CN111626321A

CN111626321A - Image data clustering method and device

Info

Publication number: CN111626321A
Application number: CN202010260470.7A
Authority: CN
Inventors: 孙林; 秦小营; 孙全党; 李文凤
Original assignee: Henan Normal University
Current assignee: Henan Normal University
Priority date: 2020-04-03
Filing date: 2020-04-03
Publication date: 2020-09-04
Anticipated expiration: 2040-04-03
Also published as: CN111626321B

Abstract

The invention relates to a method and a device for clustering image data, and belongs to the technical field of image processing. The invention firstly proposes a self recommendation strategy to determine a local central point, eliminates the subjectivity of manually selecting a cluster center, and solves the problem that the central point of a cluster with lower density is neglected; distributing residual data points by taking the local central point as a micro-cluster central point so as to generate a plurality of micro-clusters; and finally, a micro-cluster combination strategy is provided, and micro-clusters with fewer boundary points in the micro-clusters and the adjacent clusters are combined, because the micro-clusters are closer to the center of the cluster and are more likely to be in the same cluster with the micro-cluster, the clustering result of the invention is more accurate.

Description

Image data clustering method and device

Technical Field

The invention relates to a method and a device for clustering image data, and belongs to the technical field of image processing.

Background

Clustering analysis is an unsupervised classification method. The method aims to divide a data set without classification labels into a plurality of clusters, ensures that objects in the clusters are similar to each other and objects between the clusters are dissimilar, and can be applied to the fields of optimization analysis, image segmentation, bioinformatics and the like.

The image data can be classified by Clustering method, and the Clustering algorithm (DPC) is a kind of Clustering proposed by Alex and LaioThe methods, because of their effectiveness, novelty and simplicity, have become more and more widely used in recent years. The proposal of DPC is based on two assumptions: (1) the cluster center is surrounded by neighbor points with lower density than the cluster center; (2) the distance between the cluster centers is relatively large. To this end, two concepts are proposed: (1) data point x_iLocal density of in rho_iRepresents; (2) ratio data point x_iHigh density and distance x_iNearest point to x_iA distance of_iAnd (4) showing. On the basis of the above two assumptions, the authors of the DPC algorithm propose the general idea of the algorithm: first selecting rho_iValue (c),_iThe data point with the larger value is taken as the cluster center point, and then the non-center point is allocated to the cluster with the density higher than the self and the data point closest to the self. Although the DPC has good performances in all aspects, certain defects still exist: cut-off distance d_cThe choice of (a) affects the accuracy of the clustering, i.e. d_cThe small variation of the value influences the density of data points and the selection of a clustering center, and further influences the final clustering result; the clustering center needs to be selected manually, has certain subjective factors and influences the objectivity of the final clustering result. And for data sets with large density differences between clusters, DPC tends to ignore the cluster centers of the less dense clusters.

On the basis, many scholars make corresponding improvements on the density peak value clustering algorithm. For example: the high pozzing and the like provide a density peak value clustering algorithm based on density proportion, namely R-DPC, the algorithm introduces the density proportion into DPC, the identification degree of a cluster with lower density in data is improved by calculating the density ratio of sample data, and the accuracy of integral clustering is further improved; schroewina et al propose an improved density peak value clustering algorithm combined with k nearest neighbor, and provide a new local density measurement method to describe the distribution condition of each sample in space, thereby effectively improving the clustering quality; xie et al propose a density peak search and point distribution algorithm based on fuzzy weighted k-nearest neighbor technology to solve the problem of inconsistency of data point density measurement methods in DPC algorithm. The improvement improves the performance of the density peak clustering algorithm to a certain extent, but the improvement is difficult to deal with the data set with variable density and has higher time complexity.

Disclosure of Invention

The invention aims to provide a method and a device for clustering image data, which are used for solving the problem of poor image data clustering effect at present.

The present invention provides a method for clustering image data to solve the above technical problems, the method comprising the steps of:

1) acquiring image data to be clustered, performing dimension reduction processing on the image data, taking each image as a data point, and determining the density of each data point;

2) for each data point, judging whether the density of the data point is greater than that of the data points in the neighborhood, if so, recommending the data point as a local central point, taking the local central point as a center, distributing the rest data points to the local central point, and generating a micro-cluster;

3) determining the number of boundary points according to a first set proportion, selecting a corresponding number of points with larger boundary degree as the boundary points, and determining whether an adjacent cluster relation is formed between two micro clusters under the condition of not considering the boundary points;

4) determining boundary points according to a second set proportion, judging whether the determined boundary points contain local central points, if so, selecting the micro-cluster with the highest combination degree from the adjacent clusters of the micro-cluster where the local central points are located for combination, and deleting the boundary points;

and continuously determining boundary points from the rest points again according to a second set proportion to obtain a next layer of boundary points, judging whether the next layer of boundary points contain local central points, if so, selecting the micro-cluster with the highest combination degree from the adjacent clusters of the micro-cluster with the local central points to combine, deleting the layer of boundary points until the boundary points with the set times are deleted, and then combining the micro-cluster with the adjacent clusters if data points exist in the adjacent clusters.

The invention also provides a clustering device of image data, which comprises a memory, a processor and a computer program stored on the memory and running on the processor, wherein the processor is coupled with the memory, and the processor executes the computer program to realize the clustering method of the image data.

According to the image clustering method, the local central point is selected by adopting a self-recommendation strategy, so that the subjectivity of manually selecting the cluster center is eliminated, and the problem that the central point of a cluster with low density is ignored is solved; then, generating a micro-cluster by taking the local central point as the central point of the micro-cluster; and then, providing a micro-cluster combination method, and in the process of deleting boundary points layer by layer, selecting a micro-cluster with the highest combination degree from the adjacent clusters of one micro-cluster for combination, wherein the micro-cluster is closer to the center of the cluster and is more likely to be in the same cluster with the micro-cluster, so that the final clustering result is more accurate.

Further, in order to solve the problem that the truncation distance is difficult to select when calculating the density, the density value of each point in the step 1) is determined according to the mutual adjacency of the point and the mean value of the distances between the point and the middle points in the neighborhood.

Furthermore, the invention provides a formula for calculating the density of the data points, and the calculation formula of the density value is as follows:

where ρ (x)_p) Represents the data point x_pDensity of (a), MND (x)_p) Represents the data point x_pMutual adjacency of NG (x)_p) Represents the data point x_pIs in the neighborhood, | NG (x)_p) | represents a data point x_pThe number of data points in the neighborhood,

represents the data point x_pTo data point x_qWeighted Euclidean distance of (1), at data point x_pOf k neighbors, if x_qIs x of the data point_pThe first neighbor, then_pqIs that the larger the value of (A), x_qTo x_pThe farther away; if x_pIs x of the data point_qThe first one of the neighbors is,_qpthe value of (c) is.

Further, in order to reflect the internal relationship between the data points, the distance between two data points is a weighted euclidean distance, and the pearson correlation coefficient of the two points is used as a weight.

Further, in order to accurately determine whether a data point is a boundary point, a concept of data point boundary degree is provided, the larger the value of the boundary degree is, the more likely the data point becomes a boundary point, and a calculation formula of the boundary degree of the data point is as follows:

wherein bp (x)_p) Represents the data point x_pBoundary degree, | kNN (x)_p) | represents data point X_pNumber of neighbors of, kNN (x)_p) Represents the data point x_pThe number of k neighbor points of (a),

is a data point X_pAnd data point X_qPearson correlation coefficient between, m is the dimensionality of the data set.

Further, in order to accurately determine the relationship between the micro-clusters, the judgment basis of the adjacent micro-clusters is as follows:

wherein NG (x)_p) Represents the data point x_pNeighborhood of c_iDenotes the ith micro-cluster formed, R (c)_i，c_j) Represents a micro cluster c_iAnd micro-cluster c_jForm a neighboring cluster relationship.

Further, after the micro-cluster is formed, the method changes the local central point of the micro-cluster into the mean point of all data points in the micro-cluster, and if the mean point does not exist, the data point closest to the mean point in the micro-cluster is changed.

Further, the invention provides a calculation formula of the binding degree, wherein the calculation formula of the binding degree is as follows:

|c_ji represents a micro-cluster c in an initial state_jThe number of data points, | c_jtI represents the micro-cluster c after deleting the boundary point of the t layer_jThe number of the data points in the data stream,

the larger the value of (A), c_jThe closer to the cluster core region, c_jAnd c_iThe higher the degree of bonding of (a),

representing a weighted euclidean distance between two local center points.

Further, in order to avoid clustering some abnormal points individually, the method further includes a process of adjusting the clusters after the micro-cluster merging:

and after merging the homologous clusters, judging whether the number of the data points in the obtained cluster is less than or equal to 0-5% of the total number of the data points, if so, taking the cluster as an abnormal cluster, and distributing the data points in the cluster to the non-abnormal cluster where the data points closest to the cluster are located.

Drawings

FIG. 1 is a flow chart of an image data clustering method of the present invention;

fig. 2-a is a schematic diagram of a preliminary clustering result when a Pathbased data set (k is 7) is obtained in the embodiment of the present invention;

fig. 2-b is a schematic diagram of a preliminary clustering result when the Pathbased dataset (k is 11) is obtained in the embodiment of the present invention;

fig. 2-c is a schematic diagram of the preliminary clustering result of the frame data set (k 7) in the embodiment of the present invention;

fig. 2-d is a schematic diagram of the preliminary clustering result of the frame data set (k 18) in the embodiment of the present invention;

fig. 2-e is a schematic diagram of a preliminary clustering result of a helical dataset (k is 6) in the embodiment of the present invention;

fig. 2-f is a schematic diagram of a preliminary clustering result of a helical dataset (k 13) according to an embodiment of the present invention;

FIG. 3-a is a schematic diagram of a final clustering result of the Pathbased data set in the embodiment of the present invention;

FIG. 3-b is a diagram illustrating the final clustering result of the frame data set in the embodiment of the present invention;

3-c are schematic diagrams of final clustering results of the Spiral data set in the embodiment of the invention;

FIG. 4-a is a schematic diagram of a final clustering result of Olivetti Faces data sets in the embodiment of the present invention;

FIG. 4-b is a diagram illustrating the final clustering result of Olivetti Faces data sets by DPC algorithm;

FIG. 5 is a block diagram showing the structure of an image data clustering apparatus according to the present invention.

Detailed Description

The following further describes embodiments of the present invention with reference to the drawings.

Method embodiment

The invention provides a new image clustering method in consideration of the problem that the clustering effect of the DPC algorithm and other clustering algorithms is poor when clustering image data, and the method provides two improvements on the basis of the existing DPC image clustering: firstly, calculating the density of data points by adopting a weighted Euclidean distance and mutual adjacency calculation method; and secondly, automatically determining a local central point by adopting a self-recommendation strategy, forming a micro-cluster based on the determined local central point, and providing a strategy for combining the micro-cluster with micro-clusters containing fewer boundary points in adjacent clusters to obtain a final clustering result. The implementation flow is shown in fig. 1. The implementation of the method is described in detail below with respect to specific examples.

1. And acquiring image data to be clustered, and preprocessing the image data.

The number of features of the acquired image data is generally large. For example, a 90 × 150 image has 13500 features. Therefore, the dimension reduction processing needs to be performed on the image data, and the PCA algorithm is adopted for the dimension reduction processing in the embodiment. For Olivetti faces datasets, PCA is used to filter out features after the cumulative contribution rate exceeds 90%. In order to eliminate the influence on the experimental result caused by different value ranges among different characteristics, the data needs to be normalized. The normalization process uses the formula:

wherein x_ijIs the value of the ith data point on the jth attribute, x_j' is the value of the ith data point on the jth attribute after normalization.

2. The density of each data point is determined.

For a given data set x_n×m＝{x₁，x₂，...，x_i，...，x_nN is the number of data points, m is the dimension of the data points, x_i＝[x_i1，x_i2，...，x_im]The calculation formula of the data point density in the existing DPC algorithm is as follows:

wherein, formula (1) is for the data set with a larger number of samples, and formula (2) is for the data set with a smaller number of samples. D in formula (1) and formula (2)_cThe values are identical and all represent the truncation distance, d_ijRepresenting the euclidean distance between two data points.

Firstly, calculating a weighted Euclidean distance between two data points by taking a Pearson correlation coefficient between the two data points as a weight; the concept of mutual proximity is then proposed to determine how closely a data point is related to the data points in its neighborhood. The higher the value of mutual adjacency is, the more closely a data point is connected with the data points in its neighborhood, and the greater the density of the data point is. And finally, calculating the density of the data points by combining the weighted Euclidean distance and the mutual adjacency.

1) A weighted euclidean distance between the two data points is calculated.

In the embodiment, the pearson correlation coefficient is used as an influence factor, and a new calculation manner of the distance between data points is defined as follows:

wherein

Represents the data point x_pAnd data point x_qThe weighted euclidean distance between them,

represents the data point x_pAnd data point x_qPearson's correlation coefficient between, x_pjRepresents the data point x_pThe value on its jth attribute. x is the number of_qjRepresents the data point x_qThe value on its jth attribute. m is the dimension of the data set, i.e. the number of data point features.

2) Mutual adjacency of data points and data point density are calculated.

Before determining mutual proximity, mutual proximity relations are first explained. Public placeIt is well known that if x_pK of (a) has x in the neighborhood_q，x_qK of (a) also has x in the neighborhood_pThen x_pAnd x_qForm a mutual adjacent relation. Based on this, the concept of neighborhood of data points is proposed. By data point x_pFor example, x_pIs referred to as being in x_pK in the neighborhood of (a) and x_pA collection of data points that constitute a mutual adjacency. Namely NG (x)_p)＝{x_q|x_q∈kNN(x_p)∧x_p∈kNN(x_q)}，kNN(x_p) Represents a distance data point x_pThe set of the most recent top k data points, NG (x)_p) Denotes x_pA set of data points in a neighborhood. The invention provides a concept of mutual adjacency capable of reflecting the data point density, and the mutual adjacency is calculated in the following way:

MND(x_p) The larger the value of (A), x_pThe greater the density of (a).

The density of data points is inversely proportional to the average distance between data points and directly proportional to the mutual proximity of the data points themselves. Then, a new density calculation mode is proposed by combining the weighted euclidean distance and the mutual adjacency, and the formula specifically adopted is as follows:

wherein, | NG (X)_p) I represents x_pThe number of data points in the neighborhood.

3. The local center points are automatically selected using a self-recommended strategy to form micro-clusters.

For data sets with large density differences between clusters, the DPC can easily ignore the center points of the less dense clusters. Therefore, the invention provides a strategy for automatically selecting the local central point. If x_pIs greater than the density of points in its neighborhood, x is then_pRecommending a local central point (self-recommendation strategy for short), otherwise x_pLocal center points are not recommended. If there are a total of c partsThe center point represents the set of all local center points in one data set with LC. Then, LC ═ LC₁，lc₂，...，lc_i，...，lc_j，...，lc_c}。

By adopting the self-recommendation strategy, points in the cluster with relatively low density have great chance to be selected as local central points, and further the cluster with relatively low density cannot be ignored.

The present embodiment uses the DPC method to assign the remaining data points, i.e., to the cluster having the local center point with higher density and closest distance. After the allocation process is completed, a plurality of "micro-clusters" are formed. And finishing the primary clustering task. With C ═ C₁，c₂，...，c_i，...，c_j，...，c_cDenotes the preliminary clustering result (since there are c local centers, the preliminary clustering result has c micro-clusters), c_iIndicates the ith micro cluster. Since the dimensionality of the image data is higher after dimensionality reduction, the preliminary clustering result is shown on only 3 widely used synthetic data sets in this embodiment. The datasets employed include Pathbased datasets, Flame datasets and Spiral datasets. The preliminary clustering results are shown in FIGS. 2-a, 2-b, 2-c, 2-d, 2-e, and 2-f (it should be noted that different shapes in the figure represent different micro-clusters, but if two non-adjacent micro-clusters are the same shape, they also represent different micro-clusters, black pentagons represent local center points of different micro-clusters, and black numbers represent cluster numbers of different micro-clusters). As other embodiments, besides the DPC method, other existing clustering methods can be used to assign the remaining data points to the cluster where the local center point closest thereto is located, so as to form a micro-cluster.

It can be seen from the figure that for each data set, it is easy to find a suitable k value to get the correct preliminary clustering result, as long as the k value is not too large or too small relative to the total number of data points in its data set. The correct preliminary clustering result means that for data points in each micro-cluster, in the real case, the data points belong to the same cluster.

It should be noted that, in order to ensure that all local center points can form a stable cluster structure, after the remaining data points are distributed, the local center points are changed to the average values of all data points in the micro-cluster, and if such average values do not exist, the local center points are changed to the data points closest to the average values in the micro-cluster.

4. And merging the micro-clusters.

In the present invention, one data set is composed of a plurality of clusters, and one cluster is composed of a plurality of micro-clusters. Therefore, micro-clusters need to be merged into clusters, and the relationship between adjacent clusters is introduced before merging the micro-clusters.

If micro-cluster c_iUnion of data points in the neighborhood of all data points in and micro-cluster c_jIf the members of (c) have intersection, then c is considered_jAnd c_iAnd forming a neighboring cluster relationship. With R (c)_i，c_j) Denotes c_jAnd c_iAdjacent cluster relationship between them. Then

If two micro-clusters satisfy the adjacent cluster relationship, they are called paired clusters. Two micro-clusters are homologous clusters if they are paired clusters and also belong to the same cluster. If two micro-clusters are paired clusters, but they are not in the same cluster, then the two micro-clusters are non-homologous clusters. And if the two micro-clusters meet the adjacent cluster relationship, the two micro-clusters are adjacent to each other. With S_niRepresents a micro cluster c_iThe set of all neighboring clusters.

In the course of micro-cluster merging, deletion boundary points are required layer by layer, and therefore, a method of determining boundary points is described herein. The basis for determining the boundary point is the boundary degree of the data point, and the larger the boundary degree is, the more likely the point is to be a boundary point. For any data point x_pIn other words, x_pWith respect to the distance between k neighbors and x_pThe larger the distance between the adjacent neighbor of (a) and its neighbor (b), the larger the boundary degree of the point. In addition, the weighted Euclidean distance is used for makingIs the distance between two data points, the boundary degree of the data points is calculated as

Wherein,

the number of boundary points is taken as a proportion of the total number of data points. And selecting a corresponding number of points with larger boundary degree as boundary points.

The specific method of merging the micro-clusters is as follows. With c_iFor example, it may form a neighbor cluster relationship with a plurality of micro-clusters, that is, c_iThere are a plurality of adjacent clusters. In these adjacent clusters, there are c_iForm homologous clusters, some and c_iForming a non-homologous cluster. The main purpose is from c_iCan be found in the adjacent cluster of (c)_iConstructing micro-clusters of homologous clusters, and then merging c_iAnd the nanoclusters. In fact, in the region far from the cluster core (in X)₁Indicating) that non-homologous clusters are easier to construct between micro-clusters. In the region near the cluster core (by X)₂Representing), homologous clusters are very easily formed among the micro clusters. So at X₁And X₂In both places, the homologous clusters are found in different ways. Suppose X₁Consisting of multiple layers of boundary points, the three steps of micro-cluster merging are as follows.

1) And after deleting a layer of boundary points, determining whether the adjacent cluster relation is formed among all the micro clusters.

For many data sets, there is a tight cluster-to-cluster relationship. Then, the micro-clusters located on the cluster boundary are highly susceptible to adjacent cluster relationship. Once a neighbor cluster relationship is formed between two micro clusters in different clusters, the two micro clusters may further form a homologous cluster. Therefore, after deleting a layer of boundary points, determining the adjacent cluster relationship (the data points with higher boundary degree are taken as the boundary points according to a set proportion, and the set proportion is generally between 10% and 50% of the total number of the data points), so that the adjacent cluster relationship originally established between two micro clusters in the boundary area of the clusters is not established, and naturally, the two micro clusters cannot form a homologous cluster.

2) And searching for homologous clusters.

Find X₁The homologous cluster in (1). The purpose of this stage is to find X₁Homologous clusters of all micro-clusters in (a). Suppose X₁Has M boundary points (X)₁Is composed of layer-by-layer boundary points), some local central points in M boundary points are defined by l_cjRepresenting a local central point thereof. With c_iRepresentative and l_cjCorresponding micro-cluster with c_jDenotes c_iOne of the adjacent clusters of (a). Then the invention needs to determine c before these M boundary points are deleted_iWho is the homologous cluster of (a). As previously mentioned, the closer the micro-clusters are to the center of the cluster, the more likely it is that homologous clusters will be formed. While the more micro-clusters near the center of the cluster, the fewer boundary points are contained. Therefore, with | c_jI represents a micro-cluster c in an initial state_jThe number of data points in the data set. With | c_jtI denotes the micro-cluster c before M boundary points are deleted_jThe number of data points in (where t represents the boundary point that has been previously deleted t times). Then it is determined that,

the larger the value of (A), the larger the description of c_jThe closer to the center of the cluster, c is illustrated_jAnd c_iThe more likely it is that homologous clusters are formed. Furthermore, the closer the local center points of two micro-clusters are, the more likely the two micro-clusters are in the same cluster. The probability that two micro-clusters are homologous clusters is expressed in terms of binding degree, and the larger the value of the binding degree, the more likely the two micro-clusters are homologous clusters. The degree of binding was calculated as follows

Representing the sum between two local centre pointsThe distance in the form of the euler.

At c_iIn the neighboring cluster of (A), if c_jAnd c_iThe highest degree of binding, then c_jIs that c_iThe homologous cluster of (a). After finding the homologous clusters of all local center points contained in the M boundary points, deleting the M boundary points. If there is no local center point in the M boundary points, the M boundary points are directly deleted, and then the next layer of boundary points are processed, until the boundary points of the specified times are deleted, the above process is stopped. If c is_iIf there is no adjacent cluster, then c is considered_iAnd c_iAnd forming a homologous cluster.

Find X₂The homologous cluster in (1). The above deleted boundary points are X₁The data points in (1). The remaining points in the dataset are X₂Point (2). With l_ciRepresents X₂A local center point of_iIndicating the micro-cluster corresponding to this local center point. As previously mentioned, X₂The micro-clusters in (a) are very likely to constitute homologous clusters. Therefore, for c_iAll neighboring clusters of (1) are considered to be c and c as long as they have data points remaining_iAre homologous clusters. Find X₂The homologous cluster of the micro-cluster where all local center points are located.

Then, for X₁And X₂And combining two micro-clusters in the homologous cluster into one cluster.

3) Reallocating data points in an outlier cluster

After merging the micro-clusters in step 1) and step 2), if the number of data points in a certain cluster is less than the set proportion of the total number of data points, the cluster is considered to be an abnormal cluster, otherwise, the cluster is considered to be a non-abnormal cluster, the set proportion is 0% -5%, and 1% is selected in the embodiment. The reason for the occurrence of an anomalous cluster is either that the points themselves are anomalous points or that the points are relatively less closely related to the core portion of the cluster. For whatever reason, the data points in the abnormal cluster are rebinned. That is, for each data point in an abnormal cluster, it is assigned to the cluster where the data point closest to it is located (this cluster must be a non-abnormal cluster, otherwise, the cluster where the next data point closest to it is located is continuously found). And ending the whole micro-cluster merging process to obtain the final image clustering result.

The algorithm adopted in the image clustering process is named as follows: the direction of binding between micro-clusters is determined by The passage of boundary points (The direction of The between micro-clusters is determined by The passing of boundary points, DUMDPBP), and The steps of The algorithm can be summarized as follows:

device embodiment

The apparatus proposed in this embodiment, as shown in fig. 5, includes a processor and a memory, where a computer program operable on the processor is stored in the memory, and the processor implements the method of the above method embodiment when executing the computer program.

That is, the method in the above method embodiments should be understood that the flow of the image data clustering method may be implemented by computer program instructions. These computer program instructions may be provided to a processor such that execution of the instructions by the processor results in the implementation of the functions specified in the method flow described above.

The processor referred to in this embodiment refers to a processing device such as a microprocessor MCU or a programmable logic device FPGA;

the memory referred to in this embodiment includes a physical device for storing information, and generally, information is digitized and then stored in a medium using an electric, magnetic, optical, or the like. For example: various memories for storing information by using an electric energy mode, such as RAM, ROM and the like; various memories for storing information by magnetic energy, such as hard disk, floppy disk, magnetic tape, magnetic core memory, bubble memory, and U disk; various types of memory, CD or DVD, that store information optically. Of course, there are other ways of memory, such as quantum memory, graphene memory, and so forth.

The apparatus comprising the memory, the processor and the computer program is realized by the processor executing corresponding program instructions in the computer, and the processor can be loaded with various operating systems, such as windows operating system, linux system, android, iOS system, and the like.

As other embodiments, the device can also comprise a display, and the display is used for displaying the diagnosis result for the reference of workers.

Analysis of experiments

To better illustrate the effect of the present invention, the method of the present invention is compared with several existing clustering methods, and the experiment is performed on a processor with 764 bits windows operating system, Matlab 2016b software, 32GB memory, intel (r) core (tm) i7-8700 CPU @3.20 GHz. Experiments were performed using 3 different types of 2-dimensional datasets and one high-dimensional image dataset, as shown in table 1, table 1 describing the basic information of the datasets. Before the experiment begins, the data is normalized to eliminate the influence of different values on the experiment result.

TABLE 1

For the first three data sets, the clustering results of the DUMDPBP algorithm in the invention are shown in FIG. 3-a, FIG. 3-b and FIG. 3-c, and the clustering results are consistent with the actual classification results of the corresponding data sets, which shows that the DUMDPBP algorithm in the invention has good effects on Pathbased data sets, Flame data sets and Spiral data sets. Different shapes in the above figures represent different clusters, and the final clustering into three types is performed for Pathbased datasets, two types is performed for Flame datasets, and three types is performed for Spiral datasets.

For Olivetti faces datasets, because they are composed of a large number of photographs, rather than points one by one in a two-dimensional plane, and the number of clusters is as many as several tens, it is impossible to show their clustering results in different shapes as in the case of synthetic datasets. There is also no other way to illustrate a data set consisting of photographs with a large number of clusters. But in the subsequent section, the clustering result is shown by the evaluation index. Even so, we show the case of cluster center point in the clustering result. It should be noted that the Olivetti faces dataset has 400 photos, and only the photos of the first 100 persons are selected for presentation. FIG. 4-a is the clustering result of the DUMDPBP algorithm. The white dots in the upper right corner of the partial picture represent the selected local center points, and it can be seen from the figure that at least one white dot appears in each row (i.e. each cluster), that is, each cluster selects the local center point, so that each cluster cannot be ignored. Fig. 4-b is a clustering result of the DPC algorithm, and a white dot at the upper right corner in the picture represents a cluster center point selected by the DPC algorithm, and it can be seen from the figure that no cluster center point is selected in the 1 st cluster (the picture in the 1 st row), the 8 th cluster (the picture in the 8 th row), and the 10 th cluster (the picture in the 10 th row), that is, the DPC algorithm will definitely confuse data points in the three clusters with data points in other clusters. In fact, the DPC algorithm assigns all data points in the 3 rd cluster and 8 data points in the 10 th cluster to one cluster, which is a bad case. The comparison shows that the method has great advantages compared with the DPC algorithm.

In order to further know the good and bad of the clustering result of the DUMDPBP algorithm, the invention adopts three evaluation indexes commonly used by people to measure the performance of the clustering algorithm. The three evaluation indexes are: adjusting mutual information AMI, adjusting a landed coefficient ARI and a Fowles-Mallows index FMI. AMI is in the range of [ -1, 1 [)]The ARI value range is [ -1, 1 [ ]]The FMI value range is [0, 1 ]]. All three indexes are that the clustering result is consistent with the actual situation when the index value is larger. If U is equal to { U ═ U₁，…，U_RV ═ V } and V ═ V₁，…，V_CDenotes the data set X ═ X, respectively₁，x₂，…，x_nAnd (5) real division and clustering result division. Let H (U) denote the entropy of the original partition, H (V) denote the entropy of the partition of the clustering result, MI (U, V) denote the mutual information between the original partition U and the partition of the clustering result V, E {. cndot.) denote the original partition U and the partition of the clustering result VThe clustering results partition the expected mutual information between V. And a represents the condition that two data points are in the same cluster under the real condition and the clustering result is in the same cluster. And b represents the condition that the two data points are in the same cluster under the real condition and the clustering result is not in the same cluster. And c represents the condition that the two data points are not in the same cluster under the real condition and the clustering result is in the same cluster. D represents the condition that the two data points are not in the same cluster under the real condition, and the clustering result is not in the same cluster. The above three evaluation indexes are calculated as follows.

The algorithm (called DUMDPBP) of the present invention was compared with the existing 7 clustering algorithms on the above 4 data sets according to the above three criteria, and the comparison results are shown in Table 2.

TABLE 2

TABLE 3

In Table 2, the parameters of DUMDPBP on 4 data sets are shown in Table 3, including the number of neighbors of a data point (k), the number of times of finding homologous clusters (N1), the ratio of boundary points to the total number of data points when determining the relationship of the adjacent clusters (N2), finding homologous clustersThe proportion of boundary points to the total number of remaining data points when clustered (N3). For the Spiral dataset, as long as two micro-clusters form a neighbor cluster relationship, they are homologous clusters due to the obvious interval between clusters. So N is not required to search for homologous clusters₁Set to 0; for the Flame dataset, N is due to the inconspicuous inter-cluster spacing₂The arrangement is relatively large; for N₃The value of (b) is usually less than 0.5. Adjusting parameter d adopted by clustering results of DPC algorithm on Olivetti faces data set_c1.1125. Adjustment of the remaining parameters is all referred to the article published by the author Rui Liu et al in journal "Information Sciences": shared-near-based clustering by fast search and find of densitypeaks.

As can be seen from table 2, for a data set with any shape like Spiral, the values of the dumpbp and the other 5 clustering algorithms on three indexes are all 1, which indicates that the clustering results of the six clustering algorithms are completely the same as the real situation; for the data set with no obvious separation limit among the clusters like the Flame, the upper limit value 1 of the indexes is obtained on three indexes, namely DUMDPBP, FKNN-DPC and DPC, which indicates that the clustering result is completely consistent with the real situation; for the Pathbased dataset, only the values of the DUMDPBP algorithm on three indexes are all 1, while the effects of other algorithms are relatively poor; for Olivetti faces datasets, the DUMDPBP algorithm performs better than other algorithms in three indexes. Therefore, the algorithm of the invention not only has good clustering effect on the synthetic data set, but also has good clustering effect on the image data set in the real world. Therefore, the DUMDPBP algorithm provided by the invention is suitable for various types of data sets, and has a very good application prospect.

The invention combines mutual adjacency and weighted Euclidean distance, proposes a new density calculation mode, avoids artificially setting truncation distance; then, a strategy of self-recommending local central points is provided, and then the rest points are distributed, so that the local central points of the clusters with relatively low density can be selected, and the clusters with relatively low density cannot be ignored; and finally, searching the homologous clusters in the process of deleting boundary points layer by layer, combining the homologous clusters into clusters, and finishing a clustering task. Experiments show that the algorithm provided by the invention obtains better results on data sets with different distribution types, and the application field is wide.

Claims

1. A clustering method of image data, characterized by comprising the steps of:

2. The method for clustering image data according to claim 1, wherein the density value of each point in step 1) is determined based on the mutual proximity of the points and the average of the distances between the point and each point in its neighborhood.

3. The method for clustering image data according to claim 2, wherein the density value is calculated by the formula:

4. The method for clustering image data according to claim 2, wherein the distance between two data points is a weighted euclidean distance, and the pearson correlation coefficient between two data points is used as a weight.

5. The method for clustering image data according to any one of claims 1 to 4, wherein the degree of boundary of the data points is calculated by the formula:

wherein bp (x)_p) Represents the data point x_pBoundary degree, | kNN (x)_p) | represents a data point x_pNumber of neighbors of, kNN (x)_p) Represents the data point x_pThe number of k neighbor points of (a),

6. The method for clustering image data according to any one of claims 1 to 4, wherein the adjacent cluster relationship is determined by:

wherein NG (x)_p) Represents the data point x_pNeighborhood of c_iDenotes the ith micro-cluster formed, R (c)_i,c_j) Represents a micro cluster c_iAnd micro-cluster c_jForm a neighboring cluster relationship.

7. The method for clustering image data according to claim 1, further comprising performing center point adjustment on the obtained micro-cluster, changing a local center point of the micro-cluster to a mean point of all data points in the micro-cluster, and if the mean point does not exist, changing to a data point in the micro-cluster closest to the mean point.

8. The method for clustering image data according to claim 1, wherein the degree of combination is calculated by the following formula:

representing a weighted euclidean distance between two local center points.

9. The method for clustering image data according to claim 1, further comprising a process of adjusting the clusters after the merging of the micro clusters:

10. An apparatus for clustering image data, comprising a memory and a processor, and a computer program stored in the memory and running on the processor, wherein the processor is coupled to the memory, and wherein the processor executes the computer program to implement the method for clustering image data according to any one of claims 1 to 9.