Graph matching based multi-component object matching method and system

Abstract

This invention relates to the field of computer vision and pattern recognition technology, and discloses a multi-component object matching method and system based on graph matching. The method involves: identifying discrete components and component type labels in an image of an object to be detected; forming a hierarchical set structure corresponding to a preset template hierarchy; establishing undirected fully connected edges between adjacent levels and filtering invalid edges to construct a dynamic graph; searching for candidate subgraphs in the dynamic graph that match the topological structure of the preset template; calculating the affine transformation matrix between each candidate subgraph and the preset template; calculating the matching degree between the remaining candidate subgraphs and the preset template, and taking matching results with matching degrees exceeding a preset threshold as valid matches; and outputting the object position coordinates and geometric transformation parameters corresponding to each valid match. This invention, by integrating deep learning detection and graph matching technologies, and combining hierarchical dynamic graph construction and backtracking affine double-level verification, achieves a significant improvement in detection rate and processing efficiency, while effectively reducing the false matching rate.

Description

Technical Field

[0001] This invention relates to the field of computer vision and pattern recognition technology, and in particular to a multi-component object matching method and system based on graph matching. Background Technology

[0002] In fields such as industrial quality inspection, automated assembly, and electronic component testing, many target objects are composed of multiple discrete components. For example, an accessory packaging box consists of a power cord, adapter, ferrule, and mounting bracket; a gearbox consists of multiple gears, shafts, and flanges. Inspecting such objects presents the following technical challenges:

[0003] 1. Occlusion and blurring lead to missed detections: Deep learning-based detection methods struggle to fully identify objects composed of multiple discrete components (such as groups of industrial parts) when the target is partially occluded or the image is blurred. This is because these methods rely on global feature extraction, and occlusion disrupts feature integrity, resulting in a low detection rate.

[0004] 2. Lack of component relationships: Traditional detection algorithms can only identify isolated components and cannot establish spatial topological relationships between components, making it difficult to reconstruct the complete object structure. This is particularly prominent in dense or dynamic scenes due to the lack of geometric constraint mechanisms.

[0005] 3. High false match rate: In multi-object dense environments, matching methods that do not consider the geometric constraints between components are prone to invalid matches. This is due to reasons including ineffective compensation for scale / rotation differences and the sensitivity of the matching algorithm to noise.

[0006] Limitations of existing methods: While existing graph matching schemes can model spatial relationships, they neglect the scale / rotation inconsistency between deep learning recognition results and templates, leading to insufficient matching robustness. The key to overcoming these difficulties lies in integrating efficient filtering mechanisms and verification strategies, but traditional methods struggle to balance computational efficiency and accuracy.

[0007] Therefore, there is an urgent need for an efficient method that combines deep learning detection and graph matching, which can maintain a high detection rate in complex scenarios, reduce the false matching rate, and have high computational efficiency. Summary of the Invention

[0008] The purpose of this invention is to provide a multi-component object matching method and system based on graph matching, so as to solve the problems of low detection rate, lack of component geometric constraints and high mismatch rate in occluded / blurred scenes in the prior art.

[0009] To achieve this objective, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides a multi-component object matching method based on graph matching, which includes the following steps:

[0011] S1. Obtain an image of the object to be detected that contains multiple discrete components;

[0012] S2. Identify each discrete component and its corresponding component type label in the image of the object to be detected;

[0013] S3. Organize the detected discrete components into layers according to different component type labels to form a layered set structure corresponding to the preset template hierarchy.

[0014] S4. Establish undirected fully connected edges between adjacent levels and filter invalid edges to construct a dynamic graph, wherein each node in the dynamic graph corresponds to a detected discrete component.

[0015] S5. Search the dynamic graph for candidate subgraphs that match the preset template topology.

[0016] S6. Calculate the affine transformation matrix between each candidate subgraph and the preset template, and remove candidate subgraphs that fail the geometric transformation consistency test.

[0017] S7. Calculate the matching degree between the remaining candidate sub-images and the preset template, and take the matching results with matching degree exceeding the preset threshold as valid matches, and output the object position coordinates and geometric transformation parameters corresponding to each valid match.

[0018] Preferably, invalid edges are filtered using a side length ratio constraint. Specifically, filtering invalid edges using a side length ratio constraint includes the following steps:

[0019] Calculate the actual side length between each pair of nodes in adjacent levels;

[0020] The actual side length is compared with the standard side length at the corresponding position in the preset template to calculate the side length ratio;

[0021] The dynamic graph is constructed by filtering out connected edges whose side length ratio is greater than the side scaling scale threshold and constructing connected edges whose side length ratio is less than or equal to the side scaling scale threshold.

[0022] Specifically, the edge scaling threshold satisfies the following formula:

[0023] ,

[0024] Where δ is the edge scaling threshold, L' is the actual edge length, and L is the standard edge length in the preset template corresponding to the actual edge length.

[0025] Specifically, the edge scaling threshold δ is between 0.15 and 0.25.

[0026] Preferably, step S3 specifically includes:

[0027] S31. Categorize the recognition results according to the component type labels;

[0028] S32. Group components with the same component type label into the same level to form a multi-level component collection;

[0029] S33. Establish the adjacency relationship between each level to form the hierarchical set structure, which corresponds to the hierarchical structure of the preset template.

[0030] The recognition result in step S31 is specifically the result obtained after executing step S2.

[0031] Preferably, a depth-first backtracking algorithm is used to search for candidate subgraphs in the dynamic graph that match the preset template topology.

[0032] Specifically, the step of using a depth-first backtracking algorithm to search for candidate subgraphs in the dynamic graph that match the preset template topology includes:

[0033] The depth-first backtracking algorithm is used to start the search from any node in the first level of the dynamic graph, and the search path is formed by selecting nodes layer by layer in hierarchical order;

[0034] When a path that does not meet the topological constraints is encountered, the search backtracks to the previous decision point and selects other nodes to continue the search until a complete path containing nodes at all levels and whose total number of nodes is equal to the preset template is found as a candidate subgraph.

[0035] Preferably, the affine transformation matrix includes translation parameters, rotation parameters, and scaling parameters, and the affine transformation matrix is ​​used to verify the geometric transformation consistency between the candidate subgraph and the preset template.

[0036] Specifically, verifying the geometric transformation consistency between the candidate subgraph and the preset template includes:

[0037] Calculate the geometric error after affine transformation;

[0038] Determine whether the geometric error is less than a preset geometric threshold, and discard candidate sub-graphs whose geometric errors are greater than or equal to the preset geometric threshold.

[0039] Preferably, calculating the matching degree between the remaining candidate subgraphs and the preset template specifically includes:

[0040] Calculate the ratio error between the actual side length and the standard side length of each connecting edge in the remaining candidate subgraphs;

[0041] The side length matching accuracy of the candidate subgraph is calculated based on the proportional error. The final matching degree is obtained by weighted summation of the side length matching accuracies of the candidate subgraphs. The formula for calculating the side length matching accuracy is as follows:

[0042] ,

[0043] Where T is the side length matching accuracy, L' is the actual side length, and L is the standard side length in the preset template that corresponds to the actual side length.

[0044] Preferably, the output of the object position coordinates and geometric transformation parameters corresponding to each valid match specifically includes:

[0045] Sort the valid matches in descending order based on their matching degree;

[0046] Output the position coordinates of the object in the image for each valid match, the geometric transformation parameters relative to the preset template, and the matching degree.

[0047] Preferably, a deep learning detection model is used to identify the individual discrete components and their corresponding component type labels.

[0048] Specifically, the deep learning detection model is obtained through the following steps:

[0049] Historical image data containing different types of components and their corresponding component type labels are used as training data samples to train an artificial intelligence model, thus obtaining the deep learning detection model.

[0050] Preferably, the preset template includes a pre-constructed distance relationship matrix between template components, which records the standard distance, relative positional relationship, and geometric constraint information of the connecting edges between each component node in the template.

[0051] Secondly, the present invention provides a multi-component object matching system based on graph matching, comprising:

[0052] The image acquisition unit is used to acquire an image of the object to be detected, which contains multiple discrete components;

[0053] The component detection unit is used to identify each discrete component and its corresponding component type label.

[0054] The hierarchical building unit is used to organize the detected discrete components hierarchically according to different component type labels, so as to form a hierarchical set structure corresponding to the preset template hierarchy.

[0055] The graph structure generation unit is used to establish undirected fully connected edges between adjacent levels and filter invalid edges to construct a dynamic graph, wherein each node in the dynamic graph corresponds to a detected discrete component.

[0056] A subgraph search unit is used to search for candidate subgraphs in the dynamic graph that match the preset template topology.

[0057] The geometric verification unit is used to calculate the affine transformation matrix between each candidate subgraph and the preset template, and to remove candidate subgraphs that have abnormal geometric transformation consistency.

[0058] The matching evaluation unit is used to calculate the matching degree between the remaining candidate sub-images and the preset template, and to take the matching results with matching degree exceeding the preset threshold as valid matches, and output the object position coordinates and geometric transformation parameters corresponding to each valid match.

[0059] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the graph-matching-based multi-component object matching method as described above.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] On the one hand, by organizing discrete components hierarchically and filtering invalid edges based on edge length ratio constraints, a geometric constraint mechanism between components is established, which effectively solves the problem that traditional methods cannot establish spatial topological relationships, avoids blind searching, and reduces computational complexity. On the other hand, by searching candidate subgraphs through a depth-first backtracking algorithm and performing affine transformation verification, the topological integrity and geometric consistency of the matching results are ensured, and effective compensation for scale, rotation, and translation changes is achieved. Furthermore, an interpretability scoring mechanism is provided through matching degree calculation and threshold screening, which facilitates result verification and quality control.

[0062] The present invention has other features and advantages, which will be apparent from or will be set forth in detail in the accompanying drawings and the following detailed description, which together serve to explain the particular principles of the invention. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of a multi-component object matching method based on graph matching provided in an embodiment of the present invention.

[0065] Figure 2This is a schematic diagram of affine transformation according to an embodiment of the present invention.

[0066] Figure 3 This is a schematic diagram of the template diagram and dynamic diagram construction in an embodiment of the present invention.

[0067] Figure 4 This is a schematic diagram of the backtracking search path in an embodiment of the present invention.

[0068] Figure 5 This is the identification path diagram of an embodiment of the present invention.

[0069] Figure 6 This is an example diagram of the matching results in an embodiment of the present invention.

[0070] Figure 7 This is a structural block diagram of a multi-component object matching system based on graph matching provided in an embodiment of the present invention. Detailed Implementation

[0071] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0072] Example 1

[0073] Please see Figures 1-6 The multi-component object matching method based on graph matching provided in this embodiment is applicable to scenarios such as industrial quality inspection, automated assembly, and electronic component inspection, which require the identification of target objects composed of multiple discrete components. This method integrates deep learning detection and graph matching techniques to establish geometric topological constraints between components, achieving accurate identification and localization of multi-component objects in complex scenes.

[0074] The core technical idea of ​​this invention is as follows: First, a deep learning model is used to detect discrete components in an image and identify their types. Then, a hierarchical graph structure is constructed based on the component types. Within the graph, margin filtering and backtracking search are used to find candidate subgraphs that match a preset template. Finally, affine transformation verification and matching degree calculation are used to determine the final valid matching result. Figure 1 A flowchart of the multi-component object matching method based on graph matching in this embodiment is shown.

[0075] To better understand the core technical concept of this embodiment, this embodiment will be described in detail using industrial parts inspection as a specific application scenario. Specifically, in this application scenario, the object to be inspected is an accessory packaging box, which is said to consist of four components: a power cord ×1, an adapter ×1, a card sleeve ×1, and a fixing bracket ×1. Figure 6The illustration shows characters on the exterior of one component of an accessory packaging box being partially obscured by at least one other component. The edge scaling threshold δ for this application scenario is set to 0.2, and the matching score threshold is set to 0.85.

[0076] The graph-matching-based multi-component object matching method in this embodiment includes the following steps:

[0077] S1. Obtain an image of the object to be detected, which contains multiple discrete components.

[0078] Specifically, in this embodiment, production line images containing accessory packaging boxes can be acquired using an industrial camera. This step provides raw data input for the entire matching process, ensuring that subsequent detection and matching algorithms have a sufficient image information foundation.

[0079] S2. Identify each discrete component and its corresponding component type label in the image of the object to be detected.

[0080] Specifically, this embodiment uses a deep learning detection model to identify each discrete component in the image and its corresponding component type label.

[0081] In this embodiment, a YOLOv5 network is used as a deep learning detection model to identify discrete components in the image and assign a corresponding type label to each component, including power cord label 0, adapter label 1, card sleeve label 2, and mounting bracket label 3. This step transforms the complex industrial scene into structured component data. This YOLOv5 network is an adaptively trained network model.

[0082] Of course, in other implementations, visual recognition, manual recognition, or other methods can be used to identify each discrete component and its corresponding component type label in the image of the object to be detected, and this is not limited here.

[0083] S3. Organize the detected discrete components into layers according to different component type labels.

[0084] Specifically, in the example of accessory packaging box inspection, the first layer collects all power cord labels 0, the second layer collects all adapter labels 1, the third layer collects all card sleeve labels 2, and the fourth layer collects all fixing bracket labels 3, forming a hierarchical collection structure. The hierarchical organization method can significantly reduce the computational complexity.

[0085] S4. Establish undirected fully connected edges between adjacent levels and filter out invalid edges to construct a dynamic graph. Figure 3 The diagram illustrates the construction of template graphs and dynamic graphs. This step, through margin filtering, can significantly reduce invalid calculations.

[0086] S5. Search the dynamic graph for candidate subgraphs that match the preset template topology. Figure 4 An example of a backtracking search path is shown, a step that ensures the completeness and accuracy of the search results.

[0087] S6. Calculate the affine transformation matrix between each candidate subgraph and the preset template, and remove candidate subgraphs that fail the geometric transformation consistency test. Figure 2 A schematic diagram of affine transformation verification is shown. This verification mechanism effectively eliminates distorted matches and significantly improves matching accuracy.

[0088] S7. Calculate the matching degree between the remaining candidate sub-images and the preset template, and take the matching results with a matching degree exceeding a preset threshold as valid matches. Output the object position coordinates and geometric transformation parameters corresponding to each valid match. Figure 6 An example image of the matching results is shown.

[0089] Preferably, in step S2, the deep learning detection model is obtained through the following steps: using historical image data containing different types of components and their corresponding component type labels as training data samples, an artificial intelligence model is trained to obtain the deep learning detection model. The deep learning detection model obtained after training the YOLOv5 network through the model can ensure the accuracy of recognition.

[0090] Preferably, in step S3, the layered organization specifically includes:

[0091] The recognition results are categorized according to component type labels;

[0092] Group components with the same component type label into the same level to form a multi-level component collection;

[0093] Establish adjacency relationships between each level to form the hierarchical set structure, which corresponds to the hierarchical structure of the preset template.

[0094] Preferably, in step S4, invalid edges are filtered using a side length ratio constraint. Specifically, filtering invalid edges using a side length ratio constraint includes:

[0095] Calculate the actual side length between each pair of nodes in adjacent levels;

[0096] The actual side length is compared with the standard side length at the corresponding position in the preset template to calculate the side length ratio;

[0097] Filter out connected edges whose side length ratio is greater than the side scaling scale threshold, and construct the dynamic graph using connected edges whose side length ratio is less than or equal to the side scaling scale threshold.

[0098] The edge scaling threshold satisfies the following formula:

[0099] ,

[0100] Where δ is the edge scaling threshold, L' is the actual edge length, and L is the standard edge length corresponding to the actual edge length in the preset template. The edge scaling threshold δ is between 0.15 and 0.25.

[0101] Preferably, in step S5, a depth-first backtracking algorithm is used to search for candidate subgraphs in the dynamic graph that match the preset template topology. Specifically, the search for candidate subgraphs in the dynamic graph using the depth-first backtracking algorithm includes:

[0102] The depth-first backtracking algorithm is used to start the search from any node in the first level of the dynamic graph, and the search path is formed by selecting nodes layer by layer in hierarchical order;

[0103] When a path that does not meet the topological constraints is encountered, the search backtracks to the previous decision point and selects other nodes to continue the search until a complete path containing nodes at all levels and whose total number of nodes is equal to the preset template is found as a candidate subgraph.

[0104] Of course, in other implementations, algorithms such as breadth-first backtracking, iterative deepening backtracking, and best-first backtracking can be used to search for candidate subgraphs that match the preset template topology in the dynamic graph.

[0105] Preferably, in step S6, the affine transformation matrix includes translation parameters, rotation parameters, and scaling parameters, and the affine transformation matrix is ​​used to verify the geometric transformation consistency between the candidate subgraph and the preset template.

[0106] The verification of the geometric transformation consistency between the candidate subgraph and the preset template specifically includes:

[0107] Calculate the geometric error after affine transformation;

[0108] Determine whether the geometric error is less than a preset geometric threshold, and discard candidate sub-graphs whose geometric errors are greater than or equal to the preset geometric threshold.

[0109] Preferably, in step S7, calculating the matching degree between the remaining candidate subgraphs and the preset template specifically includes:

[0110] Calculate the ratio error between the actual side length and the standard side length of each connecting edge in the remaining candidate subgraphs;

[0111] The side length matching accuracy of the candidate subgraph is calculated based on the proportional error. The final matching degree is obtained by weighted summation of the side length matching accuracies of the candidate subgraphs. The formula for calculating the side length matching accuracy is as follows:

[0112] ,

[0113] Where T is the side length matching precision, L' is the actual side length, and L is the standard side length in the preset template that corresponds to the actual side length.

[0114] The output of the object position coordinates and geometric transformation parameters corresponding to each valid match specifically includes: sorting the valid matches in descending order according to the matching degree;

[0115] Output the position coordinates of the object in the image for each valid match, the geometric transformation parameters relative to the preset template, and the matching degree.

[0116] Preferably, the preset template includes a pre-constructed distance relationship matrix between template components, which records the standard distance, relative positional relationship, and geometric constraint information of the connecting edges between each component node in the template.

[0117] Figure 5 This embodiment illustrates the identification path for accessory packaging boxes. Figure 6 The corresponding matching results are shown.

[0118] Practical verification using the method described in this embodiment on 200 test images showed a detection rate of 91.4%, a coordinate error of less than 3 pixels, and an average processing time of 95ms per target. Compared to traditional methods, this invention improves the occluded scene recognition rate from 63% to 93%, an increase of 45.9%, reduces the matching processing time from 285ms to 89ms, an overall decrease of 68.8%, and reduces the false matching rate from 34% to 6.8%, a reduction of 80%.

[0119] Example 2

[0120] Please see Figure 7 The multi-component object matching system based on graph matching in this embodiment includes:

[0121] Image acquisition unit 10 is used to acquire an image of a target object containing multiple discrete components;

[0122] The component detection unit 20 is used to identify each discrete component and its corresponding component type label.

[0123] The hierarchical building unit 30 is used to organize the detected discrete components hierarchically according to different component type labels, so as to form a hierarchical set structure corresponding to the preset template hierarchical relationship;

[0124] Graph structure generation unit 40 is used to establish undirected fully connected edges between adjacent levels and filter invalid edges to construct a dynamic graph, wherein each node in the dynamic graph corresponds to a detected discrete component.

[0125] The subgraph search unit 50 is used to search for candidate subgraphs in the dynamic graph that match the preset template topology.

[0126] The geometric verification unit 60 is used to calculate the affine transformation matrix between each candidate subgraph and the preset template, and to remove candidate subgraphs that have abnormal geometric transformation consistency.

[0127] The matching evaluation unit 70 is used to calculate the matching degree between the remaining candidate sub-images and the preset template, and to take the matching results with matching degrees exceeding the preset threshold as valid matches, and output the object position coordinates and geometric transformation parameters corresponding to each valid match.

[0128] Example 3

[0129] The computer-readable storage medium of this embodiment stores a computer program thereon, which, when executed by a processor, implements the above-described multi-component object matching method based on graph matching.

[0130] Combination Figures 1-7 The present invention has the following beneficial effects:

[0131] This invention establishes a geometric constraint mechanism between components by hierarchically organizing discrete components and filtering invalid edges based on edge length ratio constraints. This effectively solves the problem that traditional methods cannot establish spatial topological relationships, avoids blind searching, and reduces computational complexity. Furthermore, it ensures the topological integrity and geometric consistency of the matching results by searching candidate subgraphs using a depth-first backtracking algorithm and performing affine transformation verification, achieving effective compensation for scale, rotation, and translation changes. Additionally, it provides an interpretability scoring mechanism through matching degree calculation and threshold screening, facilitating result verification and quality control. Practical verification shows that the overall solution of this invention improves the detection rate from 63% to 93% compared to existing technologies, reduces processing time from 285ms to 89ms, and lowers the false matching rate from 34% to 6.8%.

[0132] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-component object matching method based on graph matching, characterized in that, Includes the following steps: Acquire an image of the object to be detected, which contains multiple discrete components; Identify each discrete component and its corresponding component type label in the image of the object to be detected; The identification results are categorized according to component type labels. Components with the same component type label are grouped into the same level to form a multi-level component set. Adjacency relationships between each level are established to form a hierarchical set structure, which corresponds to the hierarchical structure of the preset template. Undirected fully connected edges are established between adjacent levels, and invalid edges are filtered to construct a dynamic graph, wherein each node in the dynamic graph corresponds to a detected discrete component. Search the dynamic graph for candidate subgraphs that match the preset template topology. Calculate the affine transformation matrix between each candidate subgraph and the preset template, and remove candidate subgraphs that fail the geometric transformation consistency test. Calculate the matching degree between the remaining candidate subgraphs and the preset template, and take the matching results with matching degree exceeding the preset threshold as valid matches, and output the object position coordinates and geometric transformation parameters corresponding to each valid match.

2. The multi-component object matching method based on graph matching as described in claim 1, characterized in that, Filtering invalid edges using side length ratio constraints specifically includes the following steps: Calculate the actual side length between each pair of nodes in adjacent levels; The actual side length is compared with the standard side length at the corresponding position in the preset template to calculate the side length ratio; Filter out connected edges whose side length ratio is greater than the side scaling scale threshold, and construct the dynamic graph using connected edges whose side length ratio is less than or equal to the side scaling scale threshold.

3. The multi-component object matching method based on graph matching as described in claim 2, characterized in that, The edge scaling threshold satisfies the following formula: ， Where δ is the edge scaling threshold, L' is the actual edge length, and L is the standard edge length in the preset template corresponding to the actual edge length.

4. The multi-component object matching method based on graph matching as described in claim 2, characterized in that, The edge scaling threshold δ is between 0.15 and 0.

25.

5. The multi-component object matching method based on graph matching as described in claim 1, characterized in that, A depth-first backtracking algorithm is used to search for candidate subgraphs in the dynamic graph that match the preset template topology.

6. The multi-component object matching method based on graph matching as described in claim 5, characterized in that, The process of using a depth-first backtracking algorithm to search for candidate subgraphs in the dynamic graph that match the preset template topology specifically includes: The depth-first backtracking algorithm is used to start the search from any node in the first level of the dynamic graph, and the search path is formed by selecting nodes layer by layer in hierarchical order; When a path that does not meet the topological constraints is encountered, the search backtracks to the previous decision point and selects other nodes to continue the search until a complete path containing nodes at all levels and whose total number of nodes is equal to the preset template is found as a candidate subgraph.

7. The multi-component object matching method based on graph matching as described in claim 1, characterized in that, The affine transformation matrix includes translation parameters, rotation parameters, and scaling parameters. The affine transformation matrix is ​​used to verify the consistency of geometric transformations between candidate subgraphs and preset templates.

8. The multi-component object matching method based on graph matching as described in claim 7, characterized in that, The verification of the geometric transformation consistency between the candidate subgraph and the preset template specifically includes: Calculate the geometric error after affine transformation; Determine whether the geometric error is less than a preset geometric threshold, and discard candidate sub-graphs whose geometric errors are greater than or equal to the preset geometric threshold.

9. The multi-component object matching method based on graph matching as described in claim 1, characterized in that, The calculation of the matching degree between the remaining candidate subgraphs and the preset template specifically includes: Calculate the ratio error between the actual side length and the standard side length of each connecting edge in the remaining candidate subgraphs; The side length matching accuracy of the candidate subgraph is calculated based on the proportional error. The final matching degree is obtained by weighted summation of the side length matching accuracies of the candidate subgraphs. The formula for calculating the side length matching accuracy is as follows: ， Where T is the side length matching accuracy, L' is the actual side length, and L is the standard side length in the preset template that corresponds to the actual side length.

10. The multi-component object matching method based on graph matching as described in claim 1, characterized in that, The output of the object position coordinates and geometric transformation parameters corresponding to each valid match specifically includes: Sort the valid matches in descending order based on their matching degree; Output the position coordinates of the object in the image for each valid match, the geometric transformation parameters relative to the preset template, and the matching degree.

11. The multi-component object matching method based on graph matching as described in claim 1, characterized in that, A deep learning detection model is used to identify the individual discrete components and their corresponding component type labels.

12. The multi-component object matching method based on graph matching as described in claim 11, characterized in that, The deep learning detection model is obtained through the following steps: Historical image data containing different types of components and their corresponding component type labels are used as training data samples to train an artificial intelligence model, thus obtaining the deep learning detection model.

13. The multi-component object matching method based on graph matching as described in claim 1, characterized in that, The preset template includes a pre-constructed distance relationship matrix between template components. The distance relationship matrix records the standard distance, relative position relationship, and geometric constraint information of the connecting edges between each component node in the template.

14. A multi-component object matching system based on graph matching, characterized in that, include: The image acquisition unit is used to acquire an image of the object to be detected, which contains multiple discrete components; A component detection unit is used to identify each discrete component and its corresponding component type label in the image of the object to be detected. The hierarchical construction unit is used to classify the recognition results according to the component type label, group components with the same component type label into the same level to form a multi-level component set, and establish the adjacency relationship between each level to form a hierarchical set structure, which corresponds to the hierarchical structure of the preset template. The graph structure generation unit is used to establish undirected fully connected edges between adjacent levels and filter invalid edges to construct a dynamic graph, wherein each node in the dynamic graph corresponds to a detected discrete component. A subgraph search unit is used to search for candidate subgraphs in the dynamic graph that match the preset template topology. The geometric verification unit is used to calculate the affine transformation matrix between each candidate subgraph and the preset template, and to remove candidate subgraphs that have abnormal geometric transformation consistency. The matching evaluation unit is used to calculate the matching degree between the remaining candidate sub-images and the preset template, and to take the matching results with matching degree exceeding the preset threshold as valid matches, and output the object position coordinates and geometric transformation parameters corresponding to each valid match.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the multi-component object matching method based on graph matching as described in any one of claims 1 to 13.

Images