Oracle bone structure identification method of glyph graph isomorphic network

Abstract

The invention belongs to the technical field of character pattern recognition, and particularly provides an oracle structure recognition method of a font pattern isomorphic network, which comprises three stages of font pattern structure feature extraction, font pattern isomorphic network model and oracle font structure recognition. Through oracle font skeleton extraction, oracle font skeleton singular point detection, skeleton burr and bifurcation optimization, and key point extraction as a node communication path as an edge, an oracle font graph structure is established, and graph structure nodes, edges and graph level features are calculated. The font graph isomorphic network integrates the advantages of a graph isomorphic network and dynamic edge and graph embedding, and learns vector representation of multi-level features such as nodes, edges and graphs of the oracle font graph structure so as to train and identify the font structure of the oracle. The method can be used for structure matching recognition of oracle character pattern images, has the advantage of being high in recognition accuracy, and is suitable for recognition of character pattern structures such as gold texts and seal scripts and recognition of social networks, biological information and molecular structures.

Description

Technical Field

[0001] This disclosure relates to the field of character image recognition technology, and in particular to a method for recognizing the structure of oracle bone script using a character image isomorphic network. Background Technology

[0002] With the rapid development of computer and artificial intelligence technologies, machine learning and data mining have shown broad application prospects in national economy, scientific research, and other fields. Machine learning, as a core technology assisting humans in efficiently processing and analyzing large-scale data, can significantly reduce manual workload, improve work efficiency and quality, and enhance the intelligence level of systems. Its greatest advantage lies in its ability to automatically extract valuable information from massive amounts of data without human intervention, thus avoiding human error. Furthermore, machine learning can tirelessly process and analyze data in large-scale data environments, consistently performing tasks, learning patterns and rules in the data, and making fast and effective decisions based on what it has learned. Traditional graph isomorphic network methods, for example, suffer from high computational complexity when processing large-scale graphs, leading to low computational efficiency. They also exhibit sensitivity to noise and outliers and poor scalability, making it difficult to meet the demands of modern applications for big data processing.

[0003] The character-graph isomorphic network recognition method constructs character graphs and uses graph isomorphism algorithms to detect oracle bone script structures. Based on features such as singular points of key skeleton points, graph structure nodes, edges, and layer levels, it determines whether the character matches the oracle bone script structure, reducing manual comparison workload and improving the efficiency and quality of oracle bone script recognition. This method can be applied to the recognition of ancient scripts such as oracle bone script and bronze inscriptions in the field of paleography, as well as the recognition of ancient books and inscriptions in the field of cultural heritage protection. It can significantly improve the performance of computers in textual research and protection, making computers more adaptable in character recognition and promoting the modernization and intelligentization of character recognition. In practical applications, it is necessary to extract the stroke and structural texture information of the characters so that the computer program can save the corresponding matching positions and images based on the recognition results.

[0004] The full-character shape structure matching and recognition of oracle bone script has significant application value in fields such as oracle bone script matching and recognition, and natural language processing. The identification of isomorphic and similar structures in oracle bone script is crucial for understanding the evolution of writing history and deciphering oracle bone script. Currently, oracle bone script shape structure matching and recognition suffers from drawbacks such as poor fine-grained structure recognition ability, weak feature vector discriminative power, and limited ability to adapt to complex topological structures. Summary of the Invention

[0005] This disclosure aims to at least solve one of the technical problems existing in the prior art, and proposes a method for recognizing the structure of oracle bone script using a glyph-graph isomorphic network, including:

[0006] S1, extract the skeleton of the oracle bone script, extract the singular points of the oracle bone script skeleton as nodes, and use the connectivity between adjacent nodes as edges to establish the oracle bone script graph structure.

[0007] S2, using a multi-layer message passing and aggregation mechanism, learns the vector representations of nodes, edges and layers of the oracle bone script graph structure to obtain a graph isomorphic network model;

[0008] S3. Extract the glyph graph structure features from the pre-constructed graph structure dictionary, train the glyph graph isomorphic network model, then use the glyph graph isomorphic network model to learn the vector representation of the glyph graph structure of oracle bone script characters, identify the oracle bone script category labels, use the graph structure dictionary matching method to verify the isomorphic glyphs of oracle bone script characters under the same label, and output the recognition results.

[0009] Preferably, S1 specifically includes:

[0010] The Zhang-Suen algorithm was used to perform skeletonization on the binary image of oracle bone script, and a skeleton image with a single pixel width was extracted.

[0011] Analyze the Moore eight-neighborhood of each pixel in the skeleton image, calculate the neighborhood difference metric, and identify singularities.

[0012] Preferably, the extraction of key points from the oracle bone script skeleton as nodes in step S1 specifically includes:

[0013] The singular points in the skeleton of oracle bone script include endpoints and intersections. Skeleton optimization is performed based on the pixel path length between endpoints and intersections, and between intersections. If the pixel path length between endpoints and intersections is lower than a threshold, the jagged edges are erased. For each pair of intersections, a circle is drawn with the coordinates of each intersection as the center and the maximum radius within the stroke as the circle radius. If the circles drawn between two intersections intersect, it is determined to be a skeleton fork, and it is erased and reconstructed.

[0014] Preferably, the method for establishing the oracle bone script character structure in S1 specifically includes the following steps:

[0015] Key points are extracted based on the optimized skeleton, and the key points are treated as nodes in the graph structure. Edge relationships are established between each pair of nodes through the adjacency connectivity in the skeleton, thus constructing an adjacency matrix.

[0016] Preferably, after step S1, the method further includes refining the structure of the oracle bone script characters, specifically including:

[0017] In the extracted oracle bone script structure, if an endpoint is connected to both other endpoints and intersections, delete the contradictory edges between endpoints; if the degree of an intersection does not satisfy the intersection property (the degree of an intersection must be greater than or equal to 3), add a self-loop edge.

[0018] Preferably, after step S1, the method further includes feature extraction of the oracle bone script character structure, specifically including:

[0019] Node features, edge features, and global graph structure features;

[0020] Graph structure node-level features include node degree, node type, and node coordinates;

[0021] The edge-level features of the graph structure include edge length, mean, maximum, minimum and standard deviation of edge curvature, mean and maximum of the derivative of curvature, and mean of the tangent angle of the edge.

[0022] Global graph-level features of a graph structure include the number of nodes, the number of inner contours, and the number of connected subgraphs.

[0023] Preferably, S2 specifically includes:

[0024] To ensure that the vector representation learned by the graph isomorphic network model achieves the graph structure discrimination capability of the WL test, a dynamic edge convolutional network is combined with the dynamic edge representation of the edges between nodes and the vector uniqueness of the graph structure accurately expressed by the graph isomorphic network. The dynamic weights of the gated layer are used to fuse the two and graph-level pooling is performed. The global graph-level features are then integrated into the perceptron classifier along with the graph embedding linear transformation.

[0025] Preferably, S2 specifically includes: in the character graph isomorphic network model, the graph structure of oracle bone script characters and the multi-level features of the graph structure are input, the model performs neighborhood information aggregation on each node, each time the node is aggregated, it obtains the information of the outermost neighboring nodes, and updates its own node information by combining the neighboring information. After the update is completed, graph-level pooling is performed on the graph structure to obtain the graph structure vector representation of the oracle bone script characters.

[0026] Preferably, S2 specifically includes:

[0027] The features of the central node are projected from the original feature space to an implicit higher-order semantic space, so that it already has task-related discriminative information before being fused with the features of its neighbors; and the injectivity of the node update function is guaranteed.

[0028] Preferably, S3 specifically includes:

[0029] Based on the oracle bone script character category labels output by the graph isomorphic network model, oracle bone script characters of the same category and structure are found in the graph structure dictionary. The number of shortlisted points is checked by using the coordinates of the corresponding nodes in the graph structure and the homography matrix, and the correctness of the graph structure recognition is determined.

[0030] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0031] 1. Fine-grained structure recognition capability: Existing technologies cannot effectively identify fine-grained basic levels such as key points, strokes, components, and structures of oracle bone script nodes. The character graph structure recognition network model proposed in this invention obtains the graph structure representation of oracle bone script characters through a designed graph structure extraction method and learns the feature vectors of the graph structure using a graph isomorphic network. This enables accurate recognition of the fine-grained structure of oracle bone script characters, overcoming the shortcomings of traditional methods in fine-grained structure recognition and providing more precise information for in-depth research on oracle bone script.

[0032] 2. Enhanced Discrimination of Feature Vectors: Mainstream graph neural networks have limitations in coarse-grained oracle bone script recognition and struggle to effectively learn and distinguish simple graph structures. The network architecture of this invention ensures superior discriminative power in feature vectors and guarantees the completeness of structural representation through graph isomorphism theory. This enables the model to better identify and distinguish different oracle bone script structures, improving the accuracy and reliability of recognition.

[0033] 3. Enhanced adaptability to complex topologies: Oracle bone script struggles to adapt to its complex and varied heterogeneous structures, such as variant characters and fragment completion scenarios. This invention employs a graph isomorphic network model that utilizes the multi-layer message passing mechanism of GIN to achieve hierarchical feature learning from local strokes to global component combinations. Furthermore, it introduces dynamic edge weights to adapt to noise interference from oracle bone rubbings, effectively addressing the complex and varied topologies of oracle bone script and improving the model's robustness and adaptability in complex scenarios.

[0034] 4. The experimental performance has reached the current best. This demonstrates the advanced nature and effectiveness of this invention in the field of oracle bone script structure recognition, and provides strong technical support for the digital research and intelligent processing of oracle bone script.

[0035] 5. Providing a technical path for oracle bone script recognition: The deep graph learning paradigm constructed in this invention, which has theoretical guarantees, breaks through the limitations of traditional methods and provides a new solution and new ideas for the structural reconstruction accuracy and semantic interpretability of oracle bone script. Attached Figure Description

[0036] Figure 1 This invention provides a method for recognizing the structure of oracle bone script using a glyph-image isomorphic network.

[0037] Figure 2 A schematic diagram of the GGIN architecture provided for an embodiment of the present invention. Detailed Implementation

[0038] To enable those skilled in the art to better understand the technical solutions of this disclosure, the disclosure will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] In the various figures, the same elements are represented by similar reference numerals. For clarity, not all parts in the figures are drawn to scale. Furthermore, some well-known parts may not be shown in the figures.

[0040] Many specific details of this disclosure, such as the structure, materials, dimensions, processing methods, and techniques of the components, are described below to provide a clearer understanding of the disclosure. However, as those skilled in the art will understand, this disclosure may be implemented without following these specific details.

[0041] Combination Figure 1 and Figure 2 This invention provides a method for recognizing the structure of oracle bone script using a glyph-image isomorphic network, comprising the following steps:

[0042] S1. Character graph structure extraction stage: Extract the skeleton of oracle bone characters, detect singular points of the skeleton of oracle bone characters, purify them into key points of the character graph skeleton, use key points as nodes, determine the connectivity between adjacent key points and use them as edges, establish the oracle bone character graph structure and extract graph structure nodes, edges and layer-level features.

[0043] S2, Character-Graph Isomorphic Network Model: Combining the advantages of graph isomorphic networks, dynamic edge networks, graph-level feature embedding, and gating layers, it learns vector representations of multi-level features such as nodes, edges, and graphs in the oracle bone character graph structure through multi-layer message passing and aggregation mechanisms.

[0044] S3. Oracle Bone Script Character Structure Recognition: Extract the character graph structure features from the oracle bone script character dataset, train the character graph isomorphic network model, learn the unique vector representation of the oracle bone script character graph structure and identify the oracle bone script category labels, and use the graph structure dictionary matching method to verify the isomorphic oracle bone script characters under the same label.

[0045] The method for establishing the structure of oracle bone script characters in step S1 specifically includes the following steps:

[0046] The Zhang-Suen algorithm is used to skeletonize the binary image of oracle bone script characters, extracting a single-pixel-width skeleton map. Then, by analyzing the Moore's eight-neighborhood (mod 8) of each pixel in the skeleton map, a neighborhood difference metric is calculated to identify endpoints and intersections as nodes in the graph structure. For each foreground pixel p = (x, y), the eight-neighborhood N8(p) = {n0,...,n7} is arranged clockwise. The neighborhood change metric is calculated, where nk ∈ {0,1} represents the neighboring pixel value, and |∗| represents the absolute value. Based on this, an adjacency matrix is ​​constructed according to the connectivity between nodes, forming the graph structure of the oracle bone script characters. This method verifies the connectivity between nodes through region blocking and path search, and optimizes the graph structure through strategies such as endpoint conflict elimination and loop detection, thereby achieving efficient representation of the topological features of the oracle bone script character graph. The resulting oracle bone script character graph structure is as follows:

[0047]

[0048] The method for optimizing the oracle bone script skeleton described in step S1 specifically includes the following steps:

[0049] First, distance detection and processing are needed for endpoints and intersections. For each endpoint, find its nearest intersection and calculate the distance between them. If the distance is greater than a set threshold, the endpoint is retained; otherwise, it is deleted. Simultaneously, for adjacent intersections, calculate their distances. If the distance is less than another set threshold, merge the two intersections into one. This process effectively simplifies the graphic structure, removes redundant endpoints and intersections, and thus optimizes the complexity and clarity of the graphic. The formulas for calculating the distance between endpoints and intersections, and the distance between intersections, are as follows:

[0050] The formula for calculating the distance between an endpoint and an intersection is: where pend is the endpoint coordinate and plink is the coordinate of the nearest intersection. If dspur is greater than the threshold t, the endpoint is retained; otherwise, it is deleted.

[0051]

[0052] The formula for calculating the distance between intersection points is as follows: The radius R(v) of the strokes in v ∈ L is defined as the radius of the largest inscribed circle centered at the coordinate system. The adjacent intersections vi and vj of coordinates pi and pj are merged when the following formula is satisfied, where θmerge represents the merging threshold.

[0053]

[0054] The method for recognizing oracle bone script structure using a isomorphic network of character shapes described in step S1 is characterized by the following steps after the character shape structure is extracted in S1:

[0055] For an endpoint vi ∈ V, if its neighborhood set N(vi) contains both endpoint vj and intersection vk ∈ L, then the contradictory edge is deleted.

[0056]

[0057] If the degree of an intersection does not satisfy the intersection property (the degree of an intersection must be greater than or equal to 3), then a self-loop edge is added. The formula is as follows:

[0058]

[0059] The preferred scheme, the method for extracting the structural features of oracle bone script characters in step S1 includes the following steps:

[0060] By using node feature vectors Defined as a concatenation of multimodal features, where W × H represents the image size, (x, y) are the coordinates of node vi, f represents the type discrimination function, and OneHot represents the class coding method.

[0061]

[0062] Feature eij integrates geometric and morphological features, and φ(eij) is the edge type encoding from vi to vj. Here, α represents a line segment. The angle between the coordinate axes, and These represent the coordinates of nodes vi and vj, respectively. It is the maximum Euclidean distance, Npix measures the pixel path length normalized by the Manhattan distance upper bound Lmax = W + H, κ and κ′ characterize the curvature and its derivative along the edge eij, θ represents the chamfer angle at each edge point, and w represents the stroke width along the path. Graph feature encoding: Graph-level features The concatenation of the original topological features is as follows:

[0063]

[0064] For each oracle bone script image structure, define a graph-level feature vector. The formula is:

[0065]

[0066] hWL represents the quantized WL graph hash sequence, C represents the number of connected components, L represents the number of cycles, Eend represents the number of endpoints, and Ecross represents the number of intersections. All of these belong to N.

[0067] As a preferred technical solution of the present invention, the method for recognizing the oracle bone script structure using a isomorphic network of character shapes is characterized in that, after extracting the character shape structure of S1, the method for improving the oracle bone script character shape structure specifically includes the following steps:

[0068] First, in the extracted oracle bone script structure, if an endpoint is connected to both other endpoints and intersections, delete the contradictory edges between endpoints; if the degree of an intersection does not satisfy the intersection property (the degree of an intersection must be greater than or equal to 3), add a self-loop edge.

[0069] As a preferred technical solution of the present invention, the method for recognizing oracle bone script structure using a character-graph isomorphic network is characterized in that, in the character-graph structure extraction of S1, the method for extracting the structural features of oracle bone script characters includes the following steps:

[0070] The graph structure features of oracle bone script characters include three aspects: node features, edge features, and global graph structure features. Node-level features include node degree, node type, and node coordinates; edge-level features include edge length, average, maximum, minimum, and standard deviation of edge curvature, average and maximum of the derivative of curvature, and average tangent angle of the edge; global graph-level features include the number of nodes, the number of inner contours, and the number of connected subgraphs.

[0071] The preferred solution, the isomorphic network model of the glyph graph described in step S2, specifically includes the following steps:

[0072] To ensure that the vector representation learned by the graph isomorphic network model achieves the graph structure discrimination capability of the WL test, a dynamic edge convolutional network is combined with the dynamic edge representation learned between nodes and the uniqueness of the vector representation accurately representing the graph structure. A gated layer with dynamic weights fuses both and performs graph-level pooling. This, along with a graph embedding linear transformation, integrates global graph-level features into the perceptron classifier. The formula is as follows:

[0073]

[0074] In the preferred embodiment, step S2, learning the vector representation of the multi-level features of the oracle bone script character structure, includes the following steps: projecting the features of the central node from the original feature space to an implicit higher-order semantic space, so that it already possesses task-related discriminative information before fusing with neighboring aggregated features. Finally, ensuring the injectivity of the node update function. The formula is as follows:

[0075]

[0076] Furthermore, in order to achieve the graph structure discrimination capability of the WL test by learning the vector representation of the graph isomorphic network model, the dynamic edge convolutional network is combined with the dynamic representation of the edges between nodes and the vector uniqueness of the graph structure accurately expressed by the graph isomorphic network. The dynamic weights of the gated layer are used to fuse the two and graph-level pooling is performed. The global graph-level features are then integrated into the perceptron classifier along with the graph embedding linear transformation.

[0077] The preferred scheme, the isomorphism verification method for oracle bone script characters in step S3, specifically includes the following steps:

[0078] Based on the general approximation theorem of multilayer perceptrons, the structure vector representations of all oracle bone script characters in a pre-constructed graph structure dictionary are input, and a function is learned to convert the structure vector representations of the oracle bone script characters into their graph structure category labels. The formula is as follows:

[0079]

[0080] First, based on the category labels output by the model, a set of matching oracle bone script character graph structures is selected from a pre-built graph structure dictionary. Then, this set is further filtered to select graph structures with the same topological structure as the input graph structure, ensuring not only category label matching but also graph structure consistency. Next, for each matching graph structure, the number of matching points is checked by comparing the coordinates of corresponding nodes in the input graph structure and the dictionary graph structure. If the number of matching point pairs exceeds a preset threshold, the graph structure is considered correctly identified; otherwise, it is considered incorrectly identified. This process fully utilizes the category and topological information in the graph structure dictionary, further ensuring the accuracy of the recognition results through geometric feature matching, thereby improving the reliability and accuracy of oracle bone script character recognition.

[0081] The method of this invention was simulated and verified on a PC, and HWOBC and the selected Oracle-50K dataset were selected for experimental verification.

[0082] The main testing platforms are as follows:

[0083] The computing platform consists of an Intel® Core™ i9-9900X CPU (3.50 GHz) and four NVIDIA GeForce RTX 2080 Ti graphics processors (each with 11GB of GDDR6 video memory).

[0084] The experiment was implemented using the PyTorch 2.5.0 deep learning framework and the CUDA 12.4 environment.

[0085] This experiment first sets the random seed for all random actions to 42. After standardization and preprocessing, the dataset is divided into training, validation, and test sets in a stratified random sampling strategy at a ratio of 70%:20%:10%, effectively maintaining the balance of sample distribution across classes. A dynamic early stopping mechanism is applied during model training: training is terminated when the classification accuracy on the validation set fails to improve for 30 consecutive training epochs, while the model with the best performance on the validation set is retained as the final model. Top-1 accuracy, precision, recall, and F1-score are collected for each model. For optimization algorithms, the Adam optimizer is used for parameter updates, and the cross-entropy loss function is used as the optimization objective for model training.

[0086] The following conclusions can be drawn from the analysis:

[0087] The proposed character graph isomorphic network model method has achieved advanced results in structural recognition experiments on the HWOBC dataset and Oracle-50K dataset, and is applicable to oracle bone script recognition.

[0088] The use of isomorphic network models to identify singularities in oracle bone script has improved both the efficiency and quality of recognition.

[0089] 3. The isomorphic network model algorithm based on character shapes is fast and has good real-time performance. It can also accurately identify complex oracle bone script structures, reducing the error rate.

[0090] 4. This method is highly adaptable. It is applicable to the identification and processing of variant characters and variant forms in ancient scripts such as oracle bone script and bronze inscriptions, as well as the study of calligraphy.

[0091] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of this disclosure, and this disclosure is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of this disclosure, and these modifications and improvements are also considered to be within the scope of protection of this disclosure.

Claims

1. A method for recognizing the structure of oracle bone script using a grapheme isomorphic network, characterized in that, include: S1. Extract the skeleton of the oracle bone script, extract the key points of the skeleton of the oracle bone script as nodes, and use the connectivity between adjacent nodes as edges to establish the structure of the oracle bone script graph. S2, using a multi-layer message passing and aggregation mechanism, learns the vector representations of nodes, edges and layers of the oracle bone script graph structure to obtain a graph isomorphic network model; S3. Extract the glyph graph structure features from the pre-constructed graph structure dictionary, train the glyph graph isomorphic network model, then use the glyph graph isomorphic network model to learn the vector representation of the glyph graph structure of oracle bone script characters, identify the oracle bone script category labels, use the graph structure dictionary matching method to verify the isomorphic glyphs of oracle bone script characters under the same label, and output the recognition results.

2. The method for recognizing oracle bone script structure using a isomorphic network of character shapes according to claim 1, characterized in that, S1 specifically includes: The Zhang-Suen algorithm was used to perform skeletonization on the binary image of oracle bone script, and a skeleton image with a single pixel width was extracted. Analyze the Moore eight-neighborhood of each pixel in the skeleton image, calculate the neighborhood difference metric, and identify singularities.

3. The method for recognizing oracle bone script structure using a glyph-image isomorphic network according to claim 1, characterized in that, The key points for extracting the skeleton of oracle bone script characters as nodes in S1 specifically include: The singular points in the skeleton of oracle bone characters include endpoints and intersections. Skeleton optimization is performed based on the pixel path length between endpoints and intersections, and between intersections. If the pixel path length between endpoints and intersections is lower than a threshold, the jagged edges are erased. For each pair of intersections, a circle is drawn with the coordinates of each intersection as the center and the maximum radius within the stroke as the circle radius. If the circles drawn between two intersections intersect, it is judged as an abnormal bifurcation of the skeleton, and it is erased and reconstructed.

4. The method for recognizing oracle bone script structure using a isomorphic network of character shapes according to claim 1, characterized in that, The method for establishing the oracle bone script graphic structure in S1 specifically includes the following steps: Key points are extracted based on the optimized skeleton, and the key points are treated as nodes in the graph structure. Edge relationships are established between each pair of nodes through the adjacency connectivity in the skeleton, thus constructing an adjacency matrix.

5. The method for recognizing oracle bone script structure using a glyph-based isomorphic network according to claim 1, characterized in that, The step S1 is followed by further refinement of the oracle bone script graphic structure, specifically including: In the extracted oracle bone script structure, if an endpoint is connected to both other endpoints and intersections, delete the contradictory edges between endpoints; if the degree of an intersection does not satisfy the intersection property (the degree of an intersection must be greater than or equal to 3), add a self-loop edge.

6. The method for recognizing oracle bone script structure using a glyph-image isomorphic network according to claim 1, characterized in that, The step S1 further includes feature extraction of the oracle bone script character structure, specifically including: Node features, edge features, and global graph structure features; Graph structure node-level features include node degree, node type, and node coordinates; The edge-level features of the graph structure include edge length, mean, maximum, minimum and standard deviation of edge curvature, mean and maximum of the derivative of curvature, and mean of the tangent angle of the edge. Global graph-level features of a graph structure include the number of nodes, the number of inner contours, and the number of connected subgraphs.

7. The method for recognizing oracle bone script structure using a glyph-image isomorphic network according to claim 1, characterized in that, S2 specifically includes: To ensure that the vector representation learned by the graph isomorphic network model achieves the graph structure discrimination capability of the WL test, a dynamic edge convolutional network is combined with the dynamic edge representation of the edges between nodes and the vector uniqueness of the graph structure accurately expressed by the graph isomorphic network. The dynamic weights of the gated layer are used to fuse the two and graph-level pooling is performed. The global graph-level features are then integrated into the perceptron classifier along with the graph embedding linear transformation.

8. The method for recognizing oracle bone script structure using a glyph-image isomorphic network according to claim 1, characterized in that, S2 specifically includes: In the graph isomorphic network model, the input is the graph structure of oracle bone script characters and the multi-level features of the graph structure. The model aggregates neighborhood information for each node. Each time a node is aggregated, it obtains information about the outermost neighboring nodes and updates its own node information by combining the neighboring information. After the update is completed, graph-level pooling is performed on the graph structure to obtain the graph structure vector representation of the oracle bone script characters.

9. The method for recognizing oracle bone script structure using a glyph-image isomorphic network according to claim 1, characterized in that, S2 specifically includes: The features of the central node are projected from the original feature space to an implicit higher-order semantic space, so that it already has task-related discriminative information before being fused with the features of its neighbors; and the injectivity of the node update function is guaranteed.

10. The method for recognizing oracle bone script structure using a glyph-image isomorphic network according to claim 1, characterized in that, S3 specifically includes: Based on the oracle bone script character category labels output by the graph isomorphic network model, oracle bone script characters of the same category and structure are found in the graph structure dictionary. The number of shortlisted points is checked by using the coordinates of the corresponding nodes in the graph structure and the homography matrix, and the correctness of the graph structure recognition is determined.

Images