Multi-label graph structure construction method based on conditional probability matrix and GCN

Abstract

The invention provides a multi-label graph structure construction method based on a conditional probability matrix and a GCN. The method comprises the following steps: classifying label data in a multi-label data set, constructing a conditional probability matrix according to a co-occurrence frequency corresponding to each label combination, and constructing a corresponding mean-standard deviation dynamic threshold, and performing binarization processing on the conditional probability matrix by using a mean-standard deviation dynamic threshold, generating a binarized adjacency matrix to obtain binary information corresponding to each label combination, inputting the binarized adjacency matrix into a preset GCN for convolution processing to obtain convolution information corresponding to each label combination, and outputting the convolution information corresponding to each label combination. According to the method, the label association features corresponding to each label combination are generated based on the binary information and the convolution information corresponding to each label combination, and the multi-label graph structure of the multi-label data set is constructed by using the label association features, so that the problem that an over-fitting phenomenon easily occurs in the traditional technology is effectively solved, and the performance and the quality of data detection are guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-label graph structures, and particularly to a method for constructing a multi-label graph structure based on a conditional probability matrix and a GCN. Background Art

[0002] Image classification, as an important branch in the field of computer vision, has been widely applied to application scenarios such as object recognition and defect detection. According to the number of labels in an image, there are two types of image classification: single-label image classification and multi-label image classification. With the development and popularization of deep learning technology, the performance of single-panel image classification methods has been sufficiently superior. Compared with single-label image classification, multi-label image classification meets more general requirements. Its implementation and realization are more difficult and complex, and thus more challenging in image processing tasks. Large-scale multi-label image classification requires determining the presence or absence of target objects in a large number of sample images. Due to the huge number of samples and different levels of people's awareness, this may lead to very low efficiency of multi-polyhedron image classification, especially for highly specialized and complex multi-label image sets. Therefore, ensuring the accuracy of image classification is particularly important.

[0003] However, existing multi-label image classification methods focus on the accuracy of label prediction while ignoring the structural information embedded in the hierarchical label space, lacking deep semantic associations and resulting in a decline in detection performance. Therefore, there is an urgent need for a method to solve the above problems.

[0004] Therefore, the present invention provides a method for constructing a multi-label graph structure based on a conditional probability matrix and a GCN. Summary of the Invention

[0005] The method for constructing a multi-label graph structure based on a conditional probability matrix and a GCN of the present invention effectively solves the problem of overfitting easily occurring in traditional technologies, and also ensures the performance and quality of data detection.

[0006] The present invention provides a method for constructing a multi-label graph structure based on a conditional probability matrix and a GCN, including:

[0007] Step 1: Classify the label data in the multi-label data set to obtain the co-occurrence frequency corresponding to each label combination, and construct the conditional probability matrix of the multi-label data set;

[0008] Step 2: Calculate the data mean and data standard deviation corresponding to each label data respectively, and construct the corresponding mean-standard deviation dynamic threshold;

[0009] Step 3: Use the mean-standard deviation dynamic threshold to perform binarization processing on the conditional probability matrix to generate a binarized adjacency matrix and obtain the binary information corresponding to each label combination;

[0010] Step 4: Input the binarized adjacency matrix into a preset GCN for convolutional processing to obtain the convolutional information corresponding to each label combination;

[0011] Step 5: Generate the label association features corresponding to each label combination based on the binary information and convolutional information corresponding to each label combination, and use the label association features to construct the multi-label graph structure of the multi-label dataset.

[0012] In an implementable manner,

[0013] The said Step 1 includes:

[0014] Step 11: Identify a number of label data included in the multi-label dataset, perform binary classification training on each label data respectively to obtain a number of labels included in each label data, and simultaneously identify the samples corresponding to the label data;

[0015] Step 12: Combine the labels according to the correspondence between the samples and labels included in the same label data to obtain a number of label combinations corresponding to each label data, and construct the combination conditions corresponding to each label combination according to the label features corresponding to each label;

[0016] Step 13: Use each combination condition to perform in-depth statistical search on the multi-label dataset respectively to obtain the co-occurrence frequency corresponding to each label combination, construct the corresponding label event according to the label combination, and construct the corresponding event probability according to the co-occurrence frequency;

[0017] Step 14: Construct the conditional probability matrix of the multi-label dataset according to the label event and event probability.

[0018] In an implementable manner,

[0019] The said Step 2 includes:

[0020] Step 21: Perform numerical sampling on each label data respectively to obtain a number of label values, divide the label data into a number of data segments based on the sampling positions corresponding to each label value, and screen the target data segments with a data segment length higher than the preset length for compensation sampling to obtain the corresponding label values;

[0021] Step 22: Calculate the data mean and data standard deviation corresponding to the label data according to the label values, use the data mean and data standard deviation to determine the dynamic threshold range corresponding to the label data, and establish the range ratio of the dynamic threshold range according to the quantity ratio of the first label values falling within the dynamic threshold range and the second label values not falling within the dynamic threshold range;

[0022] Step 23: Use the mean and standard deviation corresponding to the range for correction to obtain a number of corrected means and corrected standard deviations, screen out a number of target corrected means and target corrected standard deviations that fall within the corresponding dynamic threshold range, and generate a mean-standard deviation dynamic threshold corresponding to the label data.

[0023] In an implementable manner,

[0024] The said step 3 includes:

[0025] Step 31: Respectively use each mean-standard deviation dynamic threshold to perform binarization processing on the conditional probability matrix to obtain the assignment result of each mean-standard deviation dynamic threshold to the conditional probability matrix;

[0026] Step 32: Respectively calculate the assignment balance degree corresponding to each assignment result, and use the target assignment result with the highest assignment balance degree to perform binarization assignment on the conditional probability matrix to generate a binarized adjacency matrix;

[0027] Step 33: Respectively identify the matrix elements corresponding to each label combination in the binarized adjacency matrix, and determine the binary information corresponding to the label combination according to the assignment corresponding to the matrix elements.

[0028] In an implementable manner,

[0029] The said step 4 includes:

[0030] Step 41: Obtain a number of adjacent nodes included in the binarized adjacency matrix, input the binarized adjacency matrix into a preset GCN for convolution processing to obtain the node relationship between different adjacent nodes;

[0031] Step 42: Respectively match a corresponding label combination for each adjacent node, construct a label combination relationship according to the node relationship, and generate an external feature of each label combination;

[0032] Step 43: Capture the convolution feature corresponding to each adjacent node in the convolution processing result, and establish the convolution information of the label combination in combination with the corresponding external feature.

[0033] In an implementable manner,

[0034] The said step 5 includes:

[0035] Step 51: Establish a shadow distribution feature corresponding to the label combination according to the binary information, and establish a pixel distribution feature corresponding to the label combination according to the convolution information;

[0036] Step 52: Obtain the presentation features of each of the label combinations in the binarized neighborhood matrix respectively, and construct the label-related features corresponding to each label combination by combining the corresponding shadow distribution features and pixel distribution features;

[0037] Step 53: Draw the label images corresponding to the combined labels according to the label-related features, fuse the label images according to the presentation features to generate the multi-label graph structure of the multi-label dataset, and display it.

[0038] In an implementable manner,

[0039] It further includes:

[0040] Mark the label names corresponding to each label combination in the multi-label graph structure respectively, obtain the graph domain corresponding to each label combination, and display it.

[0041] In an implementable manner,

[0042] It further includes:

[0043] Identify the nodes and edges included in the multi-label graph structure;

[0044] Identify several first labels corresponding to each node and several second labels corresponding to each edge respectively;

[0045] Calculate the first label sparsity corresponding to each node and the second label sparsity corresponding to each edge, and determine the structural key area of the multi-label graph structure;

[0046] Highlight the key area in the multi-label graph structure.

[0047] In an implementable manner,

[0048] It further includes:

[0049] Search for the relevant information included in the multi-label graph structure according to the execution instructions issued by the user, and transmit it to the corresponding terminal for display.

[0050] In an implementable manner,

[0051] It further includes:

[0052] Identify the domain information included in the multi-label graph structure according to the graph structure domain uploaded by the user, and display it.

[0053] The achievable beneficial effects of the above technical solution are as follows: To achieve image classification and structure determination, first, the multi-label dataset is mechanically classified to determine the co-occurrence frequency of each label combination, thereby establishing a conditional probability matrix. Further, the mean and standard deviation of the label data are used to establish a mean-standard deviation dynamic threshold, thereby binarizing the conditional probability matrix to determine the binary information of each label combination. Then, a preset GCN is further used to perform convolution processing on the binarized adjacency matrix to obtain the convolution information of each label combination. Finally, the label association features of the label combination are constructed based on the binary information and convolution information of the label combination, and the multi-label graph structure is constructed using the label association feature chain. In this way, not only can multiple groups of label data be analyzed simultaneously, but also the noise in the labels can be significantly eliminated, improving the classification accuracy of the multi-label graph structure.

[0054] Other features and advantages of the present invention will be described in the following specification, and in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained by the structure specifically pointed out in the written specification and the drawings.

[0055] The technical solution of the present invention will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0057] Figure 1 is a schematic diagram of the working process of the multi-label graph structure construction method based on the conditional probability matrix and GCN in the embodiment of the present invention;

[0058] Figure 2 is a schematic diagram of the working process of step 2 of the multi-label graph structure construction method based on the conditional probability matrix and GCN in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] The following describes the preferred embodiments of the present invention with reference to the drawings. It should be understood that the preferred embodiments described herein are only for the purpose of illustrating and explaining the present invention and are not intended to limit the present invention.

[0060] Embodiment 1

[0061] This embodiment provides a multi-label graph structure construction method based on the conditional probability matrix and GCN, as Figure 1 shown, including:

[0062] Step 1: Classify the label data in the multi-label dataset to obtain the co-occurrence frequency corresponding to each label combination, and construct the conditional probability matrix of the multi-label dataset;

[0063] Step 2: Calculate the data mean and data standard deviation corresponding to each label data respectively, and construct the corresponding mean-standard deviation dynamic threshold;

[0064] Step 3: Use the mean-standard deviation dynamic threshold to perform binarization processing on the conditional probability matrix, generate a binarized adjacency matrix, and obtain the binary information corresponding to each label combination;

[0065] Step 4: Input the binarized adjacency matrix into a preset GCN for convolution processing to obtain the convolution information corresponding to each label combination;

[0066] Step 5: Generate the label association features corresponding to each label combination based on the binary information and convolution information corresponding to each label combination, and use the label association features to construct the multi-label graph structure of the multi-label dataset.

[0067] In this example, the label combination represents the label group corresponding to the data containing two labels simultaneously;

[0068] In this example, the mean-standard deviation dynamic threshold represents the result of determining the dynamic threshold of label data according to the mean and standard deviation of the data;

[0069] In this example, the binarization processing represents the result of adjusting the elements in the conditional probability matrix to only contain two values of 0 and 1;

[0070] In this example, the binary information represents the value related to a label combination in the binarized adjacency matrix;

[0071] In this example, the preset GCN represents a model used for deep learning;

[0072] In this example, the convolution information represents the result presented after convolving the label combination;

[0073] In this example, the label association feature represents the feature formed due to the label combination in a label combination.

[0074] Working principle and beneficial effects of the above technical solution: To achieve image classification and structure determination, first, the multi-label dataset is mechanically classified to determine the co-occurrence frequency of each label combination, thereby establishing a conditional probability matrix. Further, the mean and standard deviation of the label data are used to establish a mean-standard deviation dynamic threshold to binarize the conditional probability matrix, determining the binary information of each label combination. Then, a preset GCN is used to perform convolution processing on the binarized adjacency matrix to obtain the convolution information of each label combination. Finally, based on the binary information and convolution information of the label combination, the label association features of the label combination are constructed, and a multi-label graph structure is constructed using the label association feature chain. In this way, not only can multiple sets of label data be analyzed simultaneously, but also the noise in the labels can be significantly eliminated, improving the classification accuracy of the multi-label graph structure.

[0075] Example 2

[0076] Based on the method for constructing a multi-label graph structure using a conditional probability matrix and GCN in Example 1, Step 1 includes:

[0077] Step 11: Identify a number of label data included in the multi-label dataset, perform binary classification training on each of the label data to obtain a number of labels included in each of the label data, and simultaneously identify the samples corresponding to the label data;

[0078] Step 12: Combine the labels according to the correspondence between the samples and the labels included in the same label data to obtain a number of label combinations corresponding to each label data, and construct a combined condition corresponding to each label combination according to the label features corresponding to each label;

[0079] Step 13: Use each of the combined conditions to perform in-depth statistical search on the multi-label dataset to obtain the co-occurrence frequency corresponding to each label combination, construct a corresponding label event according to the label combination, and construct a corresponding event probability according to the co-occurrence frequency;

[0080] Step 14: Construct a conditional probability matrix of the multi-label dataset according to the label event and the event probability.

[0081] In this example, binary classification training represents the training process of identifying two labels in the label data;

[0082] In this example, a sample represents the basic data of the label data, which consists of the features input by the user and the input labels;

[0083] In this example, the correspondence represents the relationship between the sample and the label;

[0084] In this example, the combined condition represents the condition for generating a label combination.

[0085] The working principle and beneficial effects of the above technical solution: By performing binary classification training on the data labels in the multi-label dataset to determine the label and sample corresponding to each label data, establishing the combined condition of the label combination according to the corresponding relationship between the label and the sample, thereby counting the co-occurrence frequency of each label combination, constructing the corresponding label event and determining the event probability corresponding to each label event, and finally constructing a conditional probability matrix. The conditional probability matrix can present the event probabilities corresponding to multiple label events at the same time, which is beneficial to subsequent analysis of the characteristics of the label data.

[0086] Embodiment 3

[0087] Based on Embodiment 1, the method for constructing a multi-label graph structure based on a conditional probability matrix and a GCN is as Figure 2 shown, and Step 2 includes:

[0088] Step 21: Numerically sample each of the label data to obtain a number of label values, divide the label data into a number of data segments based on the sampling positions corresponding to each label value, and screen the target data segments with a data segment length higher than a preset length for compensation sampling to obtain the corresponding label values;

[0089] Step 22: Calculate the data mean and data standard deviation corresponding to the label data according to the label values, determine the dynamic threshold range corresponding to the label data using the data mean and data standard deviation, and establish the range ratio of the dynamic threshold range according to the quantity ratio of the first label values falling within the dynamic threshold range and the second label values not falling within the dynamic threshold range;

[0090] Step 23: Use the range ratio to correct the corresponding mean and standard deviation to obtain a number of corrected means and corrected standard deviations, screen a number of target corrected means and target corrected standard deviations falling within the corresponding dynamic threshold range, and generate the mean-standard deviation dynamic threshold corresponding to the label data.

[0091] In this example, the label value represents the value obtained by sampling from the label data;

[0092] In this example, the preset length is 5% of the label data;

[0093] In this example, the dynamic threshold range represents the range where the dynamic threshold of the label data will appear;

[0094] In this example, the quantity ratio represents the quantity ratio between the first label value and the second label value;

[0095] In this example, the range ratio represents the range value used to adjust the dynamic threshold range within the range ratio.

[0096] The working principle and beneficial effects of the above technical solution: By sampling the tag data to obtain the corresponding tag values, determining the dynamic threshold range of the tag data according to the data mean and data standard deviation of the tag data, and correcting the data mean and data standard deviation according to the proportion of the tag data falling inside and outside the dynamic threshold range, the mean-standard deviation dynamic threshold of the tag data is determined, laying a foundation for subsequent binarization processing.

[0097] Example 4

[0098] Based on Example 1, the method for constructing a multi-label graph structure based on a conditional probability matrix and GCN, step 3 includes:

[0099] Step 31: Respectively perform binarization processing on the conditional probability matrix using each of the mean-standard deviation dynamic thresholds to obtain the assignment result of each mean-standard deviation dynamic threshold to the conditional probability matrix;

[0100] Step 32: Calculate the assignment balance degree corresponding to each assignment result respectively, and use the target assignment result with the highest assignment balance degree to perform binarization assignment on the conditional probability matrix to generate a binarized adjacency matrix;

[0101] Step 33: Identify the matrix elements corresponding to each label combination in the binarized adjacency matrix respectively, and determine the binary information corresponding to the label combination according to the assignment corresponding to the matrix element.

[0102] In this example, the assignment result represents the result of inputting 0 and 1 values into the conditional probability matrix;

[0103] In this example, the assignment balance degree represents the balance ratio of 0 and 1 values in each assignment result;

[0104] In this example, the binarization assignment represents the process of assigning 0 and 1 values to the conditional probability matrix.

[0105] The working principle and beneficial effects of the above technical solution: Use the mean-standard deviation dynamic threshold to perform binarization processing on the conditional probability matrix, convert the conditional probability matrix into a simple and efficient binarized adjacency matrix, and then identify the matrix elements corresponding to each label combination in the binarized adjacency matrix to construct the binary information of the label combination. In this way, the information of the label combination can be simplified, the binary information of the label combination is determined, and the quality and efficiency of identifying the label combination information are improved.

[0106] Example 5

[0107] Based on Embodiment 1, in the method for constructing a multi-label graph structure based on a conditional probability matrix and a GCN, Step 4 includes:

[0108] Step 41: Obtain a number of adjacent nodes included in the binary adjacency matrix, input the binary adjacency matrix into a preset GCN for convolution processing, and obtain the node relationships between different adjacent nodes;

[0109] Step 42: Match a corresponding label combination to each adjacent node, construct a label combination relationship according to the node relationship, and generate an external feature for each label combination;

[0110] Step 43: Capture the convolution feature corresponding to each adjacent node in the convolution processing result, and establish the convolution information of the label combination in combination with the corresponding external feature.

[0111] In this example, an adjacent node refers to a node composed of elements with an adjacent relationship in the binary adjacency matrix;

[0112] In this example, the node relationship refers to the logical association relationship between different adjacent nodes;

[0113] In this example, the external feature refers to the feature presented by the logical relationship between different label combinations;

[0114] In this example, the convolution information refers to the information presented after convolving the information of the label combination.

[0115] The working principle and beneficial effects of the above technical solution: By inputting the binary adjacency matrix into the GCN for convolution training to determine the node relationship between adjacent nodes, then setting corresponding external features for the corresponding label combinations, and at the same time capturing its convolution information in the convolution processing result, the noise in the original information can be eliminated by convolution, and the features of the label combination at different levels can be captured, ensuring the integrity of the convolution information.

[0116] Embodiment 6

[0117] Based on Embodiment 1, in the method for constructing a multi-label graph structure based on a conditional probability matrix and a GCN, Step 5 includes:

[0118] Step 51: Establish a shadow distribution feature corresponding to the label combination according to the binary information, and establish a pixel distribution feature corresponding to the label combination according to the convolution information;

[0119] Step 52: Obtain the presentation features of each label combination in the binary neighborhood matrix respectively, and construct the label-related features corresponding to each label combination in combination with the corresponding shadow distribution feature and pixel distribution feature;

[0120] Step 53: Draw a label image corresponding to the combined label according to the label-related features, and fuse the label images according to the presentation features to generate a multi-label graph structure of the multi-label dataset and display it.

[0121] In this example, the presentation feature represents the performance of a label combination in the binary neighborhood matrix.

[0122] The working principle and beneficial effects of the above technical solution: Using binary information to establish the shadow distribution characteristics of label combinations and using convolutional information to establish the pixel distribution characteristics of label combinations, combining the presentation characteristics of label combinations in the binary neighborhood matrix to construct the label-related characteristics of each label combination, and then drawing label images according to the label-related characteristics, thereby constructing a multi-label graph structure. In this way, in-depth analysis of label combinations can be carried out to determine their characteristics in each dimension, and then a multi-label graph structure can be constructed based on their characteristics, improving the accuracy of image classification.

[0123] Example 7

[0124] Based on Example 1, the method for constructing a multi-label graph structure based on a conditional probability matrix and GCN further includes:

[0125] Mark the label names corresponding to each label combination in the multi-label graph structure respectively to obtain the graph domain corresponding to each label combination and display it.

[0126] The working principle and beneficial effects of the above technical solution: By marking the label names in the multi-label graph structure, it is convenient for users to distinguish the information corresponding to different labels.

[0127] Example 8

[0128] Based on Example 7, the method for constructing a multi-label graph structure based on a conditional probability matrix and GCN further includes:

[0129] Identify the nodes and edges included in the multi-label graph structure;

[0130] Identify several first labels corresponding to each node and several second labels corresponding to each edge respectively;

[0131] Calculate the sparsity of the first labels corresponding to each node and the sparsity of the second labels corresponding to each edge to determine the key structural areas of the multi-label graph structure;

[0132] Highlight the key areas in the multi-label graph structure.

[0133] In this example, a node represents the smallest unit that makes up a multi-label graph structure, and an edge represents the structural edge of the multi-label graph structure;

[0134] In this example, label sparsity represents the tightness of the labels corresponding to a node or an edge.

[0135] The working principle and beneficial effects of the above technical solution: By analyzing the label sparsity of the nodes and the label sparsity of the edges in the multi-label graph structure, key areas with high label sparsity are determined. To facilitate users to quickly locate the key areas, the key areas are prominently displayed in the multi-label graph structure, improving the user experience.

[0136] Embodiment 9

[0137] Based on Embodiment 1, the method for constructing a multi-label graph structure based on a conditional probability matrix and GCN is characterized in that it further includes:

[0138] Search for relevant information included in the multi-label graph structure according to the execution instruction issued by the user, and transmit it to the corresponding terminal for display.

[0139] The working principle and beneficial effects of the above technical solution: When the user issues an instruction, the instruction is promptly responded to and relevant information is filtered and transmitted to the user's terminal for display.

[0140] Embodiment 10

[0141] Based on Embodiment 1, the method for constructing a multi-label graph structure based on a conditional probability matrix and GCN further includes:

[0142] Identify the domain information included in the multi-label graph structure according to the graph structure domain uploaded by the user and display it.

[0143] The working principle and beneficial effects of the above technical solution: Identify and display the relevant information included in the multi-label graph structure according to the graph structure domain uploaded by the user in advance, improving the accuracy of automatic identification of the multi-label structure.

[0144] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A method for constructing a multi-label graph structure based on a conditional probability matrix and GCN, characterized in that: include: Step 1: Classify the label data in the multi-label data set, obtain the co-occurrence frequency corresponding to each label combination, and construct the conditional probability matrix of the multi-label data set; Step 2: Calculate the data mean and data standard deviation corresponding to each of the label data, and construct the corresponding mean-standard deviation dynamic threshold; Step 3: Binarize the conditional probability matrix using the mean-standard deviation dynamic threshold to generate a binary adjacency matrix to obtain binary information corresponding to each label combination; Step 4: Input the binary adjacency matrix into the preset GCN for convolution processing to obtain the convolution information corresponding to each label combination; Step 5: Generate label association features corresponding to each label combination based on the binary information and convolution information corresponding to each label combination, and use the label association features to construct a multi-label graph structure of the multi-label data set.

2. The method for constructing a multi-label graph structure based on a conditional probability matrix and GCN as claimed in claim 1, characterized in that: The step 1 comprises: Step 11: Identify several label data contained in the multi-label data set, perform binary classification training on each of the label data, obtain several labels contained in each of the label data, and identify samples corresponding to the label data; Step 12: combining the labels according to the correspondence between the samples and labels contained in the same label data to obtain a plurality of label combinations corresponding to each label data, and constructing a combination condition corresponding to each label combination according to the label feature corresponding to each label; Step 13: Use each of the combination conditions to perform deep statistical search on the multi-label data set to obtain the co-occurrence frequency corresponding to each of the label combinations, construct corresponding label events according to the label combinations, and construct corresponding event probabilities according to the co-occurrence frequencies; Step 14: Construct a conditional probability matrix of the multi-label data set according to the label events and event probabilities.

3. The method for constructing a multi-label graph structure based on a conditional probability matrix and GCN as claimed in claim 1, characterized in that: The step 2 comprises: Step 21: numerically sampling each of the label data to obtain a plurality of label values, dividing the label data into a plurality of data segments based on the sampling position corresponding to each of the label values, and selecting a target data segment whose data segment length is longer than a preset length for compensation sampling to obtain a corresponding label value; Step 22: Calculate the data mean and data standard deviation corresponding to the label data according to the label value, determine the dynamic threshold range corresponding to the label data by using the data mean and data standard deviation, and establish the range ratio of the dynamic threshold range according to the ratio of the number of first label values ​​falling within the dynamic threshold range and the number of second label values ​​not falling within the dynamic threshold range; Step 23: Use the range comparison to correct the corresponding mean and standard deviation to obtain several corrected means and corrected standard deviations, screen several target corrected means and target corrected standard deviations that fall within the corresponding dynamic threshold range, and generate a mean-standard deviation dynamic threshold corresponding to the label data.

4. The method for constructing a multi-label graph structure based on a conditional probability matrix and GCN as claimed in claim 1, characterized in that: The step 3 comprises: Step 31: using each of the mean-standard deviation dynamic thresholds to perform binarization processing on the conditional probability matrix, to obtain the assignment result of each of the mean-standard deviation dynamic thresholds to the conditional probability matrix; Step 32: Calculate the assignment balance corresponding to each of the assignment results respectively, and use the target assignment result with the highest assignment balance to perform binarization assignment on the conditional probability matrix to generate a binarized adjacency matrix; Step 33: Identify the matrix elements corresponding to each of the label combinations in the binary adjacency matrix, and determine the binary information corresponding to the label combination according to the values ​​corresponding to the matrix elements.

5. The method for constructing a multi-label graph structure based on a conditional probability matrix and GCN as claimed in claim 1, characterized in that: The step 4 comprises: Step 41: obtaining a plurality of adjacent nodes contained in the binary adjacency matrix, inputting the binary adjacency matrix into a preset GCN for convolution processing, and obtaining node relationships between different adjacent nodes; Step 42: Match the corresponding label combination for each adjacent node, build a label combination relationship according to the node relationship, and generate an external feature for each label combination; Step 43: Capture the convolution features corresponding to each of the adjacent nodes in the convolution processing result, and establish the convolution information of the label combination in combination with the corresponding external features.

6. The method for constructing a multi-label graph structure based on a conditional probability matrix and GCN as claimed in claim 1, characterized in that: The step 5 comprises: Step 51: establishing a shadow distribution feature corresponding to the label combination according to the binary information, and establishing a pixel distribution feature corresponding to the label combination according to the convolution information; Step 52: respectively obtain the presentation features of each label combination in the binarized neighborhood matrix, and construct label-related features corresponding to each label combination in combination with the corresponding shadow distribution features and pixel distribution features; Step 53: Draw a label image corresponding to the combined label according to the label-related features, fuse the label images according to the presentation features to generate a multi-label graph structure of the multi-label data set and display it.

7. The method for constructing a multi-label graph structure based on a conditional probability matrix and GCN as claimed in claim 1, characterized in that: Also includes: The label names corresponding to each label combination are marked in the multi-label graph structure respectively, and the graph area corresponding to each label combination is obtained and displayed.

8. The method for constructing a multi-label graph structure based on a conditional probability matrix and GCN as claimed in claim 7, characterized in that: Also includes: Identifying nodes and edges contained in the multi-label graph structure; Respectively identifying a plurality of first labels corresponding to each of the nodes and a plurality of second labels corresponding to each of the edges; Calculating the first label sparsity corresponding to each of the nodes and the second label sparsity corresponding to each of the edges, and determining the structural key area of ​​the multi-label graph structure; The key area is highlighted in the multi-label graph structure.

9. The method for constructing a multi-label graph structure based on a conditional probability matrix and GCN as claimed in claim 1, characterized in that: Also includes: The relevant information contained in the multi-label graph structure is searched according to the execution instruction issued by the user, and transmitted to the corresponding terminal for display.

10. The method for constructing a multi-label graph structure based on a conditional probability matrix and GCN as claimed in claim 1, characterized in that: Also includes: The domain information contained in the multi-label graph structure is identified according to the graph structure domain uploaded by the user and displayed.

Images