CN111626321A - Image data clustering method and device - Google Patents

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

A kind of image data clustering method and device

technical field

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

Background technique

Cluster analysis is an unsupervised classification method. The goal is to divide the data set without classification labels into several clusters, and at the same time to ensure that the objects in the clusters are similar to each other and the objects between clusters are not similar. It can be applied to optimization analysis, image segmentation, bioinformatics and many other fields.

The clustering method can be used to classify the image data. The density peak clustering algorithm (Clustering by fast search and find of density peaks, DPC) is a clustering method proposed by Alex and Laio. It is effective, novel and simple. Its application is getting wider and wider. The proposal of DPC is based on two assumptions: (1) the cluster centers are surrounded by neighbors with lower density than it; (2) the distance between the cluster centers is relatively far. To this end, two concepts are proposed: (1) the local density of the data point _xi , represented by ρ _i ; (2) the distance from the point with a higher density than the data point _xi and the closest to _xi to _xi , represented by δ _i express. On the basis of the above two assumptions, the author of the DPC algorithm proposed the general idea of the algorithm: first, select the data points with larger ρ _i and δ _i values as the cluster center points, and then assign the non-center points to the density higher than itself, And the cluster where the data point closest to itself is located. Although the performance of DPC is good in all aspects, there are still certain defects: the selection of the truncation distance dc _affects the accuracy of clustering, that is, the slight change of the _dc value affects the density of data points and the selection of cluster centers, and then Affect the final clustering result; the clustering center needs to be selected manually, which has certain subjective factors and affects the objectivity of the final clustering result. And for datasets with large density differences between clusters, DPC tends to ignore the cluster centers of clusters with low density.

On this basis, many scholars have made corresponding improvements to the density peak clustering algorithm. For example, Gao Shiying et al. proposed a density peak clustering algorithm based on density ratio, namely R-DPC, which introduced the density ratio into DPC, and improved the identification of clusters with less density in the data by calculating the density ratio of the sample data. And then improve the accuracy of the overall clustering; Xue Xiaona et al. proposed an improved density peak clustering algorithm combined with k-nearest neighbors, and gave a new local density measurement method to describe the distribution of each sample in space, which effectively improved the clustering quality. Xie et al. proposed a density peak search and point assignment algorithm based on the fuzzy weighted k-nearest neighbor technique to solve the problem of inconsistent data point density measurement methods in the DPC algorithm, which uses k-nearest neighbor information to define the local density of points and Search and discover cluster centers. The above improvements improve the performance of the density peak clustering algorithm to a certain extent, but the above improvements are difficult to deal with variable density datasets and have high time complexity.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method and device for clustering image data, so as to solve the problem that the current image data clustering effect is not good.

The present invention provides a clustering method for image data in order to solve the above-mentioned technical problems, and the clustering method comprises the following steps:

1) Obtain the image data to be clustered, perform dimensionality reduction processing on the image data, use each image as a data point, and determine the density of each data point;

2) For each data point, determine whether its density is greater than the density of the data points in its neighborhood, if so, recommend it as the local center point, and use the local center point as the center, and assign the remaining data points to the local center. point to generate microclusters;

3) Determine the number of boundary points according to the first set ratio, select a corresponding number of points with a larger boundary degree as the boundary points, and determine whether the adjacent cluster relationship is formed between the two microclusters without considering the boundary points;

4) Determine the boundary points according to the second set ratio, and judge whether the determined boundary points include local center points. If the boundary points include local center points, select the microclusters where these local center points are located from the neighboring clusters. Merge the microclusters with the highest degree of association, and delete the boundary points;

Continue to determine the boundary points again from the remaining points according to the second set ratio, obtain the next layer of boundary points, and judge whether the next layer of boundary points contains local center points. Cluster, select the microcluster with the highest degree of association from its neighboring clusters to merge, and delete the boundary points of this layer until the boundary points of the set number of times are deleted. After the boundary points of the set number of times are deleted, for the local center point For the microclusters that still exist, if there are still data points in the adjacent clusters, the microclusters are merged with their adjacent clusters.

The present invention also provides a clustering device for image data, the clustering device includes a memory and a processor, and a computer program stored on the memory and running on the processor, the processor and the processor The memory is coupled, and the processor implements the image data clustering method of the present invention when executing the computer program.

The image clustering method of the present invention adopts a self-recommendation strategy to select local center points, which eliminates the subjectivity of manually selecting cluster centers, and solves the problem that the center points of clusters with smaller density are ignored; The micro-cluster center point is used to generate micro-clusters; then a micro-cluster merging method is proposed. In the process of deleting boundary points layer by layer, the micro-clusters with the highest degree of association are selected from the neighboring clusters of a micro-cluster and merged because such micro-clusters Closer to the cluster center, it is more likely to be in the same cluster as this microcluster, making the final clustering result 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 step 1) is determined according to the mutual adjacency of the point and the average value of the distance between the point and the midpoint of its neighborhood.

Further, the present invention provides a formula for calculating the density of data points, and the formula for calculating the density value is:

表示数据点x_p到数据点x_q的加权欧式距离，在数据点x_p的k个邻居中，如果x_q是数据点的x_p第ε个邻居，那么ε_pq的值就是ε，ε的值越大，x_q到x_p的距离就越远；若x_p是数据点的x_q第ε个邻居，ε_qp的值就是ε。where ρ(x _p ) is the density of the data point x _p , MND(x _p ) is the mutual adjacency of the data point x _p , NG(x _p ) is the neighborhood of the data point x _p , |NG(x _p )| represents the number of data points in the neighborhood of data point x _p ,

Represents the weighted Euclidean distance from the data point x _p to the data point x _q . Among the k neighbors of the data point x _p , if x _q is the ε-th neighbor of the data point x _p , then the value of ε _pq is ε, ε The larger the value, the further the distance from x _q to x _p ; if x _p is the ε-th neighbor of x _q of the data point, the value of ε _qp is ε.

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

Further, in order to accurately determine whether a data point is a boundary point, the concept of the boundary degree of a data point is proposed. The larger the value of the boundary degree, the more likely the data point becomes a boundary point. The calculation formula of the boundary degree of the data point is: :

为数据点X_p和数据点X_q之间的皮尔逊相关系数，m为数据集的维度。where bp(x _p ) represents the boundary degree of the data point x _p , |kNN(x _p )| represents the number of neighbors of the data point x _p , kNN(x _p ) represents the k neighbors of the data point x _p ,

is the Pearson correlation coefficient between data point X _p and data point X _q , and m is the dimension 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:

where NG(x _p ) represents the neighborhood of the data point x _p , c _i represents the i-th micro-cluster formed, and R( _{ci , c j )} represents the adjacent cluster formed between micro-cluster c _i and micro-cluster c _j relation.

Further, after the microcluster is formed, the method changes the local center point of the microcluster to the mean point of all data points in the microcluster. If such a mean point does not exist, it is changed to the data closest to the mean point in the microcluster. point.

Further, the present invention provides the calculation formula of the degree of binding, and the calculation formula adopted by the degree of binding is:

的值越大，c_j就越靠近簇核心区域，c_j和c_i的结合度就越高，

表示两个局部中心点之间的加权欧式距离。|c _j | represents the number of data points in the micro-cluster c _j in the initial state, |c _jt | represents the number of data points in the micro-cluster c _j after deleting the boundary points of the t layer,

The larger the value of , the closer c _j is to the cluster core area, the higher the degree of association between c _j and c _i ,

Represents the weighted Euclidean distance between two local center points.

Further, in order to prevent some outliers from clustering alone, the method also includes a process of adjusting the clusters after the micro-clusters are merged:

After merging homologous clusters, determine whether the number of data points in the obtained cluster is less than or equal to 0%-5% of the total number of data points, if so, the cluster is regarded as an abnormal cluster, and the data points in the cluster are assigned to the non-anomalous cluster whose closest data point is located.

Description of drawings

Fig. 1 is the flow chart of the image data clustering method of the present invention;

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

Fig. 2-b is a schematic diagram of preliminary clustering results when the Pathbased data set (k=11) in the embodiment of the present invention;

Figure 2-c is a schematic diagram of a preliminary clustering result when the Flame data set (k=7) in the embodiment of the present invention;

Fig. 2-d is a schematic diagram of preliminary clustering results when the Flame data set (k=18) in the embodiment of the present invention;

Fig. 2-e is a schematic diagram of preliminary clustering results when the Spiral data set (k=6) in the embodiment of the present invention;

Fig. 2-f is a schematic diagram of preliminary clustering results when the Spiral data set (k=13) in the embodiment of the present invention;

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

Figure 3-b is a schematic diagram of the final clustering result of the Flame dataset in the embodiment of the present invention;

Figure 3-c is a schematic diagram of the final clustering result of the Spiral dataset in the embodiment of the present invention;

4-a is a schematic diagram of the final clustering result of the Olivetti Faces dataset in the embodiment of the present invention;

Figure 4-b is a schematic diagram of the final clustering result of the Olivetti Faces dataset using the DPC algorithm;

FIG. 5 is a structural block diagram of the image data clustering apparatus of the present invention.

Detailed ways

The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

Method embodiment

Considering the problem that the DPC algorithm and other clustering algorithms have poor clustering effect when clustering image data, the present invention provides a new image clustering method, which is based on the existing DPC for image clustering. In the above, two improvements are proposed: one is to use weighted Euclidean distance and mutual adjacency to calculate the density of data points; the other is to use its own recommendation strategy to automatically determine local center points, and then form micro-clusters based on the determined local center points, and propose The strategy of merging the micro-clusters with the micro-clusters with fewer boundary points in the neighboring clusters, and get the final clustering result. The implementation process is shown in Figure 1. The implementation process of the method will be described in detail below with respect to specific examples.

1. Obtain the image data to be clustered and preprocess the image data.

The acquired image data generally has a large number of features. For example, a 90×150 image has 13,500 features. To this end, it is necessary to perform dimension reduction processing on the image data. In this embodiment, the PCA algorithm is used for the dimension reduction processing. For the Olivetti faces dataset, PCA is used to filter out features whose cumulative contribution rate exceeds 90%. In order to eliminate the influence of different features on the experimental results due to different value ranges, it is necessary to normalize the data. The formula used for normalization is:

where x _ij is the value of the i-th data point on the j-th attribute, and x _j ′ is the value of the i-th data point on the j-th attribute after normalization.

2. Determine the density of each data point.

For a given dataset x _n×m ={x ₁ , x ₂ ,..., _xi ,...,x _n }, n is the number of data points, m is the dimension of data points, x _i =[ x _i1 , x _i2 , ..., x _im ], the calculation formula of data point density in the existing DPC algorithm is:

Among them, formula (1) is aimed at a data set with a large number of samples, and formula (2) is aimed at a data set with a small number of samples. The meaning of the d _c value in formula (1) and formula (2) is the same, and both represent the truncation distance, and d _ij represents the Euclidean distance between two data points.

The present invention firstly uses the Pearson correlation coefficient between two data points as a weight to calculate the weighted Euclidean distance between the two data points; How closely the data points are connected. The higher the value of the degree of mutuality, the closer a data point is to the data points in its neighborhood, and the greater the density of this data point. Finally, the weighted Euclidean distance and the mutual adjacency are combined to calculate the density of the data points.

1) Calculate the weighted Euclidean distance between two data points.

In this example, the Pearson correlation coefficient is used as the influencing factor to define a new calculation method for the distance between data points, as shown below:

表示数据点x_p和数据点x_q之间的加权欧式距离，

表示数据点x_p和数据点x_q之间的皮尔逊相关系数，x_pj表示数据点x_p在其第j个属性上的取值。x_qj表示数据点x_q在其第j个属性上的取值。m为数据集的维度，即数据点特征的数量。in

represents the weighted Euclidean distance between data point x _p and data point x _q ,

represents the Pearson correlation coefficient between the data point x _p and the data point x _q , and x _pj represents the value of the data point x _p on its jth attribute. x _qj represents the value of the data point x _q on its jth attribute. m is the dimension of the dataset, that is, the number of data point features.

2) Calculate the mutual adjacency of data points and the density of data points.

Before determining the degree of mutual adjacency, the mutual adjacency relationship is first described. As we all know, if there is x _q in the k-nearest neighbor of x _p , and there is also x _p in the k-nearest neighbor of x _q , then x _p and x _q form a mutual neighbor relationship. Based on this, the concept of data point neighborhood is proposed. Taking a data point x _p as an example, the neighborhood of x _p refers to a set of data points that form a mutually adjacent relationship with x _p in the k-nearest neighbors of x _p . That is, NG(x _p )={x _q |x _q ∈kNN(x _p )∧x _p ∈kNN(x _q )}, kNN(x _p ) represents the set of the first k data points closest to the data point x _p , NG(x _p ) represents the set of data points in the neighborhood of x _p . Here, the present invention proposes the concept of mutual adjacency that can reflect the density of data points. The calculation method of mutual adjacency is:

The larger the value of MND(x _p ), the greater the density of x _p .

The density of data points is inversely proportional to the average distance between data points, and proportional to the degree of mutual adjacency of the data points themselves. Then, a new density calculation method is proposed by combining the weighted Euclidean distance and the mutual adjacency. The specific formula used is:

where |NG(X _p )| represents the number of data points in the neighborhood of x _p .

3. Use the self-recommendation strategy to automatically select local center points to form micro-clusters.

For datasets with large density differences between clusters, DPC easily ignores the center points of clusters with low density. To this end, the present invention proposes a strategy of automatically selecting local center points. If the density of x _p is greater than the density of points in its neighborhood, x _p is recommended as a local center point (referred to as a self-recommendation strategy), otherwise x _p does not recommend a local center point. If there are c local center points in total, LC represents the set of all local center points in a data set. Then, LC={lc ₁ , lc ₂ , ..., lc _i , ..., lc _j , ..., lc _c }.

The self-recommendation strategy is adopted, so that the points in the clusters with relatively small density have a great chance to be selected as the local center points, so that the clusters with relatively small density will not be ignored.

In this embodiment, the DPC method is used to allocate the remaining data points, that is, the remaining data points are allocated to the cluster where the density is higher than that of the cluster where the nearest local center point is located. After the allocation process, multiple "microclusters" are formed. So far, the preliminary clustering task is completed. C ₌ { _{c 1} , _{c 2} , . There are a total of c microclusters), and ci represents the _ith microcluster. Since the dimension of image data is relatively high after dimension reduction, this embodiment only shows preliminary clustering results on three widely used synthetic datasets. The datasets used include Pathbased dataset, Flame dataset and Spiral dataset. The preliminary clustering results are shown in Figures 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 microclusters, but if two Two non-adjacent microclusters are of the same shape, which also represent different microclusters; black five-pointed stars represent the local center points of different microclusters; black numbers represent the cluster numbers of different microclusters). As another implementation manner, in addition to the DPC method, other existing clustering methods can also be used to assign the remaining data points to the cluster where the nearest local center point is located, so as to form micro-clusters.

As can be seen from the figure, for each data set, 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, it is easy to find a suitable k value to get the correct initial Clustering results. The correct preliminary clustering result means that for the data points in each microcluster, in the real situation, these 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 allocated, the local center point is changed to the mean point of all data points in the microcluster. If it exists, the local center point is changed to the data point closest to the mean point in the micro-cluster. The reason for this is that the center point is moved to the center of the micro-cluster in order to avoid being deleted as a boundary point.

4. Merge microclusters.

In the present invention, one data set is composed of multiple clusters, and one cluster is composed of multiple micro-clusters. Therefore, it is necessary to merge micro-clusters into clusters. Before merging micro-clusters, the adjacent cluster relationship is introduced.

If the union of the data points in the neighborhood of all data points in the microcluster c _i has an intersection with the members of the microcluster c _j , it is considered that c _j and c _i constitute an adjacent cluster relationship. The neighbor relationship between c _j and c _i is _represented by R(ci , c _j ). but

If two microclusters satisfy the neighbor relationship, they are called pairwise clusters. Two microclusters are homologous clusters if they are pairwise clusters and also belong to the same cluster. Two microclusters are non-homologous if they are pairwise but not in the same cluster. If two microclusters satisfy the neighbor relationship, the two microclusters are neighbors to each other. Let S _ni represent the set of all neighboring clusters of microcluster c _i .

In the process of merging microclusters, it is necessary to delete boundary points layer by layer, so the method for determining the boundary points is described here. The boundary point is determined based on the boundary degree of the data point. The greater the boundary degree, the more likely the point is a boundary point. For any data point x _p , the greater the distance between x _p and its k-nearest neighbors and the greater the distance between x _p 's neighbors and its neighbors, the greater the boundary degree of this point. In addition, taking the weighted Euclidean distance as the distance between two data points, the calculation formula of the boundary degree of the data points is:

in,

Take a percentage of the total number of data points as the number of boundary points. The corresponding number of points with larger boundary degree are selected as boundary points.

The specific method of merging microclusters is as follows. Taking _ci as an example, it can form an adjacent cluster relationship with multiple microclusters, that is to say, _ci has multiple adjacent clusters. Among these adjacent clusters, some form homologous clusters with _ci , and some form non-homologous clusters with _ci . The main purpose is to find a _microcluster that can form a homologous cluster with _ci from the neighboring clusters of ci, and then merge _ci and this microcluster. In fact, in the region far from the cluster core ₍ represented by X1), it is easier to form non-homologous clusters between microclusters. In the region near the cluster core (indicated by X ₂ ), it is easy to form homologous clusters between microclusters. So in the two places X ₁ and X ₂ , the homologous clusters are looked up in different ways. Assuming that X1 consists _of multiple layers of boundary points, the three steps of microcluster merging are as follows.

1) After deleting a layer of boundary points, it is determined whether a neighboring cluster relationship is formed between each micro-cluster.

For many datasets, the clusters are closely related to each other. Then, the micro-clusters on the cluster boundary are very easy to form neighbor-cluster relationships between them. Once the neighbor relationship is formed between two microclusters in different clusters, it is possible for the two microclusters to further form homologous clusters. Therefore, after deleting a layer of boundary points, determine the adjacent cluster relationship (according to the set ratio, the data points with higher boundary degrees are used as boundary points, and the set ratio is generally between 10% and 50% of the total number of data points. ), so that in the boundary area of the cluster, the adjacent cluster relationship that was established between the two microclusters is no longer established, and naturally, they cannot form a homologous cluster.

2) Find homologous clusters.

的值越大，就说明c_j越靠近簇中心，就说明c_j和c_i越有可能构成同源簇。此外，两个微簇的局部中心点的距离越近，这两个微簇就越有可能处于同一个簇。以结合度表示两个微簇之间是同源簇的可能性，结合度的值越大，两个微簇就越有可能是同源簇。结合度以如下方式计算Find _homologous clusters in X1. The purpose of this stage is to find homologous clusters _of all microclusters in X1. Assuming that a certain layer of X ₁ has M boundary points (X ₁ is composed of layer-by-layer boundary points), there are some local center points in the M boundary points, and l _cj represents one of the local center points. C _i represents the micro-cluster corresponding to l _cj , and c _j represents a micro-cluster in the neighboring clusters of c _i . Then, before the M boundary points are deleted, the present invention needs to determine who the homologous cluster of _ci is. As mentioned earlier, the closer the microclusters are to the center of the cluster, the more likely they are to form homologous clusters. The microclusters closer to the center of the cluster contain fewer boundary points. Therefore, |c _j | represents the number of data points in the microcluster c _j in the initial state. Let |c _jt | denote the number of data points in the microcluster c _j before the M boundary points are deleted (where t represents the boundary points that have been deleted t times before). So,

The larger the value is, the closer c _j is to the center of the cluster, and the more likely c _j and c _i are to form a homologous cluster. In addition, the closer the local center points of two microclusters are, the more likely the two microclusters are in the same cluster. The degree of binding is used to express the possibility that two microclusters are homologous clusters. The greater the value of the degree of binding, the more likely the two microclusters are homologous clusters. The degree of binding is calculated as follows

表示两个局部中心点之间的加权欧式距离。

Represents the weighted Euclidean distance between two local center points.

In the neighboring clusters of c _i , if c _j and c _i have the highest degree of binding, then c _j is the homologous cluster of c _i . After finding the homologous clusters of all the local center points contained in the M boundary points, delete the M boundary points. If there is no local center point among the M boundary points, the M boundary points are directly deleted, and then the next layer of boundary points is processed, and the above process is not stopped until the specified number of boundary points are deleted. If _ci has no adjacent clusters, it is considered that _ci and _ci constitute homologous clusters.

Find _homologous clusters in X2. _The boundary points deleted above are the data points in X1. _The rest of the points in the dataset are the points in X2. _Let l _ci represent a local center point in X ₂ , and let ci represent the micro-cluster corresponding to this local center point. As mentioned earlier, _{microclusters} in X2 can easily form homologous clusters. Therefore, for all the neighboring clusters of c _i , as long as there are data points in it, it is considered to be a homologous cluster with c _i . It is enough to find the homologous cluster of the _microcluster where all the local center points in X2 are located.

Then, for all homologous clusters in X1 and X2, merge the _two microclusters in the homologous cluster into _one cluster.

3) Reassign data points in anomalous clusters

After the micro-clusters in steps 1) and 2) are merged, 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, it is considered to be a non-abnormal cluster , the setting ratio is 0%-5%, and 1% is selected in this embodiment. Anomalous clusters appear either because those points are themselves outliers, or because those points are relatively less closely related to the core part of the cluster. Whatever the reason, reclassify the data points in the outlier clusters. That is, for each data point in the abnormal cluster, assign it to the cluster where the closest data point is located (this cluster must be a non-abnormal cluster, otherwise, continue to find the cluster where the next data point closest to it is located). At this point, the entire micro-cluster merging process ends, and the final image clustering result is obtained.

The algorithm adopted by the present invention in the image clustering process is named: the direction of the union between microclusters is determined by the passing of boundary points (DUMDPBP) by the passage of boundary points. The steps can be summarized as follows:

Device embodiment

The apparatus proposed in this embodiment, as shown in FIG. 5 , includes a processor and a memory. The memory stores a computer program that can be executed on the processor, and the processor implements the methods of the above method embodiments when executing the computer program.

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

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. Usually, the information is digitized and then stored in an electrical, magnetic, or optical medium. For example: all kinds of memories that use electrical energy to store information, RAM, ROM, etc.; all kinds of memories that use magnetic energy to store information, hard disks, floppy disks, magnetic tapes, magnetic core memories, magnetic bubble memories, U disks; use optical methods to store information of all kinds of memory, CD or DVD. Of course, there are other ways of memory, such as quantum memory, graphene memory, and so on.

The device constituted by the above-mentioned memory, processor and computer program is realized by the processor executing corresponding program instructions in the computer, and the processor can be equipped with various operating systems, such as windows operating system, linux system, android, iOS system, etc.

As other implementation manners, the apparatus may further include a display, which is used for displaying the diagnosis results for the reference of the staff.

experiment analysis

In order to better illustrate the effect of the present invention, the method adopted by the present invention is compared with several existing clustering methods below. (TM)i7-8700 CPU@3.20GHz processor. Three different types of 2-dimensional datasets and a high-dimensional image dataset were used in the experiments, as shown in Table 1, which describes the basic information of the datasets. Before the experiment starts, the data is normalized to eliminate the influence of the different value ranges between different features on the experimental results.

Table 1

For the first three data sets, the clustering results using the DUMDPBP algorithm in the present invention are shown in Figure 3-a, Figure 3-b and Figure 3-c. The clustering results are consistent with the actual classification results of the corresponding data sets. It shows that the DUMDPBP algorithm adopted in the present invention has a good effect on the Pathbased data set, the Flame data set and the Spiral data set. The different shapes in the above figure represent different clusters. For the Pathbased dataset, they are finally clustered into three categories, for the Flame dataset, they are finally clustered into two categories, and for the Spiral dataset, they are finally clustered into three categories.

For the Olivetti faces dataset, because it is composed of a large number of photos, rather than points in a two-dimensional plane, and the number of clusters is as many as dozens, it cannot be displayed in different shapes like a synthetic dataset. its clustering results. There is no other way to illustrate a dataset with a large number of clusters and consisting of photos. However, in subsequent chapters, the evaluation indicators will be used to show the clustering results. Even so, we show the case of cluster center points in the clustering results. It should be noted that the Olivetti faces dataset has 400 photos, and only the photos of the first 100 people are selected for display here. Figure 4-a is the clustering result of DUMDPBP algorithm. The white dots in the upper right corner of some pictures represent the selected local center points. It can be seen from the figure that at least one white dot appears in each row (that is, each cluster), that is to say, each cluster is selected. the local center points so that each cluster is not ignored. Figure 4-b is the clustering result of the DPC algorithm. The white dot in the upper right corner of the picture represents the cluster center point selected by the algorithm. It can be seen from the figure that the first cluster (the photo in the first row), the eighth cluster Clusters (photos in row 8) and 10 clusters (photos in row 10) did not select the cluster center point, which means that the DPC algorithm will definitely compare the data points in these three clusters with other Data points in clusters are confused. In fact, the DPC algorithm assigns all the data points in the 3rd cluster and 8 data points in the 10th cluster into one cluster, which is a bad situation. By comparison, it can be found that the present invention has great advantages over the DPC algorithm.

In order to further know whether the clustering result of the DUMDPBP algorithm is good or bad, the present invention adopts three evaluation indexes commonly used by everyone to measure the performance of the clustering algorithm. Three evaluation indicators are: adjusted mutual information AMI, adjusted Rand coefficient ARI and Fowlkes-Mallows index FMI. The value range of AMI is [-1, 1], the value range of ARI is [-1, 1], and the value range of FMI is [0, 1]. In these three indicators, the larger the index value, the more consistent the clustering result is with the actual situation. _{Let U = { U 1} , _. . Let H(U) represent the entropy of the original partition, H(V) represent the entropy of the clustering result partition, MI(U, V) represent the mutual information between the original partition U and the clustering result partition V, and E{·} represents The expected mutual information between the original partition U and the clustering result partition V. Use a to indicate that the two data points are in the same cluster in reality, and the clustering result is also in the same cluster. Let b represent the fact that two data points are in the same cluster in reality, but the clustering result is not in the same cluster. Use c to indicate that the two data points are not in the same cluster in reality, and the clustering result is in the same cluster. Denote by d that the two data points are not in the same cluster in reality, nor are they in the same cluster in the clustering result. Then the above three evaluation indicators are calculated as follows.

According to the above three indicators, the algorithm mentioned in the present invention (referred to as DUMDPBP) is compared with the existing seven clustering algorithms on the above four data sets, and the comparison results are shown in Table 2.

Table 2

table 3

In Table 2, the parameters of DUMDPBP on the 4 datasets are shown in Table 3, including the number of neighbors of the data point (k), the number of times to find homologous clusters (N1), and the number of data points that border points occupy when determining the relationship between neighbor clusters. The proportion of the total amount (N2), the proportion of boundary points to the total amount of remaining data points when searching for homologous clusters (N3). For the Spiral dataset, since there are obvious gaps between clusters, as long as two microclusters form an adjacent cluster relationship, they are homologous clusters. Therefore, there is no need to search for homologous clusters, and N ₁ is set to 0; for the Flame data set, since the interval between clusters is not obvious, N ₂ is set relatively large; for the value of N ₃ , it is generally less than 0.5. The adjustment parameter dc = _1.1125 is adopted for the clustering result of the DPC algorithm on the Olivetti faces dataset. The adjustment of the remaining parameters all refer to the paper published by the author Rui Liu et al in the journal "Information Sciences": Shared-nearest-neighbor-based clustering by fast search and find of densitypeaks.

It can be seen from Table 2 that for a dataset of arbitrary shape such as Spiral, the values of DUMDPBP and the other five clustering algorithms are all 1 on the three indicators, indicating that the clustering results of these six clustering algorithms are consistent with the real situation. Exactly the same; for a dataset like Flame, where there is no obvious separation boundary between clusters, DUMDPBP, FKNN-DPC and DPC all achieved the upper limit of 1 in the three indicators, indicating that the clustering results are completely consistent with the real situation; For the Pathbased data set, only the DUMDPBP algorithm has a value of 1 on the three indicators, while the effect of other algorithms is relatively poor; for the Olivetti faces data set, compared with other algorithms, the DUMDPBP algorithm is in the three indicators. outperformed other algorithms. Therefore, the algorithm of the present invention has good clustering effect not only on synthetic datasets, but also on image datasets in the real world. Therefore, the DUMDPBP algorithm proposed by the present invention is suitable for various types of data sets, and the application prospect is very good.

The invention combines mutual adjacency and weighted Euclidean distance to propose a new density calculation method to avoid artificially setting cutoff distances; next, a strategy of self-recommending local center points is proposed, and then remaining points are allocated, so that clusters with relatively small density can also be selected Local center point, so that clusters with relatively small density will not be ignored; finally, in the process of deleting boundary points layer by layer, look for homologous clusters, merge homologous clusters into clusters, and complete the clustering task. And through experiments, it is found that the algorithm proposed in the present invention has achieved good results on data sets of different distribution types, and has a wide range of applications.

Claims (10)

1. A clustering method of image data, characterized by 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.

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:

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.

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),

is a data point x_pAnd data point x_qPearson correlation coefficient between, m is the dimensionality of the data set.

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:

|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.

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:

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.

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.

Images