A game model plagiarism detection method based on three-dimensional model geometric features

Abstract

The application discloses a kind of game model plagiarism detection methods based on three-dimensional model geometric features, comprising the following steps: S1, input processing: obtain original three-dimensional game model and suspected infringement three-dimensional model, as input data;S2, point cloud preprocessing: two three-dimensional models are respectively subjected to approximate surface uniform sampling, and model data is sampled into point cloud form;S3, feature extraction: the two point clouds after preprocessing are input into the twin neural network of shared weight, and the multi-level feature representation of two point clouds is respectively extracted using feature extractor.The application relates to the technical field of copyright protection of three-dimensional model, and the beneficial effects of the application are efficient three-dimensional model similarity detection: the application improves the detection efficiency and accuracy of three-dimensional game model similarity by introducing twin network architecture and improving its structure, combined with deep learning method to extract point cloud features.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of copyright protection of three-dimensional models, in particular to a game model plagiarism detection method based on geometric features of three-dimensional models. BACKGROUND

[0002] With the rapid development of the digital entertainment industry, three-dimensional game models have been widely used in game design, film production and virtual reality. However, the problem of copyright infringement is becoming increasingly prominent. In copyright disputes, accurately determining the similarity of two three-dimensional models is crucial to determine whether there is an infringement.

[0003] Current methods for copyright protection of three-dimensional models mainly include digital watermarking methods and zero-watermarking methods. Digital watermarking methods identify copyright by embedding specific information in the model, but this method may affect the quality of the model, and the watermark may be destroyed when the model is modified. The zero-watermarking method does not modify the model, but extracts the features of the model to generate a watermark. However, the robustness and attack resistance of this method may still need to be improved for complex transformations and modifications that game models may face. Therefore, there is an urgent need for an efficient, objective and quantifiable three-dimensional model similarity detection method to assist in copyright protection and infringement determination.

[0004] Invention patent: model retrieval method and device based on three-dimensional model features, equipment and medium, application number: 202311684188.1, this invention patent has similar places with the present application, but the target is completely different, and the technical implementation method is also different, which is explained as follows: The scheme of this patent needs to scale the three-dimensional model to a unit cube bounding box, adjust the camera position and orientation, and change the coordinates of all points, while the scheme of the present application only needs to sample the model into a point cloud; The scheme of this patent uses a polynomial function to fit the preprocessed point cloud to extract features and obtain a distribution feature vector, then inputs the preprocessed point cloud into a pre-constructed twin neural network model to extract features and obtain a three-dimensional model representation feature vector, and then adds the original feature vector obtained by the two. The present application directly uses an improved twin neural network to extract features from two input point clouds, using deep learning technology; The method of this patent needs to calculate the original feature vector of the point cloud data of the three-dimensional model to be queried, then compare the distance between the original feature vector to be queried and the compressed vector stored in the model database to search for similar models, while the scheme of the present application directly uses two three-dimensional models as a sample pair, and directly compares the similarity of the two by the scheme of the present application. SUMMARY

[0005] The purpose of the present application is to solve the above problems, and a game model plagiarism detection method based on geometric features of three-dimensional models is designed.

[0006] The technical scheme of the present application for achieving the above-mentioned purpose is a game model plagiarism detection method based on the geometric features of a three-dimensional model, comprising the following steps:

[0007] S1, input processing: obtaining an original three-dimensional game model and a suspected infringing three-dimensional model as input data;

[0008] S2, point cloud preprocessing: uniformly sampling the approximate surfaces of the two three-dimensional models, and sampling the model data into point cloud form;

[0009] S3, feature extraction: inputting the two preprocessed point clouds into a shared weight twin neural network, and using a feature extractor to extract multi-level feature representations of the two point clouds respectively;

[0010] S4, feature fusion: fusing the feature representations of the two point clouds, including feature splicing, element-by-element multiplication and difference calculation, to form a comprehensive feature vector;

[0011] S5, similarity discrimination: inputting the fused feature vector into a classifier to output a probability value for discriminating whether the two models are similar, and determining as infringing if the probability value exceeds a set threshold.

[0012] Further, in S2, the calculation of the uniformly sampled points on the approximate surface is as follows:

[0013] ;

[0014] Where P is the set of sampled point clouds; m is the total number of triangular faces in the three-dimensional model; are the three vertices of the th triangular face respectively; is a randomly generated value taken from a uniform distribution .

[0015] Further, in S3, the specific process is as follows: inputting the two sampled point clouds as inputs, and using a shared weight feature extractor to extract features from the two point clouds respectively;

[0016] The feature extractor calculates multi-level features step by step, including the following modules:

[0017] Local feature extraction module;

[0018] Global feature extraction module;

[0019] Edge feature extraction module;

[0020] Feature aggregation module;

[0021] The local feature extraction module specifically includes the following steps:

[0022] Anchor sampling:

[0023] From the input point cloud data, anchor points are selected by farthest point sampling to form an anchor point set:

[0024] ;

[0025] wherein, represents an anchor point set, containing the coordinates of the selected anchor points; represents the three-dimensional coordinates of the anchor points; represents the index of the anchor points in the original point cloud set; is a point set of point cloud data, containing all points in the point cloud;

[0026] Edge feature extraction;

[0027] The edge feature between the anchor points and their neighborhood points is calculated to reflect the local geometric relationship of the anchor points:

[0028] ;

[0029] wherein, ;

[0030] wherein, represents the edge feature matrix of the anchor point ; represents the relative displacement between the anchor point and its neighborhood point , which is also called edge feature here; is the edge set between the anchor point and the neighborhood point; is the three-dimensional coordinates of the neighborhood point;

[0031] Edge feature projection;

[0032] Project the edge feature to a high-dimensional feature space:

[0033] ;

[0034] wherein, is the projected edge feature of the th anchor point; is the edge feature projection matrix; is the high-dimensional representation dimension of the edge feature;

[0035] Local attention calculation;

[0036] ;

[0037] wherein, represents the edge feature extracted by the th layer local Transformer; denotes the mapping function of the layer-wise local Transformer;

[0038] neighborhood feature aggregation;

[0039] After each layer of local Transformer processing, local information is extracted and summarized from the neighborhood features of each anchor point:

[0040] ;

[0041] wherein, is the local feature of all anchor points, containing the geometric relationship information of the anchor point and its neighborhood; denotes the aggregation and concatenation operation of the features;

[0042] Further, the global feature extraction module specifically includes the following steps:

[0043] anchor point feature update;

[0044] The updated anchor point feature is updated by the calculated anchor point feature and the aggregated edge feature:

[0045] ;

[0046] wherein, denotes the updated anchor point feature; is the anchor point feature of the previous layer; is the feature projection matrix;

[0047] global attention calculation;

[0048] ;

[0049] wherein, denotes the anchor point feature extracted by the layer-wise global Transformer;

[0050] Further, the edge feature extraction module specifically includes the following steps:

[0051] calculate the global point pair correlation graph;

[0052] Through point pair correlation calculation, the global point pair relationship is generated to measure the global geometric relationship of the point cloud. First, the point pair correlation needs to be calculated. For every two points and in the point cloud, the query and the key value are generated through linear mapping. The point pair correlation measurement is defined as:

[0053] ;

[0054] in, Point and The correlation; and It is a high-dimensional vector obtained through shared linear layers;

[0055] Then, the normalized point-pair correlation plot is calculated:

[0056] ;

[0057] in, It is a point The normalized correlation plot represents the points Weight distribution of neighboring points; It represents the dimensions of the query vector and the key vector;

[0058] Construct a global correlation matrix;

[0059] Integrate all point-to-point correlation maps into a global correlation matrix. :

[0060] ;

[0061] in, each line Point Normalized correlation plot; It refers to the number of points in a point cloud.

[0062] Edge point filtering;

[0063] By analyzing the global correlation matrix The columns are summed to calculate the global contribution of each point. Points with higher global contributions are selected as edge points for sampling. The global contribution of each point is calculated as follows:

[0064] ;

[0065] in, Point Global contribution; It is the global correlation matrix The first in row and number Column elements;

[0066] Edge feature generation;

[0067] For each edge point sampled in the edge point filtering The set of its neighborhood points is selected using the k-NN method. For point and its neighboring points , compute attention weights:

[0068] ;

[0069] where, is the attention weight, indicating the contribution of the neighborhood point to the edge point; and are the query and key value vectors of the point and the neighborhood point ;

[0070] Then, the neighborhood point features are weighted and summed according to the attention weights to generate the feature representation of the edge point:

[0071] ;

[0072] where, is the feature generated by the neighborhood point through linear mapping.

[0073] Further, the global edge feature is aggregated;

[0074] For all the selected edge point features , respectively, through max-pooling and mean-pooling, and the results of the two are spliced to form the final feature representation:

[0075] ;

[0076] ;

[0077] ;

[0078] where, represents the max-pooling result of the edge point feature; represents the mean-pooling result of the edge point feature; represents the global edge feature; represents the vector splicing operation.

[0079] The feature aggregation module specifically includes the following steps:

[0080] Intermediate feature aggregation;

[0081] The multi-layer local features generated by the local feature extraction module are spliced with the multi-layer global features generated by the global feature extraction module according to the layer to construct a unified multi-layer feature representation:

[0082] ;

[0083] Wherein, Z is the spliced multi-layer feature representation, which contains the comprehensive information of local and global features. represents the local feature of the layer, represents the global feature of the layer, ;

[0084] Unified global feature generation;

[0085] Through a single-layer multi-layer perception, abbreviated as: MLP, the spliced features are linearly mapped and activated to generate a unified global feature representation:

[0086] ;

[0087] Wherein, is the fused global feature, is an activation function;

[0088] Complete feature generation

[0089] The unified global feature representation is spliced with the global edge feature to form a complete feature representation:

[0090] ;

[0091] Wherein, represents the final feature representation, which contains local, global and edge information, and can more comprehensively describe the geometric features of the three-dimensional model.

[0092] Further, in S4, the features extracted from the two point clouds are fused according to the following steps:

[0093] Feature splicing: Splice the feature vectors of the two point clouds in the channel dimension.

[0094] Element-level operation: Calculate the element-wise product and absolute difference of the two feature vectors to form a comprehensive feature representation.

[0095] Comprehensive feature vector: The above results are spliced again to obtain a complete comprehensive feature vector.

[0096] ;

[0097] Wherein, represents the feature representation after the fusion of the features of the two point clouds; and are the feature vectors of the two point clouds, respectively; represents the element-wise product.

[0098] Further, in S5, the classifier is a multi-layer perceptron containing multiple fully connected layers, activation functions, batch normalization, and Dropout layers, used to progressively extract and fuse input features and finally output normalized similarity probability values.

[0099] Further, the method further comprises a model training process, specifically comprising:

[0100] Loss function: a contrastive loss function is used to measure the difference between the predicted results and the true labels.

[0101] ;

[0102] where y is the true label, is the boundary of the loss function. and are the final feature representations of the two input point clouds, respectively;

[0103] Optimizer: AdamW or SGD optimizer is used, combined with weight decay and learning rate scheduler, to optimize model parameters;

[0104] Data augmentation: in the training process, point cloud data is randomly rotated, scaled, translated, jittered, locally clipped, and randomly discarded to enhance the robustness of the model.

[0105] The game model plagiarism detection method based on three-dimensional model geometric features provided by the technical scheme of the application has the following beneficial effects:

[0106] 1. Efficient three-dimensional model similarity detection: the present application introduces a twin network architecture and improves its structure, combines deep learning method to extract point cloud features, and improves the detection efficiency and accuracy of three-dimensional game model similarity. Compared with traditional methods, this technology can quickly and accurately identify the similarity between models, and is suitable for large-scale model library comparison and retrieval.

[0107] 2. Strong robustness, suitable for complex plagiarism behavior: the method of the present application can maintain high detection performance when facing local modification, geometric deformation and other complex infringement behaviors of the model.

[0108] 3. No need to pre-align the input three-dimensional point cloud model, i.e. no need to rotate, scale or align the model, but can directly calculate the similarity. This feature significantly simplifies the processing flow, avoids the dependence on model position and scale in traditional methods, improves the detection efficiency, and enhances the robustness of the model under different transformation conditions.

[0109] 4. No need to embed watermark, protect the integrity of the original model: Unlike traditional digital watermarking technology, the present invention does not need to embed any additional information in the model, avoiding the modification of the original model. This not only simplifies the copyright protection process, but also maintains the integrity and quality of the model, suitable for high-precision model application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0110] Figure 1 is a flowchart of a game model plagiarism detection method based on the geometric features of three-dimensional models according to the present invention.

[0111] Figure 2 is a flowchart of an improved twin neural network for a game model plagiarism detection method based on the geometric features of three-dimensional models according to the present invention. DETAILED DESCRIPTION

[0112] The present invention will be described in detail below with reference to the accompanying drawings, such as Figures 1-2 A game model plagiarism detection method based on the geometric features of three-dimensional models, comprising the following steps:

[0113] S1, input processing: obtaining the original three-dimensional game model and the suspected infringing three-dimensional model as input data.

[0114] S2, point cloud preprocessing: uniformly sampling the approximate surface of the two three-dimensional models, and sampling the model data into point cloud form.

[0115] S3, feature extraction: inputting the two preprocessed point clouds into a twin neural network with shared weights, and using a feature extractor to extract multi-level feature representations of the two point clouds.

[0116] S4, feature fusion: fusing the feature representations of the two point clouds, including feature splicing, element-wise multiplication and difference calculation, to form a comprehensive feature vector.

[0117] S5, similarity discrimination: inputting the fused feature vector into a classifier to output a probability value for discriminating whether the two models are similar, and determining as infringing if the value exceeds a set threshold. Supervised training under the guidance of contrastive loss.

[0118] In S1, the input data is the original three-dimensional game triangle mesh model and the suspected infringing triangle mesh model.

[0119] In S2, the calculation of uniformly sampling points on the approximate surface is as follows:

[0120] ;

[0121] P = {p1, p2, p3, …, pn}, where P is the set of sampled point cloud, containing all the points sampled from the mesh model, each point is a three-dimensional coordinate, representing the position on the model surface; m is the total number of triangular faces in the mesh model; are the three vertices of the ith triangle, represented as three-dimensional coordinates; is a randomly generated value taken from a uniform distribution , used to calculate the random point inside the triangle.

[0122] In S3, the specific process is to use the shared weight feature extractor to extract features from the two point clouds obtained by sampling as input.

[0123] The feature extractor calculates multi-level features step by step, including the following modules:

[0124] 1) Local feature extraction module

[0125] 2) Global feature extraction module

[0126] 3) Edge feature extraction module

[0127] 4) Feature aggregation module

[0128] The local feature extraction module specifically includes the following steps:

[0129] 1) Anchor point sampling

[0130] From the input point cloud data, anchor points are selected by farthest point sampling to form an anchor point set:

[0131] ;

[0132] where y represents the anchor point set, containing the coordinates of the selected anchor points; represents the three-dimensional coordinates of the anchor point; represents the index of the anchor point in the original point cloud set; is the point set of point cloud data, containing all the points in the point cloud.

[0133] 2) Edge feature extraction

[0134] Calculate the edge feature between the anchor point and its neighborhood points, which is used to reflect the local geometric relationship of the anchor point:

[0135] ;

[0136] where:

[0137] where, represents the edge feature matrix of the anchor point ; represents the edge feature matrix of the anchor point and its neighborhood points The relative displacement between them is also referred to here as edge feature; It is the set of edges between the anchor point and its neighboring points; These are the three-dimensional coordinates of the neighboring points.

[0138] 3) Edge feature projection

[0139] edge features Projected onto a high-dimensional feature space:

[0140] ;

[0141] in, It is the projected edge feature of the m-th anchor point; The edge feature projection matrix; It is the high-dimensional representation dimension of the edge features.

[0142] 4) Local attention calculation

[0143] ;

[0144] in, Indicates the first Edge features extracted by the local Transformer layer; Indicates the first Mapping functions for layer-local Transformers.

[0145] 5) Neighborhood feature aggregation

[0146] After processing each local Transformer layer, local information is extracted and summarized from the neighborhood features of each anchor point:

[0147] ;

[0148] in, It is a local feature of all anchor points, containing information about the geometric relationship between the anchor point and its neighborhood; This represents the aggregation and splicing operations of features.

[0149] The global feature extraction module specifically includes the following steps:

[0150] 1) Anchor point feature update

[0151] Update the anchor point features by combining the anchor point features and aggregate edge features calculated in claim 6:

[0152] ;

[0153] in, This represents the updated anchor point features; It is the anchor point feature of the previous layer; It is the characteristic projection matrix.

[0154] 2) Global Attention Calculation

[0155] ;

[0156] in, Indicates the first Anchor point features extracted by the global Transformer layer.

[0157] The edge feature extraction module specifically includes the following steps:

[0158] 1) Calculate the global point-pair correlation map

[0159] By calculating point-pair correlation, a global metric for point-pair relationships is generated to capture the global geometric relationships of the point cloud. This first requires calculating point-pair correlation for every two points in the point cloud. and Generate queries through linear mapping and key value The point-to-point correlation metric is defined as:

[0160] ;

[0161] in, Point and The correlation; and It is a high-dimensional vector obtained through shared linear layers.

[0162] Then, the normalized point-pair correlation plot is calculated:

[0163] ;

[0164] in, It is a point The normalized correlation plot represents the points Weight distribution of neighboring points; It represents the dimensions of the query vector and the key vector.

[0165] 2) Construct a global correlation matrix

[0166] Integrate all point-to-point correlation maps into a global correlation matrix. :

[0167] ;

[0168] in, each line Point Normalized correlation plot; is the number of points of the point cloud.

[0169] 3) Edge point screening

[0170] The global contribution of each point is calculated by summing the columns of the global correlation matrix , and the points with higher global contribution are selected as edge points for sampling. The global contribution of each point is calculated as follows:

[0171] ;

[0172] wherein represents the global contribution of point ; is the element in the th row and the th column of the global correlation matrix .

[0173] 4) Edge feature generation

[0174] For each edge point sampled in step 3), a set of neighborhood points is selected by the k-NN method. For point and its neighborhood points , the attention weight is calculated as follows:

[0175] ;

[0176] wherein is the attention weight, representing the contribution of the neighborhood point to the edge point; and are the query and key value vectors of point and neighborhood point , respectively.

[0177] Then, the neighborhood point features are weighted and summed according to the attention weight to generate the feature representation of the edge point:

[0178] ;

[0179] wherein is the feature generated by linear mapping of the neighborhood point.

[0180] 5) Global edge feature aggregation

[0181] For all the selected edge point features , maximum pooling and mean pooling are performed respectively, and the results of the two are spliced to form the final feature representation:

[0182] ;

[0183] ;

[0184] ;

[0185] wherein, represents the max-pooling result of edge point features; represents the mean-pooling result of edge point features; represents the global edge feature; represents the vector concatenation operation.

[0186] The feature aggregation module specifically includes the following steps:

[0187] 1) Intermediate feature aggregation

[0188] The multi-layer local features generated by the local feature extraction module are concatenated with the multi-layer global features generated by the global feature extraction module by layer to construct a unified multi-layer feature representation:

[0189] ;

[0190] wherein, Z is the concatenated multi-layer feature representation, which contains the comprehensive information of local and global features; represents the local feature of the i-th layer, represents the global feature of the i-th layer, . 2) Unified global feature generation

[0191] The concatenated features are linearly mapped and activated by a single-layer multi-layer perception (MLP) to generate a unified global feature representation:

[0192]

[0193] ;

[0194] wherein, is the fused global feature, is an activation function.

[0195] 3) Complete feature generation

[0196] The unified global feature representation is concatenated with the global edge feature to form a complete feature representation:

[0197] ;

[0198] wherein, ​​The final feature representation, which contains local, global and edge information, can more comprehensively describe the geometric features of the three-dimensional model.

[0199] The two point cloud features are fused according to the following steps:

[0200] 1) Feature splicing: Splice the feature vectors of the two point clouds in the channel dimension.

[0201] 2) Element-level operation: Calculate the element-wise product and absolute difference of the two feature vectors to form a comprehensive feature representation.

[0202] 3) Comprehensive feature vector: Splice the above results again to get the complete comprehensive feature vector.

[0203] ;

[0204] wherein, represents the feature representation after the fusion of the two point cloud features; and are the feature vectors of the two point clouds, respectively; and ⊙ represents the element-wise product.

[0205] In S5, the classifier is a multilayer perceptron containing multiple fully connected layers, activation functions, batch normalization and Dropout layers, which is used to gradually extract and fuse the input features, and finally output the normalized similarity probability value.

[0206] Model training, specifically including:

[0207] 1) Loss function: Use the contrastive loss function to measure the difference between the predicted result and the true label. The formula of the loss function is:

[0208] ;

[0209] wherein, is the true label (1 for similar and 0 for dissimilar), is the boundary of the loss function. and are the final feature representations of the two input point clouds, respectively.

[0210] 2) Optimizer: Use AdamW or SGD optimizer combined with weight decay and learning rate scheduler to optimize model parameters.

[0211] 3) Data augmentation: In the training process, randomly rotate, scale, translate, jitter, locally shear and randomly discard the point cloud data to enhance the robustness of the model.

[0212] The above method periodically evaluates the model performance during the training process, saves the best model, and evaluates the accuracy, ROC curve and AUC value of the model on the test set to verify the effectiveness of the model.

[0213] In the present embodiment, as shown in Figure 1 、 2 , the present application provides a game model plagiarism detection method based on three-dimensional model geometric features, comprising:

[0214] Step S1, input processing: obtaining an original three-dimensional game model and a suspected infringing.obj format triangular mesh model, taking them as input data.

[0215] Step S2, point cloud preprocessing: uniformly sampling the approximate surfaces of the two three-dimensional models, and sampling the model data into point cloud form.

[0216] Step S3, feature extraction: inputting the two preprocessed point clouds into a shared weight twin neural network, and using a feature extractor to extract multi-level feature representations and of the two point clouds respectively.

[0217] Step S4, feature fusion: fusing the feature representations of the two point clouds, including feature splicing, element-by-element multiplication and difference calculation, to form a comprehensive feature vector.

[0218] Step S5, similarity discrimination: inputting the fused feature vector into a classifier to output a probability value for discriminating whether the two models are similar, and determining as infringing if the probability value exceeds a set threshold. According to and , the contrast loss is calculated, and the training is supervised under the guidance of the contrast loss.

[0219] The input model in step S1 is labeled with true value (1 for similar and 0 for dissimilar) when input as training data, and the ratio of positive and negative sample pairs in the training set is .

[0220] In the above step S2, the model is sampled into a point cloud of 2048 points.

[0221] The above technical solution only embodies the preferred technical solution of the present application, and some changes made by the skilled in the art to some parts thereof also embody the principles of the present application and are within the protection scope of the present application.

Claims

1. A game model plagiarism detection method based on three-dimensional model geometric features, characterized in that, The method comprises the following steps: S1, input processing: obtaining an original three-dimensional game model and a suspected infringing three-dimensional model as input data; S2, point cloud preprocessing: uniformly sampling the approximate surfaces of the two three-dimensional models respectively, and sampling the model data into point cloud form; S3, feature extraction: inputting the two preprocessed point clouds into a shared weight twin neural network, and using a feature extractor to extract multi-level feature representations of the two point clouds respectively; S4, feature fusion: fusing the feature representations of the two point clouds, including feature splicing, element-wise multiplication and difference calculation, to form a comprehensive feature vector; S5, similarity discrimination: inputting the fused feature vector into a classifier to output a probability value for discriminating whether the two models are similar, and determining that it is infringing if the probability value exceeds a set threshold; In S3, the specific process is: taking the two sampled point clouds as input, and using a shared weight feature extractor to extract features from the two point clouds respectively; The feature extractor calculates multi-level features step by step, including the following modules: a local feature extraction module; a global feature extraction module; an edge feature extraction module; a feature aggregation module; The local feature extraction module specifically includes the following steps: anchor point sampling; selecting anchor points from the input point cloud data to form an anchor point set through farthest point sampling: ； wherein, represents a set of anchor points, containing coordinates of selected anchor points; represents three-dimensional coordinates of an anchor point; represents an index of an anchor point in the original set of point clouds; is a set of points of point cloud data, containing all points in the point cloud; edge feature extraction; calculating edge features between anchor points and their neighborhood points to reflect the local geometric relationship of anchor points: ； wherein ; wherein, represents an edge feature matrix of the anchor point ; represents a relative displacement between the anchor point and its neighborhood points , also referred to as edge features herein; is a set of edges between the anchor point and the neighborhood points; is a three-dimensional coordinate of the neighborhood points; edge feature projection; edge features projected into a high-dimensional feature space: ； wherein, is a projected edge feature of the th anchor point; is an edge feature projection matrix; is a high-dimensional representation dimension of the edge feature. local attention calculation; ； wherein, represents the first layer local Transformer extracted edge feature; represents the first layer local Transformer mapping function; neighborhood feature aggregation; after each layer of local Transformer processing, local information is extracted and summarized from the neighborhood features of each anchor point: ； wherein, is the local feature of all anchor points, containing the geometric relationship information of anchor points and their neighborhood; denotes the aggregation and stitching operation of features; The global feature extraction module specifically includes the following steps: anchor point feature updating; updating the anchor point features with the calculated anchor point features and aggregated edge features: ； wherein, denotes the updated anchor feature; is the anchor feature of the previous layer; is the feature projection matrix; global attention calculation; ； wherein, representing the anchor point features extracted by the layer-global Transformer; The edge feature extraction module specifically includes the following steps: calculating a global point pair correlation graph; By point pair correlation calculation, a global point pair relationship metric is generated for capturing the global geometric relationship of the point cloud. First, the point pair correlation needs to be calculated, for each two points in the point cloud and , the query and key value are generated by linear mapping, and the point pair correlation metric is defined as: ； wherein, representative points and correlation; and are high-dimensional vectors computed through shared linear layers; then calculating a normalized point pair correlation graph: ； wherein, is a normalized correlation map of points representing a weight distribution of points with neighboring points; is the dimension of the query vector and the key vector; constructing a global correlation matrix; integrating all point-to-point correlation maps into a global correlation matrix : ； wherein, each row of represents a point of a normalized correlation plot; is the number of points of the point cloud; edge point screening; The global contribution degree of each point is calculated by summing the columns of the global correlation matrix The point with higher global contribution degree is taken as an edge point for sampling, wherein the global contribution degree of each point is calculated as follows: ； in, Point Global contribution; It is the global correlation matrix The first in row and number Column elements; edge feature generation; For each edge point sampled in the edge point screening , a set of neighborhood points is selected by k-NN method , for point and its neighborhood points , the attention weight is calculated: ； wherein, is an attention weight, representing the contribution of the neighborhood point to the edge point; and are the query and key value vectors of the point and neighborhood point respectively. then, the neighborhood point features are weighted and summed according to the attention weights to generate the feature representation of the edge point: ； wherein, is a feature generated by a neighborhood point through a linear mapping; The feature aggregation module includes: For all the edge point features filtered out , aggregate them by max-pooling and mean-pooling respectively, and concatenate the results of both to form the final feature representation: ； ； ； wherein, represents a max-pooling result of edge point features; represents a mean-pooling result of edge point features; represents a global edge feature; represents a vector concatenation operation; The feature aggregation module specifically includes the following steps: intermediate feature aggregation; the multi-layer local features generated by the local feature extraction module the multi-layer global features generated by the global feature extraction module concatenate the multi-layer local features and the multi-layer global features by layer to construct a unified multi-layer feature representation ； Wherein, Z is the spliced multi-layer feature representation, which contains the comprehensive information of local and global features. represents the local feature of the first layer layer local feature, represents the local feature of the first layer layer global feature, ; uniform global feature generation; performing linear mapping and activation on the spliced features through a single-layer multi-layer perceptron (MLP) to generate a uniform global feature representation: ； wherein, is the fused global feature, is an activation function; complete feature generation; unifying global feature representation with global edge features to form a complete feature representation: ； wherein, The final feature representation indicates the final feature representation, contains local, global and edge information, and can more comprehensively describe the geometric features of the three-dimensional model. 2.The method according to claim 1, wherein, In S2, the calculation of the uniformly sampled points on the approximate surface is as follows: ； wherein, is a set of sampled point clouds; is the total number of triangular faces in the three-dimensional model; are the three vertices of the th triangle, respectively; is a randomly generated numerical value taken from a uniform distribution . 3.The game model plagiarism detection method based on three-dimensional model geometric features according to claim 1, characterized in that, The fusion of the feature representations of the two point clouds specifically includes: extracting the features and fusing them according to the following steps: feature splicing: splicing the feature vectors of the two point clouds in the channel dimension; element-level operation: calculating the element-wise product and absolute difference of the two feature vectors to form a comprehensive feature representation; comprehensive feature vector: splicing the results of splicing the two feature vectors in the channel dimension, the element-wise product of the two feature vectors, and the element-wise absolute difference of the two feature vectors again to obtain a complete comprehensive feature vector; ； wherein, denotes the feature representation of the two point clouds after feature fusion; and are the feature vectors of the two point clouds, respectively; denotes the element-wise product. 4.The game model plagiarism detection method based on three-dimensional model geometric features according to claim 1, characterized in that, In S5, the classifier is a multi-layer perceptron containing multiple fully connected layers, activation functions, batch normalization, and Dropout layers, used to gradually extract and fuse input features, and finally output normalized similarity probability values.

5. The game model plagiarism detection method based on the geometric features of a three-dimensional model according to claim 1, characterized in that, The method also includes a model training process, specifically including: Loss function: using a contrastive loss function to measure the difference between the predicted results and the true labels; ； wherein, is a true label, is a boundary of the loss function; and are final feature representations of the two input point clouds, respectively. Optimizer: using AdamW or SGD optimizer, combined with weight decay and learning rate scheduler, to optimize model parameters; Data augmentation: in the training process, randomly rotate, scale, translate, jitter, local shear, and randomly discard the point cloud data to enhance the robustness of the model.

6. The game model plagiarism detection method based on the geometric features of a three-dimensional model according to claim 1, characterized in that, The method periodically evaluates the performance of the model during the training process and evaluates the accuracy, ROC curve, and AUC value of the model on the test set.

Images