed from the low-attention area queue. This step is repeated until the high-attention area queue is empty.

[0020] Preferably, the specific process of the method for calculating texture roughness in step (2) is as follows:

[0021] The weights in the Laplacian matrix for triangular meshes use both tangent weights and cotangent weights to achieve a more stable mesh deformation effect. The calculation method is:

[0022]

[0023] Among them, D ij , D ii They represent the values ​​of the i-th row and j-th column and the i-th row and i-th column on the Laplacian matrix applicable to the triangular mesh, μ, φ, γ, and ρ represent the values ​​of cotβ ij , cotβ′ ij , tanβ ij tanβ′ ij Taking different weights, β ij and β′ ij It is side v i v j The two opposite angles, V i is the i-th neighbor vertex v in the neighborhood ball i The set of all adjacent neighboring vertices;

[0024] The i-th neighboring vertex v in the neighborhood ball i Texture roughness U i Defined as:

[0025]

[0026] Among them, E i 、Ej represents the i-th and j-th neighboring vertices v in the neighborhood ball i 、v j The discrete Gaussian curvature of .

[0027] Preferably, the specific method of using spectral clustering to fuse the features of the three-dimensional mesh model with texture color and generate a zero watermark in step (3) is as follows:

[0028] The features obtained in each neighborhood ball are concatenated in the order of spherical surface integral invariant, texture roughness, first, second, third order moments of color and color entropy to form a model feature vector η. The spectral clustering method is used for all model feature vectors to classify the model feature vectors into two categories.

[0029] The modulus of the mean vectors of the two clusters obtained after clustering is completed is calculated respectively. The zero watermark bit corresponding to the model feature vectors of all clusters belonging to the mean vector with a larger modulus is 1, and the zero watermark bit corresponding to the model feature vectors of all clusters belonging to the mean vector with a smaller modulus is 0. Therefore, all neighborhood balls generate a corresponding zero watermark bit, and the final zero watermark is formed in the order of the original neighborhood balls. After being encrypted with the private key, it is registered in the IPR information database with a timestamp.

[0030] Preferably, the specific steps of detecting the watermark in step (4) are as follows:

[0031] The zero watermark generated by the same method as the watermark generation process is matched with all the registered zero watermarks in the IPR information database. The correlation coefficient Corr is used to represent the similarity between the two. A threshold Δ is set. When Corr>Δ, it is considered that the object has a watermark; when Corr≤Δ, it is considered that the object does not have a watermark.

[0032] The beneficial effects brought by adopting the above technical solution are:

[0033] 1. Watermarking technology based on a single principle is often more robust to certain types of attacks, but less robust to other types of attacks. Therefore, the idea of ​​feature fusion is used to extract the surface geometric features, texture features and color features of the 3D model to generate a zero watermark, so that the zero watermark can represent more model characteristics and be robust to more types of attacks.

[0034] 2. In the feature clustering stage, the unsupervised spectral clustering method replaces the previous practice of setting thresholds to classify feature values ​​and generate zero watermarks. Instead, it classifies and generates zero watermarks based on the distance of feature vectors in high-dimensional space, which can better reflect the relationship between features and improve the robustness of zero watermarks.

[0035] 3. Change the distribution of rays used to determine the watermark embedding position so that it is no longer uniformly distributed, but the distribution relationship between all rays is comprehensively considered. Combined with color saturation, more rays are distributed in brighter areas on the 3D grid model with textured colors, and more model features are extracted, thereby capturing the most conspicuous features of the 3D model. The generated zero watermark can better reflect the main features of the model; BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is the overall flow chart of zero watermark generation and detection based on feature clustering;

[0037] Figure 2 It is a flow chart for generating a three-dimensional mesh model with texture color;

[0038] Figure 3 It is a normalization flow chart of a three-dimensional mesh model with texture color;

[0039] Figure 4 It is a schematic diagram of texture roughness angle. DETAILED DESCRIPTION

[0040] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0041] The present invention discloses a zero watermark method for multi-feature clustering, such as Figure 1 As shown, the following steps are included:

[0042] (1) Reconstruct the object in three dimensions to obtain a three-dimensional mesh model with texture and color;

[0043] (2) The three-dimensional mesh model with texture and color is normalized, and the spherical surface integral invariant, texture roughness, first, second, and third order moments of color, and color entropy are calculated with the neighborhood sphere as the unit, which are used as the geometric, texture, and color features of the three-dimensional mesh model with texture and color, respectively;

[0044] (3) Cluster the features of the 3D mesh model with texture color using the spectral clustering method, generate a zero watermark, and register it in the IPR information database;

[0045] (4) Generate a zero watermark for the object to be detected according to the method of steps (1) to (3) above, and calculate the correlation coefficient with the zero watermark in the IPR information database to determine whether the object has a zero watermark.

[0046] like Figure 2As shown, the method of three-dimensional reconstruction in step (1) is to use a three-dimensional scanner to scan the object, perform preprocessing, generate a three-dimensional mesh model, and simultaneously collect the object image, perform texture color mapping, and generate a three-dimensional mesh model with texture color through collinear equation calculation, so that the zero watermark can represent more model features and be robust to more types of attacks;

[0047] like Figure 3 As shown, the specific process of normalizing the three-dimensional mesh model with texture color in step (2) is as follows:

[0048] The coordinate system transformation on the three-dimensional mesh model with texture color can ensure translation invariance; the vertex coordinates on the three-dimensional mesh model with texture color are transformed into coordinates with the center of mass as the origin, and the coordinates of the center of mass O are calculated (o x ,o y ,o z ) is:

[0049]

[0050] Among them, S T is the total area of ​​the triangular patches on the 3D mesh model with texture color, N f is the total number of triangle patches on the 3D mesh model with texture color, S i is the area of ​​the i-th triangle patch on the surface of the three-dimensional mesh model with texture color, Respectively represent the three vertices A on the i-th triangle patch on the three-dimensional mesh model with texture color i , B i , C i The coordinates of

[0051] The coordinate V of the kth vertex on the 3D mesh model with texture color k =(x k ,y k ,z k ) After the coordinate system is transformed, the coordinates of the vertex are changed to V′ k =(x′ k ,y′ k ,z′ k ), which is calculated as follows:

[0052]

[0053] The scale invariance of the scale can be ensured by unifying the scale of the three-dimensional mesh model with texture color. The scale factor λ is calculated using the distance between the two farthest vertices on the three-dimensional mesh model with texture color to normalize the size of the model. The calculation method is:

[0054]

[0055] in, It means calculating the sum of the squares of the coordinates of all vertices on the three-dimensional mesh model with texture color and taking the maximum value. It means calculating the square sum of the coordinates of all vertices on the three-dimensional mesh model with texture color and taking the minimum value, N v Represents the total number of vertices on a three-dimensional mesh model with texture color;

[0056] The coordinate V″ of the kth vertex on the 3D mesh model with texture color after the mesh scale is unified k =(x″ k ,y″ k ,z″ k ) is calculated as:

[0057]

[0058] The continuous principal component analysis method is used to standardize the mesh rotation on the three-dimensional mesh model with texture color, which can ensure rotation invariance; the influence matrix M is calculated by:

[0059]

[0060] Among them, f(A i ) means to convert A i The coordinates of B are considered as column vectors and the inner product is calculated, f(B i ) means to convert B i The coordinates of are considered as column vectors and the inner product is calculated, f(C i ) means to convert C i The coordinates of are considered as column vectors and the inner product is calculated. Indicates that A i , B i , C i The coordinates of are considered as column vectors and the average is calculated before calculating the inner product;

[0061] Calculate the eigenvalues ​​of M, sort them, and extract the corresponding eigenvectors. Then the eigenvectors are normalized to unit length. The ordered and standardized eigenvector matrix is ​​the rotation matrix G. The coordinates V″′ of the kth vertex on the three-dimensional mesh model with texture color after the standard rotation are k =(x″′ k ,y″′ k ,z″′ k ) is calculated as:

[0062]

[0063] in, represents the multiplication of G with the matrix consisting of the coordinates of the kth vertex on the three-dimensional mesh model with texture color;

[0064] All vertex coordinates of the three-dimensional grid model with texture color in the three-dimensional rectangular coordinate system are converted into vertex coordinates in the spherical coordinate system. A series of rays are uniformly emitted from the center of the model in sequence, and the intersection points of the rays and the three-dimensional grid model with texture color are calculated. Then, the distance between each two intersection points is calculated to obtain the shortest distance. A neighborhood sphere is constructed with the intersection point as the sphere center and the shortest distance as the radius. The attention coefficient Att is used to represent the attention-attracting ability of the neighborhood sphere, and its calculation method is:

[0065]

[0066] Where L represents the number of neighborhood balls, and the rectangular coordinates of the intersection of the ray and the surface of the three-dimensional grid model with texture color are regarded as a 3×1 column vector X, X i The rectangular coordinate column vector representing the intersection of the i-th ray and the surface of the three-dimensional mesh model with texture color, X j The rectangular coordinate column vector representing the intersection of the jth ray and the surface of the three-dimensional mesh model with texture color, ||X i -X j ||2 means calculate X i With X j The Euclidean norm of the difference, σ is used to control the local range of the Gaussian kernel function, F i 、F j Indicates the color saturation of the i-th and j-th neighborhood balls;

[0067] A threshold τ is set. When the attention coefficient Att>τ of the neighborhood ball, the neighborhood ball is considered to be a high-attention area; when the attention coefficient Att≤τ of the neighborhood ball, the neighborhood ball is considered to be a low-attention area. After classifying all neighborhood balls into two queues of high-attention area and low-attention area, a neighborhood ball is taken out from the high-attention area queue to generate a ball with the intersection of the ray and the surface of the three-dimensional grid model with texture color as the center and H as the radius. The rays of all neighborhood balls in the low-attention area queue within the ball are moved by an angle ξ toward the direction of the ray of the high-attention neighborhood ball, and the neighborhood ball is removed from the low-attention area queue. This step is repeated until the high-attention area queue is empty.

[0068] Step (2) calculates the spherical surface integral invariant of each intersection point where the ray intersects with the surface of the three-dimensional mesh model with texture color, thereby obtaining the geometric features of the three-dimensional model with texture color, and the calculation method is:

[0069] For each intersection point P of the neighborhood sphere's surface and the model i, we can calculate the area of ​​each spherical cap and the intersection point P i , the plane P formed by the center of mass O and the point T on the neighborhood sphere i OT, find all planes P i OT and Planar P i+1 The angle between OT i , and the next plane of the last plane is the first plane and the included angle is an acute angle, then the sum of the spherical cap areas can be approximately calculated, that is, the spherical surface integral invariant SIV; the calculation method of the spherical surface integral invariant SIV is:

[0070]

[0071] Among them, N C is the total number of intersections between the surface of the three-dimensional mesh model with texture color and the neighborhood sphere, S i is the area of ​​the i-th triangle patch on the surface of the three-dimensional mesh model with texture color, R is the radius of the neighborhood sphere, and h i From point P i The length of the projection of the vector to T in the direction of the ray.

[0072] like Figure 4 As shown, the specific process of the method for calculating the texture roughness in step (2) is as follows:

[0073] The weights in the Laplacian matrix for triangular meshes use both tangent weights and cotangent weights to achieve a more stable mesh deformation effect. The calculation method is:

[0074]

[0075] Among them, D ij , D ii They represent the values ​​of the i-th row and j-th column and the i-th row and i-th column on the Laplacian matrix applicable to the triangular mesh, μ, φ, γ, and ρ represent the values ​​of cotβ ij , cotβ′ ij , tanβ ij tanβ′ ij Taking different weights, β ij and β′ ij It is side v i v j The two opposite angles, V i is the i-th neighbor vertex v in the neighborhood ball i The set of all adjacent neighboring vertices;

[0076] The i-th neighboring vertex v in the neighborhood ball i Texture roughness U i Defined as:

[0077]

[0078] Among them, E i 、E j represents the i-th and j-th neighboring vertices v in the neighborhood ball i 、v j The discrete Gaussian curvature of .

[0079] Step (2) uses the first, second, and third order moments of color and color entropy to characterize the color features of the three-dimensional grid model with texture color. The specific calculation method is as follows:

[0080] Convert RGB color space to YCbCr color space, the conversion formula is:

[0081]

[0082] Among them, C Y , C Cb , C Cr They represent the values ​​of the brightness element Y and the two chromaticity elements Cb and Cr of the vertex color of the three-dimensional mesh model with texture color in the YCbCr color space. R , C G , C B Respectively represent the values ​​of the red, green and blue color components of the vertex colors of the three-dimensional mesh model with texture color in the RGB color space;

[0083] A color histogram is constructed for the brightness element Y and two chromaticity elements Cb and Cr of the neighboring vertices in each neighborhood ball, so as to calculate the first, second and third order moments of the color. The first order color moment reflects the brightness, the second order color moment reflects the color distribution range, and the third order color moment reflects the symmetry of the color distribution. The calculation formula is:

[0084]

[0085]

[0086]

[0087] Among them, θ i Represents the first-order color moment of the i-th color channel of the vertex color in the neighborhood ball, n i Represents the grayscale level of the i-th color channel of the vertex color in the neighborhood ball, p ij represents the probability of the pixel with gray value j appearing in the i-th color channel component of the vertex color in the neighborhood ball, χ i Represents the second-order color moment of the i-th color channel of the vertex color in the neighborhood ball, δ i The third-order color moment of the i-th color channel of the vertex color in the neighborhood ball;

[0088] Calculate the average values ​​of the first, second, and third order moments of the colors of all the neighboring vertices in the neighborhood ball, and use them as the first, second, and third order moments of the colors of the neighborhood ball;

[0089] The color histogram is regarded as the probability density function of pixels of different colors appearing in the color space. The color entropy En of the i-th color channel of each neighborhood ball is i The calculation method is:

[0090]

[0091] The first, second, and third order moments and color entropy of the three components Y, Cb, and Cr in the neighborhood sphere are combined into a 12-dimensional vector CO, which is composed of:

[0092] CO=[θ Y ,θ Cb ,θ Cr ,x Y ,x Cb ,x Cr ,δ Y ,δ Cb ,δ Cr ,En Y ,En Cb ,En Cr ]

[0093] CO is used as the color feature of the 3D mesh model with texture color.

[0094] Step (3) uses spectral clustering to fuse the features of the three-dimensional mesh model with texture color. The specific method of generating a zero watermark is as follows:

[0095] The features obtained in each neighborhood ball are concatenated in the order of spherical surface integral invariant, texture roughness, first, second, third order moments of color and color entropy to form a model feature vector η. The spectral clustering method is used for all model feature vectors to classify the model feature vectors into two categories.

[0096] The modulus of the mean vectors of the two clusters obtained after clustering is completed is calculated respectively. The zero watermark bit corresponding to the model feature vectors of all clusters belonging to the mean vector with a larger modulus is 1, and the zero watermark bit corresponding to the model feature vectors of all clusters belonging to the mean vector with a smaller modulus is 0. Therefore, all neighborhood balls generate a corresponding zero watermark bit, and the final zero watermark is formed in the order of the original neighborhood balls. After being encrypted with the private key, it is registered in the IPR information database with a timestamp.

[0097] The specific steps of step (4) watermark detection are as follows: the zero watermark generated by the same method as the watermark generation process is matched with all the registered zero watermarks in the IPR information database, and the correlation coefficient Corr is used to represent the similarity between the two, which is calculated as follows:

[0098]

[0099] in, represents the constructed i-th zero watermark, represents the mean of the constructed zero watermark, represents the i-th zero watermark of the original zero watermark in the IPR database, represents the mean of the original zero watermarks in the IPR database;

[0100] A threshold Δ is set. When Corr>Δ, the object is considered to have a watermark; when Corr≤Δ, the object is considered to have no watermark.

[0101] The embodiments are only for illustrating the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A zero watermark method based on multi-feature clustering, characterized in that: The following steps are involved: (1) Reconstruct the object in three dimensions to obtain a three-dimensional mesh model with texture and color; (2) The three-dimensional mesh model with texture and color is normalized, and the spherical surface integral invariant, texture roughness, first, second, and third order moments of color, and color entropy are calculated with the neighborhood sphere as the unit, which are used as the geometric, texture, and color features of the three-dimensional mesh model with texture and color, respectively; (3) Cluster the features of the 3D mesh model with texture color using the spectral clustering method, generate a zero watermark, and register it in the IPR information database; (4) Generate a zero watermark for the object to be detected according to the method of steps (1) to (3) above, and calculate the correlation coefficient with the zero watermark in the IPR information database to determine whether the object has a zero watermark.

2. According to claim 1, a multi-feature clustering zero watermark method is characterized in that: The three-dimensional reconstruction method in step (1) is to use a three-dimensional scanner to scan the object, perform preprocessing, generate a three-dimensional mesh model, and simultaneously collect the object image, perform texture color mapping, and calculate the collinear equation to generate a three-dimensional mesh model with texture color.

3. According to the multi-feature clustering zero watermark method of claim 1, it is characterized in that: The specific process of step (2) is as follows: All vertex coordinates of the three-dimensional grid model with texture color in the three-dimensional rectangular coordinate system are converted into vertex coordinates in the spherical coordinate system. A series of rays are uniformly emitted from the center of the model in sequence, and the intersection points of the rays and the three-dimensional grid model with texture color are calculated. Then, the distance between each two intersection points is calculated to obtain the shortest distance. A neighborhood sphere is constructed with the intersection point as the sphere center and the shortest distance as the radius. The attention coefficient Att is used to represent the attention-attracting ability of the neighborhood sphere, and its calculation method is: Where L represents the number of neighborhood balls, and the rectangular coordinates of the intersection of the ray and the surface of the three-dimensional grid model with texture color are regarded as a 3×1 column vector X, X i The rectangular coordinate column vector representing the intersection of the i-th ray and the surface of the three-dimensional mesh model with texture color, X j The rectangular coordinate column vector representing the intersection of the jth ray and the surface of the three-dimensional mesh model with texture color, ||X i -X j ||2 means calculate X i With X j The Euclidean norm of the difference, σ is used to control the local range of the Gaussian kernel function, F i 、F j Indicates the color saturation of the i-th and j-th neighborhood balls; A threshold τ is set. When the attention coefficient Att>τ of the neighborhood ball, the neighborhood ball is considered to be a high-attention area; when the attention coefficient Att≤τ of the neighborhood ball, the neighborhood ball is considered to be a low-attention area. After classifying all neighborhood balls into two queues of high-attention area and low-attention area, a neighborhood ball is taken out from the high-attention area queue to generate a ball with the intersection of the ray and the surface of the three-dimensional grid model with texture color as the center and H as the radius. The rays of all neighborhood balls in the low-attention area queue within the ball are moved by an angle ξ toward the direction of the ray of the high-attention neighborhood ball, and the neighborhood ball is removed from the low-attention area queue. This step is repeated until the high-attention area queue is empty.

4. The zero watermark method of multi-feature clustering according to claim 1, characterized in that: The specific process of the method for calculating texture roughness in step (2) is as follows: The weights in the Laplacian matrix for triangular meshes use both tangent weights and cotangent weights to achieve a more stable mesh deformation effect. The calculation method is: Among them, D ij , D ii They represent the values ​​of the i-th row and j-th column and the i-th row and i-th column on the Laplacian matrix applicable to the triangular mesh, respectively. μ, φ, γ, and ρ represent the values ​​of cotβ ij , cotβ′ ij , tanβ ij tanβ′ ij Taking different weights, β ij and β′ ij It is side v i v j The two opposite angles, V i is the i-th neighbor vertex v in the neighborhood ball i The set of all adjacent neighboring vertices; The i-th neighboring vertex v in the neighborhood ball i Texture roughness U i Defined as: Among them, E i 、E j represents the i-th and j-th neighboring vertices v in the neighborhood ball i 、v j The discrete Gaussian curvature of .

5. The zero watermark method of multi-feature clustering according to claim 1, characterized in that: The specific method of using spectral clustering to fuse the features of the three-dimensional grid model with texture color and generate a zero watermark in step (3) is as follows: The features obtained in each neighborhood ball are concatenated in the order of spherical surface integral invariant, texture roughness, first, second, third order moments of color and color entropy to form a model feature vector η. The spectral clustering method is used for all model feature vectors to classify the model feature vectors into two categories. The modulus of the mean vectors of the two clusters obtained after clustering is completed is calculated respectively. The zero watermark bit corresponding to the model feature vectors of all clusters belonging to the mean vector with a larger modulus is 1, and the zero watermark bit corresponding to the model feature vectors of all clusters belonging to the mean vector with a smaller modulus is 0. Therefore, all neighborhood balls generate a corresponding zero watermark bit, and the final zero watermark is formed in the order of the original neighborhood balls. After being encrypted with the private key, it is registered in the IPR information database with a timestamp.

6. The zero watermark method of multi-feature clustering according to claim 1, characterized in that: The specific steps of detecting the watermark in step (4) are as follows: The zero watermark generated by the same method as the watermark generation process is matched with all the registered zero watermarks in the IPR information database. The correlation coefficient Corr is used to represent the similarity between the two. A threshold Δ is set. When Corr>Δ, it is considered that the object has a watermark; when Corr≤Δ, it is considered that the object does not have a watermark.

Images