Image clustering method based on granular ball structure and clustering evaluation index

Abstract

The invention discloses an image clustering method based on a granular ball structure and a clustering evaluation index, and the method comprises the steps: obtaining an image sample, extracting pixel points in an image, and obtaining a group of two-dimensional data sets D after preprocessing; taking all data of the two-dimensional data set D as an initial pellet, and performing continuous splitting to obtain a stable pellet set; filtering the stable granular ball set, respectively calculating the central points of all granular balls in the stable granular ball set, and constructing a graph structure comprising all the central points by using a minimum spanning tree algorithm based on the central points of all the granular balls to obtain a minimum spanning tree MST; different numbers of maximum weight edges are removed from a minimum spanning tree (MST) to obtain different clustering results, average clustering evaluation index scores corresponding to the different clustering results are calculated, the clustering result corresponding to the minimum average clustering evaluation index score is obtained, and then the clustering result of the image is obtained. According to the invention, the accuracy of an image clustering result and the robustness of the system can be improved.

Description

Technical Field

[0001] The invention belongs to the technical field of image cluster analysis, and in particular relates to an image clustering method based on a granular sphere structure and a clustering evaluation index. Background Art

[0002] Image clustering is an unsupervised learning method in computer vision and image processing. Its main task is to automatically group a set of images according to the similarity of their contents. It analyzes the features of the image (such as color, texture, shape, structure, etc.) and groups similar images together to complete the classification and grouping tasks. Image clustering is widely used in image segmentation, image retrieval, medical image analysis, satellite remote sensing image processing, and image classification. Although image clustering has a wide range of applications in automated image analysis, how to evaluate the clustering effect remains a challenging problem. Internal clustering validity indicators are a common method for evaluating clustering quality, which aims to judge the quality of clustering results by the compactness within the cluster and the separation between clusters. Internal validity indicators rely entirely on the data characteristics of the clustering results themselves without the use of external label information, so they are particularly important for unsupervised learning tasks. Traditional internal clustering evaluation methods include Davies-Bouldin Index (DBI), Calinski-Harabasz Index (CH), Silhouette Index (SIL), and so on.

[0003] The SIL method is a cohesive effectiveness index widely used in clustering effect evaluation. It measures the balance between the compactness of a data point and its cluster and the separation between clusters, and provides a comprehensive score to help the present invention evaluate the quality of clustering results. SIL is calculated for each data point and the overall score of the cluster is obtained. Its value ranges from -1 to +1, where +1 means that the data point is well clustered in the correct cluster, with high compactness within the cluster and significant separation from other clusters. 0 means that the data point is on the boundary of two clusters, and the similarity within and between clusters is similar. -1 means that the data point is incorrectly clustered into a dissimilar cluster, with poor compactness within the cluster and small separation from other clusters. The closer the value of the silhouette coefficient is to +1, the better the clustering effect; the closer it is to -1, the poorer the clustering effect, and the data point may be misclassified. SIL may not perform well when dealing with non-spherical clusters because it assumes that the shape of the cluster is spherical or uniform. In the case of noisy data, SIL may be affected because noise points may cause inaccurate calculations of intra-cluster compactness and inter-cluster separation. In the SIL method, all data points will participate in the calculation of compactness and separation, which leads to low efficiency when processing large data information. Based on this problem, the recent LCCV (Local Cores-based index, LCCV) clustering verification method has improved the SIL method. The LCCV method first processes the original data information and obtains their natural neighbor relationship, and then obtains the local density core points through the natural neighbor relationship, that is, some data points with high local density. These points are used as representative points. On this basis, a local core-based clustering effectiveness index LCCV is proposed to improve the performance of the SIL method. LCCV selects the local core points with the local maximum density as representative points, and constructs a graph between local cores to calculate compactness and separation. This takes into account the core local structure of the original data set, can identify some non-convex cluster structures, and improves the evaluation efficiency.

[0004] Although LCCV has the above advantages, the process of obtaining local density core points is complicated. It involves calculating natural neighbor relationships, building more relationship matrices, finding local density core points through natural neighbor relationships, building directed saturated neighbor graphs on the core points, and then using hierarchical clustering methods to merge the initial core points. The whole process is relatively complicated, and each merge requires re-composition. In addition, LCCV can recognize data containing spherical and non-spherical shapes, but cannot process noisy data. In actual application scenarios, image data is often complex, large in scale and contains some noise data. This method is inefficient and cannot be applied to actual image clustering analysis. Summary of the invention

[0005] The purpose of the present invention is to solve or alleviate some limitations or problems of the prior art, and to propose an image clustering method based on granular structure and clustering evaluation index, which can be better applied to the clustering result evaluation of complex images and can also perform cluster analysis on large-scale data sets more efficiently.

[0006] The present invention adopts a new way to obtain representative points. The method uses a coarse-grained granular sphere structure to represent the original data information, and constructs a minimum spanning tree on the granular sphere structure. The natural distance relationship and the ability to quickly divide different clusters of the minimum spanning tree are utilized. The intra-cluster compactness and inter-cluster separation are calculated by iteratively cutting the edges of the minimum spanning tree, and the average evaluation scores corresponding to the clustering results of different numbers of clusters are obtained, so as to efficiently evaluate the best clustering result.

[0007] In order to solve the above problems, the present invention provides an image clustering analysis method based on particle-sphere structure, the method comprising:

[0008] S1: Obtain image samples, extract pixel points in the image, and obtain a set of two-dimensional data sets D after preprocessing;

[0009] S2: Take all the data of the two-dimensional data set D as an initial sphere and add it to the sphere set, and continuously split each sphere in the sphere set until the number of spheres no longer changes, thus obtaining a stable sphere set;

[0010] S3: Filter the stable sphere set and calculate the center points of all spheres in the stable sphere set respectively. Based on the center points of each sphere, a graph structure including all center points is constructed using the minimum spanning tree algorithm to obtain the minimum spanning tree MST.

[0011] S4: different clustering results are obtained by removing different numbers of maximum weight edges from the minimum spanning tree, and the average clustering evaluation index scores corresponding to the different clustering results are calculated, and the clustering result corresponding to the minimum average clustering evaluation index score is obtained;

[0012] S5: According to the one-to-one correspondence between the spheres and the centers of the spheres, each sphere is assigned to the same cluster as the center of each sphere in the corresponding clustering result, so as to obtain the final clustering result of the graph.

[0013] Beneficial effects of the present invention:

[0014] 1. The present invention uses an adaptive sphere structure to represent the original data. The sphere generation process is adaptive, does not require any parameters, and does not require complex splitting logic and complex matrix relationship construction. The number of spheres is much smaller than the original data set, and can reduce the impact of some noise data. Therefore, clustering validity analysis based on spheres is very efficient, and is particularly suitable for clustering analysis of large-scale data sets.

[0015] 2. The present invention constructs a minimum spanning tree on the sphere. This construction process is fast and only needs to be constructed once. Different numbers of clusters are obtained by iteratively cutting edges, and the average clustering evaluation index scores of different numbers of clusters are calculated. This method makes full use of the characteristics of the minimum spanning tree, accurately obtains the compactness and separation, and greatly improves the accuracy of the evaluation.

[0016] 3. The calculation method of compactness and separation in the present invention is not affected by the cluster shape and is suitable for the evaluation of data sets of any shape, thereby improving the flexibility of practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flowchart of steps of an embodiment of the present invention;

[0018] Figure 2 A schematic diagram of the logic flow of the particle-sphere splitting in an embodiment of the present invention;

[0019] Figure 3 This is an example diagram of the process of spherule splitting in an embodiment of the present invention;

[0020] Figure 4 An example diagram of constructing a minimum spanning tree MST in an embodiment of the present invention;

[0021] Figure 5 This is an example diagram of the original data set used in the simulation verification of the present invention;

[0022] Figure 6 This is an example diagram of simulation results in simulation verification of the present invention. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] The embodiment of the present invention proposes an image clustering method based on particle-sphere structure and clustering evaluation index, referring to Figure 1 As shown, the method includes:

[0025] S1: Obtain image samples, extract pixel points in the image, and obtain a two-dimensional data set D after preprocessing.

[0026] Specifically, in some embodiments, a set of medical images, such as CT scan images, is obtained. Each image is composed of a series of pixels, and each pixel has its corresponding color (or grayscale value). The image data is converted into a set of discrete data, and after preprocessing, a two-dimensional data set D = {p1, p2, p3, ..., p N}, where p i represents the i-th data point of this two-dimensional array, and N represents the number of p in the two-dimensional data set D.

[0027] Figure 2 Schematic diagram of the logic flow of particle-sphere splitting in an embodiment of the present invention.

[0028] S2: Take all the data of the two-dimensional data set D as an initial sphere and add it to the sphere set, and continuously split each sphere in the sphere set until the number of spheres no longer changes, that is, all spheres no longer split, and finally obtain a stable sphere set, which contains all the spheres obtained by splitting.

[0029] In a preferred embodiment, reference Figure 2 As shown, S2 is specifically:

[0030] S201: Initially, the sphere set is empty. All data of the two-dimensional data set D is taken as an initial sphere and added to the sphere set. The initial sphere is pre-split, and the new sphere obtained by the split is added to the sphere set so that the sphere set can be traversed and split subsequently.

[0031] S202: Traverse all the spheres in the sphere set, calculate the center c of the current sphere, and calculate the distribution metric value DM of the sphere based on the center point c of the current sphere.

[0032] Specifically: Assume that the current sphere includes n data points x1, x2, x3, ..., x n ; Calculate the average value of all data points in the current sphere to get the center c of the sphere. The calculation formula is:

[0033]

[0034] Among them, c represents the center of the sphere, x j represents the jth data point in the sphere, n represents the number of all data points in the current sphere, r is the radius of the sphere, and r is defined as the farthest distance between the data point in the sphere and the center.

[0035] According to the center point c of the current sphere, the distribution measure (DM) of the sphere is calculated. DM represents the compactness of the data inside the sphere. The present invention believes that the more compact the data inside the sphere is, the better the quality of the sphere is, and it is more conducive to using the sphere center to characterize the data.

[0036] The calculation formula of the distribution measure DM is:

[0037]

[0038] Among them, s represents the sum of the distances of all data points in the current sphere to the sphere center c, and x i represents the i-th data point in the sphere, c represents the center point of the sphere, and n represents the number of all data points in the sphere.

[0039] S203: According to the distribution metric value DM of the current sphere, it is determined whether the current sphere can be split. If it can be split, the new sphere after the split is added to the sphere set. If it cannot be split, it is determined whether all the spheres in the sphere set are no longer split.

[0040] Figure 3 This is an example diagram of the process of spherule splitting in an embodiment of the present invention. Figure 3-1 In the data point, data point represents the data point, and granular-ball center represents the granular-ball center. Figure 3-3 In the figure, So split ball A represents the current ball (parent ball), retain BallA1, and A2 represents the child ball after the split. Figure 3-4 In the above figure, isolated granular-ball means an isolated granular-ball, i.e. a granular-ball with only one data point.

[0041] In some embodiments, reference Figure 3 As shown in Figure 2, the specific process of spherule fission is:

[0042] First, the present invention needs to pre-split the current ball (parent ball) into two child balls, and decide whether to divide the parent ball into the two child balls by judging the relationship between the DM value of the child ball and the DM value of the parent ball.

[0043] The present invention represents this parent ball as GB A , calculate GB A GB A The farthest data point q, then calculate GB A The point p farthest from q, select p and q and GB A The midpoints c1 and c2 of the line connecting the center point c are the two reference points of the two new spheres that were pre-split;

[0044] Calculate GB in sequence A The distance between the data point and c1 and c2 is shorter than that between c1 and c2, so the data point is divided into the same sub-particle ball of c1. If the distance between the data point and c1 is shorter, the data point is divided into the same sub-particle ball of c2. In this way, GB A All data points are divided into two sub-spheres;

[0045] The two sub-balls obtained by division are marked as GB A1 and GB A2 , calculate GB A , GB A1 and GB A2 DM value, if GB A1 and GB A2 The DM value is less than GB A DM value, that is, the distribution metric values ​​DM of the two child balls are both smaller than the distribution metric value DM of the parent ball, the present invention splits the current parent ball into the two child balls and adds them to the particle ball set, otherwise it does not split.

[0046] S204: When the number of spheres in the sphere set no longer changes (ie, all spheres no longer split), the sphere generation ends, and the final stable sphere set (GBS) is obtained.

[0047] In the existing technology, the local density core points are used as representative points to participate in the calculation and clustering of clustering evaluation indicators. The process of obtaining local density core points in this method is complicated and involves calculating natural neighbor relationships. This process requires iterative calculation of the distances between many point pairs to determine the neighbor relationship, and requires the construction of more relationship matrices. Then, the local density core points are found through natural neighbor relationships. The whole process has a high time complexity and is very inefficient in obtaining local density core points on large-scale data sets. In fact, the existing technology needs to rebuild the complex graph structure and calculate the indicators every time when evaluating different clustering results, so this step is also inefficient.

[0048] The present invention uses the center of the particle ball to construct the minimum spanning tree. The particle ball generation process is very fast. Only simple splitting judgment conditions are needed to continuously split the particle ball, so the particle ball center can be quickly obtained and used as a representative point. Based on the particle ball center, only one minimum spanning tree needs to be constructed, which can be used for the evaluation of each subsequent clustering result.

[0049] S3: Filter the stable sphere set and calculate the center points of all spheres in the stable sphere set respectively. Based on the center points of each sphere, a minimum spanning tree algorithm is used to construct a graph structure including all center points to obtain a minimum spanning tree MST.

[0050] The step S3 specifically includes:

[0051] S301: Filter out isolated spheres in the stable sphere set, where an isolated sphere is a sphere containing only one data point. Based on the center points of each sphere, a graph structure including all center points is constructed using a minimum spanning tree algorithm to obtain a minimum spanning tree MST.

[0052] S302: Calculate the center points of all spheres in the stable sphere set (GBS) respectively, which is expressed as centers, where |centers|=m, and m represents the number of center points of all spheres in the stable sphere set (GBS).

[0053] Figure 4 This is an example diagram of constructing a minimum spanning tree in an embodiment of the present invention. Figure 4 In the undirected weighted graph, the top three edges in terms of weight are removed, such as Figure 4 If we add three red edges, we can get four clusters, C1 to C4. MST stands for minimum spanning tree.

[0054] S303: Reference Figure 4 As shown in FIG. 1 , based on the center points of each sphere, a minimum spanning tree algorithm is used to construct a graph structure including all center points to obtain a minimum spanning tree MST.

[0055] Minimum spanning tree is a graph algorithm that connects all points in a graph with minimal edge weight. D =(V TD ,E TD ) represents the minimum spanning tree corresponding to the two-dimensional data set D, G D =(V D ,E D ) represents the undirected weighted graph corresponding to the minimum spanning tree, V D represents the set of vertices in an undirected weighted graph, E D represents the set of edges in an undirected weighted graph, w e represents the weight of the edge, w e Usually expressed as the distance between data points.

[0056] In order to facilitate the subsequent cutting of different numbers of the maximum weight edges of the minimum spanning tree to obtain different clustering results, the embodiment of the present invention divides the minimum spanning tree T D Each edge in the e Sort in ascending order.

[0057] Since it is unknown how many clusters the image dataset actually has, the purpose of the present invention is to combine the clustering evaluation index to obtain the optimal number of clusters and the corresponding clustering results. Therefore, it is necessary to analyze different numbers of clusters, and then use the clustering evaluation index to analyze the clustering results corresponding to different numbers of clusters. The existing clustering algorithm only involves clustering, and no indicators are applied for analysis during the clustering process, so it is impossible to determine whether the obtained clustering result is the optimal clustering result. In order to obtain the best clustering result for the current data set, the embodiment of the present invention designs a clustering evaluation index GBVI, calculates the clustering evaluation index score for each clustering result, and obtains the clustering result corresponding to the minimum average clustering evaluation index score.

[0058] S4: different clustering results are obtained by removing different numbers of maximum weight edges from the minimum spanning tree, and the average clustering evaluation index scores corresponding to the different clustering results are calculated, and the clustering result corresponding to the minimum average clustering evaluation index score is obtained. As a preferred embodiment, S4 specifically includes:

[0059] S401: Minimum spanning tree T D Each edge in the corresponding undirected weighted graph has a weight w e Sort in ascending order.

[0060] In order to evaluate different clustering results, different clustering results are obtained by removing different numbers of edges from the MST for analysis.

[0061] S402: Remove k-1 edges ranked top by weight in the undirected weighted graph, Where m represents the total number of spheres in the stable sphere set, we get Clustering results.

[0062] Specifically, k-1 edges ranked top by weight in the undirected weighted graph are removed to obtain a k-connected graph. k needs to be iterated from 2 to Where m represents the total number of spheres in the stable sphere set, thus we get Clustering results.

[0063] Need to Analyze each of the clustering results and find the best clustering result.

[0064] S403: Calculate the The average clustering evaluation index GBVI score of each clustering result in the clustering results is obtained, and the clustering result corresponding to the minimum average clustering evaluation index score is obtained.

[0065] Specifically, k clusters are currently obtained, and the present invention needs to calculate the index score corresponding to each cluster, specifically, calculating the intra-cluster compactness (Com) and the inter-cluster separation (Sep).

[0066] Furthermore, the compactness Com(C i ), i∈[1,k], and its calculation formula is as follows:

[0067]

[0068] Among them, Com(C i ) represents the compactness of the i-th cluster, C i represents the i-th cluster, T i represents the i-th cluster Ci The minimum spanning tree of j Indicates T i The jth edge of this minimum spanning tree, w(e j ) indicates T i The weight of the j-th edge in .

[0069] Furthermore, the separation degree Sep(C i ), i∈[1,k], and its calculation formula is as follows:

[0070]

[0071] Among them, Com(C i ) represents the compactness of the i-th cluster, C i represents the i-th cluster, dist(y i ,y j ) represents the i-th cluster C i Any data point y in i and the jth cluster C j Any data point y in j The distance between (i.e., Euclidean distance), y i Represents any data point in the i-th cluster, y i Represents any data point in the i-th cluster.

[0072] Furthermore, based on the obtained compactness Com(C i ) and resolution Sep(C i ), calculate the i-th cluster C i The clustering evaluation index score GBVI (C i ), the specific calculation formula is:

[0073]

[0074] The clustering evaluation index score GBVI (C i ). Next, the present invention needs to calculate the clustering evaluation index scores GBVI of all the remaining clusters in the current clustering result. The k clustering evaluation index scores GBVI obtained are averaged to obtain the average clustering evaluation index, i.e., avgGBVI(k), which is calculated as follows:

[0075]

[0076] The GBVI (C i) is calculated based on the intra-cluster compactness and inter-cluster separation. Cluster analysis believes that the closer the distance between the points in each cluster, the lower the compactness, which means that the members in the cluster are more similar. This is the intra-cluster compactness. The larger the minimum distance between different clusters, the higher the separation, which means that the distance between clusters is farther and the separation is more obvious. This is the intra-cluster separation. i ) is defined by the ratio of compactness to separation. The present invention hopes that the smaller the compactness value, the better, and the larger the separation value, the better. Therefore, the present invention believes that the smaller the GBVI (C i ) value represents the current cluster C i In order to evaluate the quality of the overall cluster, the present invention calculates the average GBVI of all clusters of the current clustering result, that is, avgGBVI(k), so as to evaluate the overall clustering result more reasonably and effectively.

[0077] After calculating the average index scores for different clustering results, the clustering result corresponding to the smallest average index score is obtained. Specifically, according to the definition of GBVI and avgGBVI, the present invention expects to obtain a small index score, and the smaller the index score, the better the clustering result effect. Therefore, the present invention evaluates the overall clustering result by calculating the average value of the index score, and selects the clustering result corresponding to the smallest average index score avgGBVI.

[0078] As an example, a certain image data set is subjected to sphere splitting to obtain a stable sphere set, and a minimum spanning tree MST is constructed according to the center points of each sphere in the stable sphere set. When clustering is performed according to the minimum spanning tree, it is not known how many clusters MST can actually be clustered into at the beginning, and the actual cluster division of MST is not known, so the image data set is initially clustered into 2 clusters (i.e., k=2, k represents the number of clusters) to obtain the corresponding clustering result; clustering is performed using k=3, that is, the image data set is clustered into three clusters to obtain the corresponding clustering result; in this way, different clustering cluster numbers k are iterated to obtain different clustering results, and each clustering result of the present invention is evaluated using the clustering evaluation index GBVI; finally, each clustering cluster number k is iterated and calculated to completion, and the minimum average clustering evaluation index score GBVI(k) is the index score that is considered to be the best, so the k in the GBVI(k) of this score is the optimal number of clusters of this data set analyzed using the index in the embodiment of the present invention, and the clustering result corresponding to this optimal number of clusters is the final clustering result.

[0079] S5: According to the one-to-one correspondence between the spheres and the centers of the spheres, each sphere is assigned to the same cluster as the center of each sphere in the corresponding clustering result, so as to obtain the final clustering result of the graph.

[0080] Simulation verification:

[0081] Simulation software or platform: The simulation system is an HP 288ProG6Microtower PC with Windows 10 Professional operating system, equipped with Intel(R)Core(TM)i710700CPU@2.10GHz, 32GB RAM, and the simulation software is PyCharm 2023.1.4.

[0082] Figure 5 An example diagram of the original data set used in the simulation verification of the present invention.

[0083] Reference Figure 5 As shown, the present invention uses the original data set from the Kaggle data set website (https: / / www.kaggle.com) during simulation.

[0084] Figure 6 An example diagram of the results of the present invention in simulation verification.

[0085] Reference Figure 6 As shown, the present invention uses the center of the sphere to represent the original data set, which is effective and can remove some noise data to obtain a clearer data distribution. It is conducive to building a more accurate clustering evaluation index on this basis. And the amount of data in the center of the sphere is much less than the original data, which provides a good solution for clustering large data sets.

[0086] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which can include: ROM, RAM, disk or CD, etc.

[0087] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An image clustering method based on particle-sphere structure and clustering evaluation index, characterized in that: include: S1: Obtain image samples, extract pixel points in the image, and obtain a set of two-dimensional data sets after preprocessing; S2: Take all the data in the two-dimensional data set as an initial sphere and add it to the sphere set, and continuously split each sphere in the sphere set until the number of spheres no longer changes, thus obtaining a stable sphere set; S3: Filter the stable sphere set and calculate the center points of all spheres in the stable sphere set respectively. Based on the center points of each sphere, a graph structure including all center points is constructed using the minimum spanning tree algorithm to obtain the minimum spanning tree MST. S4: different clustering results are obtained by removing different numbers of maximum weight edges from the minimum spanning tree MST, and the average clustering evaluation index scores corresponding to the different clustering results are calculated, and the clustering result corresponding to the minimum average clustering evaluation index score is obtained; S5: According to the one-to-one correspondence between the spheres and the sphere centers, each sphere is assigned to the same cluster as the sphere centers in the corresponding clustering result, so as to obtain the final clustering result of the image.

2. The image clustering method based on particle-sphere structure and clustering evaluation index according to claim 1, characterized in that: The S2 specifically includes: Initially, the sphere set is empty. All data of the two-dimensional data set D is taken as an initial sphere and added to the sphere set. The initial sphere is pre-split, and the new sphere obtained by the split is added to the sphere set. Traverse all the balls in the ball set, calculate the center c of the current ball, and calculate the distribution metric DM of the ball based on the center point c of the current ball; According to the distribution measurement value DM of the current sphere, it is necessary to determine whether the current sphere can be split. If it can be split, the new sphere after the split is added to the sphere set. If it cannot be split, it is necessary to determine whether all the spheres in the sphere set will not split again. When all the spheres in the sphere set no longer split, the sphere generation ends and a stable sphere set is obtained.

3. The image clustering method based on particle-sphere structure and clustering evaluation index according to claim 1, characterized in that: The S4 specifically includes: The minimum spanning tree T D Each edge in the corresponding undirected weighted graph has a weight w e Sort in ascending order; Remove the k-1 edges ranked top by weight in the undirected weighted graph, Where m represents the total number of spheres in the stable sphere set, we get Clustering results; Calculate the The average clustering evaluation index score of each clustering result in the clustering results is obtained, and the clustering result corresponding to the minimum average clustering evaluation index score is obtained.

4. The image clustering method based on particle-sphere structure and clustering evaluation index according to claim 3, characterized in that: According to the For each of the clustering results, the clustering evaluation index score of each clustering result is calculated and the average value is obtained to obtain the average clustering evaluation index score.

5. The image clustering method based on granular sphere computing according to claim 4 is characterized in that: The cluster evaluation index score of each cluster includes the compactness and separation of each cluster, which are respectively expressed as: Among them, Com(C i ) represents the compactness of the i-th cluster, C i represents the i-th cluster, T i represents the i-th cluster C i The minimum spanning tree of j Indicates T i The jth edge of this minimum spanning tree, w(e j ) indicates T i The weight of the jth edge in ; Among them, Com(C i ) represents the compactness of the i-th cluster, C i represents the i-th cluster, dist(y i ,y j ) represents the i-th cluster C i Any data point y in i and the jth cluster C j Any data point y in j The distance between i Represents any data point in the i-th cluster, y i Represents any data point in the i-th cluster.

6. The image clustering method based on granular sphere computing according to claim 2, characterized in that: Calculate the average value of all data points in the sphere and get the center c of the sphere.

7. The image clustering method based on granular sphere computing according to claim 2 or 6, characterized in that: According to the center point c of the current ball, the distribution measurement value DM of the current ball is calculated, and the calculation formula is: Among them, s represents the sum of the distances of all data points in the current sphere to the sphere center c, and x i represents the i-th data point in the sphere, c represents the center point of the sphere, and n represents the number of all data points in the sphere.

Images