Deep learning based zero-watermarking method for linear vector map data

Abstract

The application discloses a deep learning-based line vector map data zero watermark method, comprising the following steps: learning deep features of line elements based on a transformer network, and extracting high-dimensional features of line elements in a line vector map data set; aggregating the high-dimensional features of all line elements into a global feature through a feature aggregation network; converting the global feature into a binary hash code through forward propagation of a hash layer; generating a zero watermark based on XOR operation between the binary hash code and a copyright image, and storing the zero watermark in a database of a copyright center. The application learns deep features of data as digital fingerprints in a manner of simulating human cognition, realizes strong differentiation, high robustness of copyright authentication of line vector map data under the premise of ensuring zero distortion of original data, and has the advantages of absolute non-loss, strong differentiation ability, excellent robustness, high security and good practicability.

Description

Technical Field

[0001] This invention relates to the field of vector map watermarking technology, and in particular to a zero-watermarking method for linear vector map data based on deep learning. Background Technology

[0002] Vector geographic data is a core asset for digital maps, GIS applications, and spatial analysis. With the widespread adoption of data sharing and cloud services, the illegal copying, dissemination, and tampering of vector data have become increasingly easy, severely damaging the rights of data producers. Traditional digital watermarking technology achieves copyright protection by embedding identification information into the data (such as fine-tuning vertex coordinates), but this poses a potential threat to the geometric fidelity of high-precision vector data, and the embedded information can be removed through targeted attacks.

[0003] Zero-watermarking technology offers a new approach to resolving this contradiction. It does not modify the original data but instead utilizes the stable and discriminative features of the data itself to construct the watermark. Existing zero-watermarking methods for vector data mostly rely on manually designed geometric features (such as angle statistics, length ratios, topological relationships, etc.), which have limited feature discrimination capabilities and insufficient robustness to common geometric attacks (such as vertex addition / deletion, smoothing, and noise addition). Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a simple and secure method for zero-watermarking linear vector map data based on deep learning.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is: a zero-watermarking method for linear vector map data based on deep learning, comprising the following steps:

[0006] Step 1, Line Feature Deep Extraction: Based on the transformer network, learn the deep features of line features to extract the high-dimensional features of line features in the dataset of the line vector map;

[0007] Step 2, Feature aggregation based on attention mechanism: Through the feature aggregation network, the high-dimensional features of all line elements are aggregated into a global feature;

[0008] Step 3, Generate hash code: Through forward propagation of the hash layer, the global features are converted into binary hash code;

[0009] Step 4, Generate Zero Watermark: A zero watermark is generated based on the XOR operation between the binary hash code and the copyrighted image, and stored in the copyright center's database. The specific process of step 1 in the above deep learning-based zero watermarking method for line vector map data is as follows:

[0010] Step 11, Data Preprocessing and Augmentation: Represent each line feature in the dataset as a series of node coordinates to obtain a variable-length coordinate sequence, and normalize the node coordinates.

[0011] Step 12, Construct a feature extraction network: Construct a feature extraction network based on a Transformer encoder to process coordinate sequences and capture global dependencies;

[0012] The feature extraction network consists of an input embedding layer, a Transformer encoder, and a pooling layer:

[0013] The input embedding layer projects the coordinates of each node into a high-dimensional space through a linear layer and adds learnable positional encoding;

[0014] The Transformer encoder consists of a multi-layer self-attention mechanism and a feedforward network;

[0015] The pooling layer aggregates the information of the entire coordinate sequence and outputs a feature vector of fixed dimensions.

[0016] Step 13, Training the feature extraction network: The training uses a triplet loss function to construct triplet pairs and uses gradient descent to optimize the network parameters.

[0017] The specific process of step 11 in the above-mentioned deep learning-based zero-watermarking method for linear vector map data is as follows:

[0018] Step 111: For a line feature, calculate the minimum X-coordinate of the line feature across all nodes. and the maximum value of the X coordinate Calculate the minimum Y-coordinate of the line feature across all nodes. and the maximum value of the Y coordinate Get the coordinates of the bottom left corner of the bounding box. , ) and the coordinates of the upper right corner ( , );

[0019] Step 112: Calculate the width of the bounding box and the height of the enclosure The formula is as follows:

[0020] ;

[0021] ;

[0022] Step 113: Perform normalization calculations for the first... There are 1 node, and the formula is as follows:

[0023] ;

[0024] ;

[0025] in, For the first The normalized x-coordinates of each node For the first The x-coordinate of each node, For the first The normalized ordinates of each node For the first The ordinate of each node;

[0026] Step 114: During training, randomly apply operations such as adding nodes, deleting nodes, and perturbing nodes to perform data augmentation.

[0027] The specific steps in step 12 of the above-mentioned deep learning-based zero-watermarking method for linear vector map data are as follows:

[0028] First, a linear layer is used to project the two-dimensional coordinates onto... 3D space For the node feature dimension, for the ... The coordinate vector of each node The linear layer performs the following calculations, as shown in the formula below:

[0029] ;

[0030] in, It is a learnable weight matrix; It is a learnable bias vector; It is a two-dimensional coordinate projection onto The high-dimensional projection vector obtained from the 1D space;

[0031] Next, add learnable positional encoding. Specifically, define a learnable parameter matrix. For a line with Extract the line features of each node. The former Row, for each node 3D coordinate projection vector and its corresponding The positional encoding vectors are summed element by element to obtain the node feature vector for each node;

[0032] The pooling layer operations in a feature extraction network consist of four stages:

[0033] Preparation phase: Create a learnable vector, denoted as [CLS] token, whose dimensions are the same as the node feature dimensions. same;

[0034] Input phase: Before inputting the coordinate sequence into the Transformer encoder, insert the [CLS] token at the very beginning of the coordinate sequence;

[0035] Processing phase: The coordinate sequence containing [CLS] is passed through the Transformer encoder. During the self-attention mechanism, the [CLS] token interacts with the node feature vectors of all nodes in the coordinate sequence, aggregating the context information of the entire coordinate sequence.

[0036] Output stage: After the Transformer encoder outputs, the output of the first position, which is the vector corresponding to the [CLS] token, is taken as the representation of the entire coordinate sequence.

[0037] The specific steps of step 13 in the above-mentioned deep learning-based zero-watermarking method for linear vector map data are as follows:

[0038] Step 131: Collect line feature data for different geographic entity types; for each line feature... Constructing positive samples and negative samples Construct triples based on line features, positive samples, and negative samples. , , );

[0039] Step 132: Select the triplet loss function. Its goal is to make and The distance is much smaller than and The distance is given by the following formula:

[0040] ;

[0041] in, It is a distance function; Represents a constant greater than 0; This is a function to find the maximum value.

[0042] Step 133: , , Each fixed-length global feature vector is obtained through a Transformer encoder. , , , Indicates and The corresponding global feature vector, Indicates and The corresponding global feature vector, Indicates and The corresponding global feature vector; , , The loss is calculated by feeding the triplet loss function, then backpropagation is performed to calculate the gradient; then gradient clipping is performed, the optimizer updates the parameters, the gradient is cleared to zero, and the next batch of iterations begins.

[0043] The specific steps of step 2 in the above-mentioned deep learning-based zero-watermarking method for linear vector map data are as follows:

[0044] Step 21, Feature-level feature extraction: Based on the feature extraction network, high-dimensional features are extracted for each line feature in the dataset, resulting in a high-dimensional feature set of all line features in the dataset. , , Indicates the first High-dimensional features of individual line elements The total number of line elements;

[0045] Step 22, Feature Aggregation: Based on the feature aggregation network, the high-dimensional features of all line features are aggregated into a global feature vector representing the entire dataset. Attention pooling is then used to calculate the attention score for each line feature, as shown in the following formula:

[0046] ;

[0047] in, For the first Attention score of each line element This is the weight matrix. For bias vectors, For the weight vector, It is the hyperbolic tangent function. For the first High-dimensional characteristics of individual line elements;

[0048] calculate The formula is as follows:

[0049] ;

[0050] in, For the first The weight of the attention score of each line element relative to the total attention scores of all line elements. It is an exponential function. For the summation function, For the first Attention score of each line element express The value ranges from 1 to ;

[0051] The global feature vector of the dataset is obtained by weighted summation. The formula is as follows:

[0052] ;

[0053] in, This represents the global feature vector of the dataset. express The value ranges from 1 to .

[0054] The specific steps of step 3 in the above-mentioned deep learning-based zero-watermarking method for linear vector map data are as follows:

[0055] Step 31, Modify the model architecture: Add a fully connected layer as a hash layer after the existing feature extraction network and feature aggregation network. The output dimension of the hash layer is the length of the binary sequence.

[0056] Step 32, Design the loss function Loss function It consists of two parts: similarity preservation loss and quantization loss;

[0057] Step 33, Model Training: Use pre-trained feature extraction and feature aggregation networks as a foundation, and randomly initialize the hash layer; fix the parameters of the feature extraction and feature aggregation networks, and use the loss function. Conduct training;

[0058] Step 34, Generate Binary Sequence: After training, for new vector map data, the output vector of the hash layer is obtained through forward propagation of the feature extraction network, feature aggregation network, and hash layer; the binary sequence is obtained by applying the sign function.

[0059] Step 35, perform discrimination evaluation: use Hamming distance to compare binary sequences, and evaluate discrimination by calculating the ratio of intra-class distance to inter-class distance.

[0060] In the aforementioned deep learning-based zero-watermarking method for linear vector map data, step 32 involves a similarity preservation loss. This ensures that similar samples have similar hash codes, while dissimilar samples have dissimilar hash codes. For a pair of samples, if they belong to the same copyright, their hash code distance is minimized; if they belong to different copyrights, their hash code distance is maximized. The formula is as follows:

[0061] ;

[0062] in, Numerical labels for sample pairs, Indicates positive sample pairs. Indicates negative sample pairs; It is the first The hash layer output vector of each sample; It is the first The hash layer output vector of each sample; It is the margin parameter; It is a distance function;

[0063] Quantifying loss To encourage hash layer outputs to be close to binary values ​​and reduce information loss during binarization, the formula is as follows:

[0064] ;

[0065] in, It is a vector consisting entirely of 1s; express The absolute value; Represents the square of the L2 norm;

[0066] The calculation formula is as follows:

[0067]

[0068] in, This is a hyperparameter.

[0069] In the aforementioned deep learning-based zero-watermarking method for linear vector map data, step 34, the formula for obtaining the binary sequence using the sign function is as follows:

[0070] ;

[0071] in, It is a binary sequence The Dimensional value, It is the first The hash layer output vector of the nth sample Dimensional value.

[0072] The specific steps of step 4 in the above-mentioned deep learning-based zero-watermarking method for linear vector map data are as follows:

[0073] Step 41: Convert the binary sequence The sequences are concatenated to form a 256×256 long sequence, creating a 256×256 two-dimensional matrix. ;

[0074] Step 42: Generate zero watermark , ;in This is an XOR operation; This is a binary copyright image, with a size of 256×256; It is a two-dimensional matrix with a size of 256×256;

[0075] Step 43: Remove the zero watermark Registered with the copyright center and stored in the copyright center's database.

[0076] The beneficial effects of this invention are as follows:

[0077] 1. This invention abandons the traditional approach of modifying data and learns deep features of data as digital fingerprints by simulating human cognition. Under the premise of ensuring zero distortion of the original data, it achieves strong distinguishability and robust copyright authentication of linear vector map data, providing a new technical solution for the protection of geographic data property rights in an open and shared environment. It is very suitable for the security protection of vector geographic data.

[0078] 2. Compared with the prior art, the present invention has the advantages of being absolutely non-destructive, having strong distinguishing ability, excellent robustness, high security, and good practicality;

[0079] Absolutely lossless: It uses zero-watermark technology, and the original data remains unchanged throughout the entire process, perfectly maintaining the geometric accuracy and attribute integrity of the vector data;

[0080] Strong discriminative ability: By using deep learning to automatically learn the deep semantic features and stable structural features of data, its discriminative ability far exceeds that of traditional features based on manual rules, and it can effectively identify highly similar data from different sources.

[0081] Excellent robustness: By focusing on key features through adversarial training (data augmentation) and attention mechanisms, the generated hash codes exhibit excellent stability against common geometric attacks (vertex addition / deletion, coordinate perturbation, data simplification), format conversion, and pruning operations that do not affect the main structure.

[0082] High security: Copyright information is hidden within the zero watermark, and the generation of the zero watermark depends on the characteristics of the data itself. Attackers cannot deduce the other party from the data or the zero watermark alone; verification requires both the data and the correct zero watermark.

[0083] High practicality: The binary hash code and zero watermark are small in size, making them easy to store and transmit. The copyright detection process involves efficient forward computation, making it suitable for rapid copyright screening and verification of large-scale datasets. Attached Figure Description

[0084] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0085] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0086] like Figure 1 As shown, the deep learning-based zero-watermarking method for linear vector map data includes the following steps:

[0087] Step 1, Line Feature Deep Extraction: Based on the transformer network, learn the deep features of line features to extract the high-dimensional features of line features in the dataset of the line vector map.

[0088] The specific process of step 1 is as follows:

[0089] Step 11, Data Preprocessing and Augmentation: Represent each line feature in the dataset as a series of node coordinates to obtain a variable-length coordinate sequence, and normalize the node coordinates.

[0090] The specific process of step 11 is as follows:

[0091] Step 111: For a line feature, calculate the minimum X-coordinate of the line feature across all nodes. and the maximum value of the X coordinate Calculate the minimum Y-coordinate of the line feature across all nodes. and the maximum value of the Y coordinate Get the coordinates of the bottom left corner of the bounding box. , ) and the coordinates of the upper right corner ( , );

[0092] Step 112: Calculate the width of the bounding box and the height of the enclosure The formula is as follows:

[0093] ;

[0094] ;

[0095] Step 113: Perform normalization calculations for the first... There are 1 node, and the formula is as follows:

[0096] ;

[0097] ;

[0098] in, For the first The normalized x-coordinates of each node For the first The x-coordinate of each node, For the first The normalized ordinates of each node For the first The ordinate of each node;

[0099] Step 114: During training, randomly apply operations such as adding nodes, deleting nodes, and perturbing nodes to perform data augmentation.

[0100] Step 12, Construct a feature extraction network: Construct a feature extraction network based on a Transformer encoder to process coordinate sequences and capture global dependencies;

[0101] The feature extraction network consists of an input embedding layer, a Transformer encoder, and a pooling layer:

[0102] The input embedding layer projects the coordinates of each node into a high-dimensional space through a linear layer and adds learnable positional encodings; the specific steps of the linear projection in the input embedding layer are as follows:

[0103] The specific steps in step 12 are as follows:

[0104] First, a linear layer is used to project the two-dimensional coordinates onto... 3D space For the node feature dimension, for the ... The coordinate vector of each node The linear layer performs the following calculations, as shown in the formula below:

[0105] ;

[0106] in, It is a learnable weight matrix; It is a learnable bias vector; It is a two-dimensional coordinate projection onto The high-dimensional projection vector obtained from the 1D space;

[0107] Next, add learnable positional encoding. Specifically, define a learnable parameter matrix. For a line with Extract the line features of each node. The former Row, for each node 3D coordinate projection vector and its corresponding The positional encoding vectors are summed element by element to obtain the node feature vector for each node.

[0108] The Transformer encoder consists of a multi-layer self-attention mechanism and a feedforward network.

[0109] Pooling layers aggregate information from the entire coordinate sequence, outputting a fixed-dimensional feature vector. The operations in the pooling layers of the feature extraction network include four stages:

[0110] Preparation phase: Create a learnable vector, denoted as [CLS] token, whose dimensions are the same as the node feature dimensions. same;

[0111] Input phase: Before inputting the coordinate sequence into the Transformer encoder, insert the [CLS] token at the very beginning of the coordinate sequence;

[0112] Processing phase: The coordinate sequence containing [CLS] is passed through the Transformer encoder. During the self-attention mechanism, the [CLS] token interacts with the node feature vectors of all nodes in the coordinate sequence, aggregating the context information of the entire coordinate sequence.

[0113] Output stage: After the Transformer encoder outputs, the output of the first position, which is the vector corresponding to the [CLS] token, is taken as the representation of the entire coordinate sequence.

[0114] Step 13, Training the feature extraction network: The training uses a triplet loss function to construct triplet pairs and uses gradient descent to optimize the network parameters.

[0115] The specific steps of step 13 are as follows:

[0116] Step 131: Collect line feature data for different geographic entity types; for each line feature... Constructing positive samples and negative samples Construct triples based on line features, positive samples, and negative samples. , , );

[0117] Step 132: Select the triplet loss function. Its goal is to make and The distance is much smaller than and The distance is given by the following formula:

[0118] ;

[0119] in, It is a distance function; Represents a constant greater than 0; This is a function to find the maximum value.

[0120] Step 133: , , Each fixed-length global feature vector is obtained through a Transformer encoder. , , , Indicates and The corresponding global feature vector, Indicates and The corresponding global feature vector, Indicates and The corresponding global feature vector; , , The loss is calculated by feeding the triplet loss function, then backpropagation is performed to calculate the gradient; then gradient clipping is performed, the optimizer updates the parameters, the gradient is cleared to zero, and the next batch of iterations begins.

[0121] Training techniques: using The optimizer employs a warm-up strategy and learning rate decay; it also implements hard sample mining, prioritizing negative samples that the current model misjudges.

[0122] Evaluation and Validation Phase: Reserve 20% of the training data as a validation set. Triples are also constructed on the validation set, and then the loss on the triples on the validation set is directly monitored.

[0123] Step 2, Feature aggregation based on attention mechanism: Through the feature aggregation network, the high-dimensional features of all line elements are aggregated into a global feature.

[0124] The specific steps of step 2 are as follows:

[0125] Step 21, Feature-level feature extraction: Based on the feature extraction network, high-dimensional features are extracted for each line feature in the dataset, resulting in a high-dimensional feature set of all line features in the dataset. , , Indicates the first High-dimensional features of individual line elements The total number of line elements;

[0126] Step 22, Feature Aggregation: Based on the feature aggregation network, the high-dimensional features of all line features are aggregated into a global feature vector representing the entire dataset. Attention pooling is then used to calculate the attention score for each line feature, as shown in the following formula:

[0127] ;

[0128] in, For the first Attention score of each line element This is the weight matrix. For bias vectors, For the weight vector, It is the hyperbolic tangent function. For the first High-dimensional characteristics of individual line elements;

[0129] calculate The formula is as follows:

[0130] ;

[0131] in, For the first The weight of the attention score of each line element relative to the total attention scores of all line elements. It is an exponential function. For the summation function, For the first Attention score of each line element express The value ranges from 1 to ;

[0132] The global feature vector of the dataset is obtained by weighted summation. The formula is as follows:

[0133] ;

[0134] in, This represents the global feature vector of the dataset. express The value ranges from 1 to .

[0135] Collect multiple vector map datasets, each with copyright tags, and construct positive sample pairs (belonging to the same copyright) and negative sample pairs (belonging to different copyrights). Using contrastive loss or triplet loss, train the entire feature aggregation network (feature extractor + attention pooling) as a large, end-to-end trainable model. Given a dataset, output the global feature vector for that dataset.

[0136] Step 3, Generate hash code: Through forward propagation of the hash layer, the global features are converted into binary hash code.

[0137] The specific steps of step 3 are as follows:

[0138] Step 31, Modify the model architecture: Add a fully connected layer as a hash layer after the existing feature extraction network and feature aggregation network. The output dimension of the hash layer is the length of the binary sequence. The hash layer uses an activation function (tanh) to restrict the output to the range of [-1, 1].

[0139] Step 32, Design the loss function Loss function It consists of two parts: similarity preservation loss and quantization loss.

[0140] In step 32, the similarity preservation loss This ensures that similar samples have similar hash codes, while dissimilar samples have dissimilar hash codes. For a pair of samples, if they belong to the same copyright, their hash code distance is minimized; if they belong to different copyrights, their hash code distance is maximized. The formula is as follows:

[0141] ;

[0142] in, Numerical labels for sample pairs, Indicates positive sample pairs. Indicates negative sample pairs; It is the first The hash layer output vector of each sample; It is the first The hash layer output vector of each sample; It is the margin parameter; It is a distance function;

[0143] Quantifying loss To encourage hash layer outputs to be close to binary values ​​and reduce information loss during binarization, the formula is as follows:

[0144] ;

[0145] in, It is a vector consisting entirely of 1s; express The absolute value; Represents the square of the L2 norm;

[0146] The calculation formula is as follows:

[0147]

[0148] in, This is a hyperparameter.

[0149] Step 33, Model Training: Use pre-trained feature extraction and feature aggregation networks as a foundation, and randomly initialize the hash layer; fix the parameters of the feature extraction and feature aggregation networks, and use the loss function. During training, select Adam or SGD as the optimizer and appropriately reduce the learning rate (e.g., 0.001).

[0150] Step 34, Generate binary sequence: After training, for new vector map data, the output vector of the hash layer is obtained through forward propagation of the feature extraction network, feature aggregation network and hash layer; the binary sequence is obtained by applying the symbol function.

[0151] The formula for obtaining a binary sequence using the sign function is:

[0152] ;

[0153] in, It is a binary sequence The Dimensional value, It is the first The hash layer output vector of the nth sample Dimensional value.

[0154] Step 35, perform a discrimination assessment: use Hamming distance to compare binary sequences. Map data with the same copyright should have a smaller Hamming distance, while data with different copyrights should have a larger Hamming distance. Discrimination can be assessed by calculating the ratio of intra-class distance to inter-class distance.

[0155] Step 4, Generate Zero Watermark: Generate a zero watermark based on the XOR operation between the binary hash code and the copyrighted image, and store it in the copyright center's database.

[0156] The specific steps of step 4 are as follows:

[0157] Step 41: Convert the binary sequence The sequences are concatenated to form a 256×256 long sequence, creating a 256×256 two-dimensional matrix. ;

[0158] Step 42: Generate zero watermark , ;in This is an XOR operation; This is a binary copyright image, with a size of 256×256; It is a two-dimensional matrix with a size of 256×256;

[0159] Step 43: Remove the zero watermark Registered with the copyright center and stored in the copyright center's database.

[0160] During copyright detection, firstly, for a given vector map data, the forward propagation of the aforementioned feature extraction network, feature aggregation network, and hash layer yields the output vector of the hash layer. Then, a sign function is applied to obtain a new binary sequence. ;

[0161] Then, the new binary sequence The sequences are pieced together to form a new 256×256 two-dimensional matrix. ;

[0162] Next, the relevant zero-watermark data is retrieved from the copyright database. The new copyright image is calculated. , ,in, For a new two-dimensional matrix;

[0163] Finally, based on the new copyrighted image The clarity of the image is used to determine the copyright of the data. This is used to determine the copyright of new images. The method for determining the clarity is: directly observe the reconstructed new copyright image. Clarity and structural integrity; or calculation The structural similarity index with the standard copyrighted image is used; if the similarity exceeds a preset threshold, the copyright is determined to be matched.

Claims

1. A method for zero-watermarking linear vector map data based on deep learning, characterized in that, Includes the following steps: Step 1, Line Feature Deep Extraction: Based on the transformer network, learn the deep features of line features to extract the high-dimensional features of line features in the dataset of the line vector map; Step 2, Feature aggregation based on attention mechanism: Through the feature aggregation network, the high-dimensional features of all line elements are aggregated into a global feature; Step 3, Generate hash code: Through forward propagation of the hash layer, the global features are converted into binary hash code; Step 4, Generate Zero Watermark: Generate a zero watermark based on the XOR operation between the binary hash code and the copyrighted image, and store it in the copyright center's database.

2. The method for zero-watermarking linear vector map data based on deep learning according to claim 1, characterized in that, The specific process of step 1 is as follows: Step 11, Data Preprocessing and Augmentation: Represent each line feature in the dataset as a series of node coordinates to obtain a variable-length coordinate sequence, and normalize the node coordinates. Step 12, Construct a feature extraction network: Construct a feature extraction network based on a Transformer encoder to process coordinate sequences and capture global dependencies; The feature extraction network consists of an input embedding layer, a Transformer encoder, and a pooling layer: The input embedding layer projects the coordinates of each node into a high-dimensional space through a linear layer and adds learnable positional encoding; The Transformer encoder consists of a multi-layer self-attention mechanism and a feedforward network; The pooling layer aggregates the information of the entire coordinate sequence and outputs a feature vector of fixed dimensions. Step 13, Training the feature extraction network: The training uses a triplet loss function to construct triplet pairs and uses gradient descent to optimize the network parameters.

3. The method for zero-watermarking linear vector map data based on deep learning according to claim 2, characterized in that, The specific process of step 11 is as follows: Step 111: For a line feature, calculate the minimum X-coordinate of the line feature across all nodes. and the maximum value of the X coordinate Calculate the minimum Y-coordinate of the line feature across all nodes. and the maximum value of the Y coordinate Get the coordinates of the bottom left corner of the bounding box. , ) and the coordinates of the upper right corner ( , ); Step 112: Calculate the width of the bounding box and the height of the enclosure The formula is as follows: ； ； Step 113: Perform normalization calculations for the first... There are 1 node, and the formula is as follows: ； ； in, For the first The normalized x-coordinates of each node For the first The x-coordinate of each node For the first The normalized ordinates of each node For the first The ordinate of each node; Step 114: During training, randomly apply operations such as adding nodes, deleting nodes, and perturbing nodes to perform data augmentation.

4. The method for zero-watermarking linear vector map data based on deep learning according to claim 3, characterized in that, The specific steps in step 12 are as follows: First, a linear layer is used to project the two-dimensional coordinates onto... 3D space For the node feature dimension, for the ... The coordinate vector of each node The linear layer performs the following calculations, as shown in the formula below: ； in, It is a learnable weight matrix; It is a learnable bias vector; It is a two-dimensional coordinate projection onto The high-dimensional projection vector obtained from the 1D space; Next, add learnable positional encoding. Specifically, define a learnable parameter matrix. For a line with Extract the line features of each node. The former Row, for each node 3D coordinate projection vector and its corresponding The positional encoding vectors are summed element by element to obtain the node feature vector for each node; The pooling layer operations in a feature extraction network consist of four stages: Preparation phase: Create a learnable vector, denoted as [CLS] token, whose dimensions are the same as the node feature dimensions. same; Input phase: Before inputting the coordinate sequence into the Transformer encoder, insert the [CLS] token at the very beginning of the coordinate sequence; Processing phase: The coordinate sequence containing [CLS] is passed through the Transformer encoder. During the self-attention mechanism, the [CLS] token interacts with the node feature vectors of all nodes in the coordinate sequence, aggregating the context information of the entire coordinate sequence. Output stage: After the Transformer encoder outputs, the output of the first position, which is the vector corresponding to the [CLS] token, is taken as the representation of the entire coordinate sequence.

5. The method for zero-watermarking linear vector map data based on deep learning according to claim 4, characterized in that, The specific steps of step 13 are as follows: Step 131: Collect line feature data for different geographic entity types; for each line feature... Constructing positive samples and negative samples ; Construct triples based on line features, positive samples, and negative samples. , , ); Step 132: Select the triplet loss function. Its goal is to make and The distance is much smaller than and The distance is given by the following formula: ； in, It is a distance function; Represents a constant greater than 0; This is a function to find the maximum value. Step 133: , , Each fixed-length global feature vector is obtained through a Transformer encoder. , , , Indicates and The corresponding global feature vector, Indicates and The corresponding global feature vector, Indicates and The corresponding global feature vector; , , The loss is calculated by feeding the triplet loss function, then backpropagation is performed to calculate the gradient; then gradient clipping is performed, the optimizer updates the parameters, the gradient is cleared to zero, and the next batch of iterations begins.

6. The method for zero-watermarking linear vector map data based on deep learning according to claim 5, characterized in that, The specific steps of step 2 are as follows: Step 21, Feature-level feature extraction: Based on the feature extraction network, high-dimensional features are extracted for each line feature in the dataset, resulting in a high-dimensional feature set of all line features in the dataset. , , Indicates the first High-dimensional features of individual line elements The total number of line elements; Step 22, Feature Aggregation: Based on the feature aggregation network, the high-dimensional features of all line elements are aggregated into a global feature vector representing the entire dataset. An attention pooling mechanism is then used to calculate the attention score for each line element, as shown in the following formula: ； in, For the first Attention score of each line element This is the weight matrix. For bias vectors, For the weight vector, It is the hyperbolic tangent function. For the first High-dimensional characteristics of individual line elements; calculate The formula is as follows: ； in, For the first The weight of the attention score of each line element relative to the total attention scores of all line elements. It is an exponential function. For the summation function, For the first Attention score of each line element express The value ranges from 1 to ; The global feature vector of the dataset is obtained by weighted summation. The formula is as follows: ； in, This represents the global feature vector of the dataset. express The value ranges from 1 to .

7. The method for zero-watermarking linear vector map data based on deep learning according to claim 6, characterized in that, The specific steps of step 3 are as follows: Step 31, Modify the model architecture: Add a fully connected layer as a hash layer after the existing feature extraction network and feature aggregation network. The output dimension of the hash layer is the length of the binary sequence. Step 32, Design the loss function Loss function It consists of two parts: similarity preservation loss and quantization loss; Step 33, Model Training: Use pre-trained feature extraction and feature aggregation networks as a foundation, and randomly initialize the hash layer; fix the parameters of the feature extraction and feature aggregation networks, and use the loss function. Conduct training; Step 34, Generate Binary Sequence: After training, for new vector map data, the output vector of the hash layer is obtained through forward propagation of the feature extraction network, feature aggregation network, and hash layer; the binary sequence is obtained by applying the sign function. Step 35, perform discrimination evaluation: use Hamming distance to compare binary sequences, and evaluate discrimination by calculating the ratio of intra-class distance to inter-class distance.

8. The method for zero-watermarking linear vector map data based on deep learning according to claim 7, characterized in that, In step 32, the similarity preservation loss This ensures that similar samples have similar hash codes, while dissimilar samples have dissimilar hash codes. For a pair of samples, if they belong to the same copyright, their hash code distance is minimized; if they belong to different copyrights, their hash code distance is maximized. The formula is as follows: ； in, Numerical labels for sample pairs, Indicates a positive sample pair. Indicates negative sample pairs; It is the first The hash layer output vector of each sample; It is the first The hash layer output vector of each sample; It is the margin parameter; It is a distance function; Quantifying loss To encourage hash layer outputs to be close to binary values ​​and reduce information loss during binarization, the formula is as follows: ； in, It is a vector consisting entirely of 1s; express The absolute value; Represents the square of the L2 norm; The calculation formula is as follows: ； in, This is a hyperparameter.

9. The method for zero-watermarking linear vector map data based on deep learning according to claim 8, characterized in that, In step 34, the formula for obtaining the binary sequence by applying the sign function is: ； in, It is a binary sequence The Dimensional value, It is the first The hash layer output vector of the nth sample Dimensional value.

10. The method for zero-watermarking linear vector map data based on deep learning according to claim 9, characterized in that, The specific steps of step 4 are as follows: Step 41: Convert the binary sequence The sequences are concatenated to form a 256×256 long sequence, creating a 256×256 two-dimensional matrix. ; Step 42: Generate a zero watermark , ;in This is an XOR operation; This is a binary copyright image, with a size of 256×256; It is a two-dimensional matrix with a size of 256×256; Step 43: Remove the zero watermark Register with the copyright center and store information in the copyright center's database.

Images