A Hiring-Based Density Clustering Method and System

Abstract

This invention provides a hiring-based density clustering method and system. The method includes: acquiring several facial features and obtaining the feature density corresponding to the facial features; selecting cluster root node features from the several facial features based on the feature density; constructing several sequentially connected supervisory levels based on the cluster root node features, with each supervisory level including several candidate node features; filtering the candidate node features based on their association relationships to identify whether the candidate node features are to be retained, thereby obtaining facial clusters. By acquiring the feature density and selecting the starting point of clustering, i.e., the cluster root node features, based on the feature density, the dependence on the K value in traditional clustering algorithms is avoided. Clusters of arbitrary shapes can be autonomously discovered using only the cluster root node features. By filtering the candidate node features based on their association relationships, the overgeneralization and mis-clustering that may occur when the similarity transmission chain is too long is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a hired density clustering method and system. BACKGROUND

[0002] Clustering is an unsupervised learning algorithm, which aims to divide an unlabeled dataset into several groups (called "clusters") so that the data objects in the same cluster have high similarity and the data objects between different clusters have significant difference. Clustering plays an important role in image feature grouping, user behavior analysis and other applications, and can mine the potential rules of data to provide structured interpretation for downstream business.

[0003] Existing clustering algorithms can be divided into two categories: neural network-based clustering algorithms (such as DeepCluster, DEC, etc.) and traditional clustering algorithms (such as K-means, spectral clustering, DBSCAN, etc.). The former performs well in high-dimensional feature extraction, but has limitations such as dependence on a large amount of resources, lack of interpretability in the clustering process, and overfitting on small-scale data. In contrast, traditional clustering algorithms have the advantages of computational efficiency, logical transparency, and no training dependence, and are still irreplaceable in real-time systems, edge computing and other resource-constrained scenarios.

[0004] Traditional clustering algorithm systems such as K-means and spectral clustering require the number of clusters K to be preset, but in actual applications, the data distribution is often unknown, such as the number of faces in the face picture clustering division process. Blindly specifying K values will lead to distortion of the clustering structure, i.e., there are different face pictures corresponding to different faces in the same feature set during face picture division. SUMMARY

[0005] In view of the shortcomings of the prior art, the purpose of the present application is to provide a hired density clustering method and system, which aims to solve the technical problem that the traditional clustering algorithm system needs to preset the number of clusters K, but the uncertainty of data distribution leads to distortion of the clustering structure when K is blindly specified, i.e., there are different face pictures corresponding to different faces in the same feature set during face picture division.

[0006] In order to achieve the above purpose, in a first aspect, the present application provides a hired density clustering method, comprising the following steps:

[0007] extracting a plurality of face features corresponding to the face pictures from a plurality of face pictures, and obtaining feature densities corresponding to the face features;

[0008] Based on the feature density, a cluster root node feature is selected from the plurality of face features, and a plurality of layers of a supervisor hierarchy is constructed based on the cluster root node feature, the cluster root node feature and the plurality of layers of the supervisor hierarchy forming a tree-shaped supervisor structure, and the supervisor hierarchy includes a plurality of candidate node features.

[0009] The candidate node features are subjected to association relationship screening to identify whether the candidate node features are reserved node features, and if the candidate node features are not reserved node features, the candidate node features are removed, and the tree-shaped supervisor structure is updated to a face cluster.

[0010] Further, the step of obtaining the feature density corresponding to the face feature comprises:

[0011] One of the face features is selected as a reference feature, and the remaining face features are selected as reference features.

[0012] The feature similarity between the reference feature and the reference feature is obtained.

[0013] The feature similarity is compared with a first similarity threshold to select a plurality of preselected features corresponding to the reference feature from a plurality of reference features.

[0014] Based on the reference feature and the number of preselected features, the feature density corresponding to the reference feature is determined.

[0015] Further, the calculation formula of the feature similarity is:

[0016] ,

[0017] wherein, represents the feature similarity between the reference feature and the i-th reference feature, represents the reference feature, represents the i-th reference feature, represents the dot product operation, represents the vector length.

[0018] Further, the number of supervisor layers is 1-4 layers.

[0019] Further, the step of selecting a cluster root node feature from a plurality of face features based on the feature density comprises:

[0020] Based on the feature density, a plurality of face features are sorted to select a face feature with the largest feature density as a cluster root node feature.

[0021] Further, the step of constructing a plurality of layers of the main supervisor hierarchy in sequence based on the cluster root node features comprises:

[0022] Based on the feature similarity, the remaining face features are divided into a plurality of candidate node features and a plurality of alternative node features, the plurality of candidate node features are associated with the cluster root node features to complete single clustering, and the plurality of candidate nodes constitute the main supervisor hierarchy.

[0023] The candidate node features are subjected to a plurality of single clustering to form a plurality of main supervisor hierarchies.

[0024] Further, the step of screening the association relationship of the candidate node features to identify whether the candidate node features are reserved node features comprises:

[0025] When the number of the main supervisor hierarchies is less than 4 layers, a first main similarity between the candidate node features and the cluster root node features is obtained;

[0026] From the upper level main supervisor hierarchy of the candidate node features, other candidate node features having an association relationship with the candidate node features are selected as first comparison features, and a first time similarity between the candidate node features and the first comparison features is obtained;

[0027] Based on the first main similarity and the first time similarity, a first final similarity is determined, the first final similarity is compared with a second similarity threshold, and if the first final similarity is greater than the second similarity threshold, the candidate node features are determined as reserved node features.

[0028] Further, the step of screening the association relationship of the candidate node features to identify whether the candidate node features are reserved node features further comprises:

[0029] When the number of the main supervisor hierarchies is 4 layers, and the candidate node features are located in the main supervisor hierarchy of the 4th layer, a second main similarity between the candidate node features and the cluster root node features is obtained;

[0030] From the upper level main supervisor hierarchy of the candidate node features, other candidate node features having an association relationship with the candidate node features are selected as second comparison features, and a second time similarity between the candidate node features and the second comparison features is obtained;

[0031] determining a second final similarity based on the second primary similarity and the second secondary similarity, comparing the second primary similarity and the second final similarity with a third similarity threshold respectively, if the second final similarity is greater than the third similarity threshold and the second primary similarity is greater than the third similarity threshold, determining the candidate node feature as a reserved node feature.

[0032] Further, the formula for obtaining the first final similarity is:

[0033]

[0034] wherein, the first final similarity is represented by S1, the first primary similarity is represented by S1p, the first primary weight is represented by w1p, the jth first secondary similarity is represented by S1j, the jth first secondary weight is represented by w1j, the number of first secondary similarities is represented by N1.

[0035] In a second aspect, an embodiment of the present application provides a hired density clustering system, applied to the hired density clustering method as described in the first aspect above, and the system comprises:

[0036] a processing module, configured to extract a plurality of face features corresponding to a plurality of face pictures from the plurality of face pictures, and obtain feature densities corresponding to the face features;

[0037] a first screening module, configured to select a cluster root node feature from the plurality of face features based on the feature densities, and construct a plurality of layers of successively connected supervisor hierarchies based on the cluster root node feature, wherein the cluster root node feature and the plurality of layers of supervisor hierarchies form a tree-shaped supervisor structure, and the supervisor hierarchies comprise a plurality of candidate node features;

[0038] a second screening module, configured to perform association relationship screening on the candidate node features to identify whether the candidate node features are reserved node features, and if a candidate node feature is not a reserved node feature, the candidate node feature is removed, so as to update the tree-shaped supervisor structure into a face clustering.

[0039] In a third aspect, an embodiment of the present application provides a computer, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the hired density clustering method as described in the first aspect above when executing the computer program.

[0040] ​In a fourth aspect, the embodiments of the present application provide a storage medium having a computer program stored thereon, which, when executed by a processor, implements the employed density clustering method according to the first aspect.

[0041] Compared with the prior art, the present application has the beneficial effects that: by acquiring the feature density, the starting point of clustering, i.e., the cluster root node feature, is selected based on the feature density, the dependence on the K value in the traditional clustering algorithm system is avoided, and only the cluster root node feature is needed to autonomously discover clusters of any shape, i.e., to autonomously complete the classification of face clusters corresponding to different faces; meanwhile, after the construction of the tree-shaped supervisor structure is completed, by screening the association relationship of the candidate node features, the situation of over-generalization and misclustering caused by too long similarity transmission chain is avoided, i.e., the situation that the similarity between the candidate node features in the later supervisor level and the cluster root node features is too low, resulting in the error aggregation of face features of different faces is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 a flowchart of the employed density clustering method in the first embodiment of the present application;

[0043] Figure 2 a structural block diagram of the employed density clustering system in the second embodiment of the present application;

[0044] The following detailed description will further illustrate the present application in combination with the above-mentioned drawings. DETAILED DESCRIPTION

[0045] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the related drawings. Several embodiments of the present application are shown in the drawings. However, the present application can be realized in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present application more thorough and comprehensive.

[0046] It should be noted that when an element is referred to as being "fixedly attached" to another element, it can be directly on the other element or there can be an intervening element. When an element is referred to as being "connected" to another element, it can be directly connected to the other element or intervening elements can be present. The terms "vertical", "horizontal", "left", "right", and the like as used herein are for purposes of illustration only.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application. The use herein of the terms "and / or" includes a set of one or more associated listed items.

[0048] Referring to Figure 1 The first embodiment of the application provides a hired density clustering method, comprising the following steps:

[0049] S10: extracting a plurality of face features corresponding to a plurality of face pictures from the plurality of face pictures, and obtaining feature densities corresponding to the face features;

[0050] In this embodiment, the face features are obtained by a multi-layer convolutional neural network. It can be understood that the face pictures and the face features are in one-to-one correspondence.

[0051] The step S10 comprises:

[0052] S110: selecting one of the face features as a reference feature, and selecting the remaining face features as reference features;

[0053] The remaining face features are the face features other than the reference feature.

[0054] S120: obtaining a feature similarity between the reference feature and the reference features;

[0055] The calculation formula of the feature similarity is:

[0056] ,

[0057] wherein, denotes the feature similarity between the reference feature and the i-th reference feature, denotes the reference feature, denotes the i-th reference feature, denotes the dot product operation, denotes the vector length. In some embodiments, the feature similarity can also be formed by obtaining the Euclidean distance between the reference feature and the reference feature.

[0058] S130: comparing the feature similarity with a first similarity threshold to select a plurality of preselected features corresponding to the reference feature from the plurality of reference features;

[0059] In this embodiment, the first similarity threshold is 0.8.

[0060] S140: determining a feature density corresponding to the reference feature based on the reference feature and the number of the preselected features;

[0061] Suppose there are 10 facial features, when feature 1 is the reference feature, the feature similarity between feature 4, feature 5 and feature 1 is greater than 0.8, and the feature density of feature 1 is 3.

[0062] S20: selecting a cluster root node feature from the plurality of facial features based on the feature density, and constructing a plurality of layers of successively connected supervisor levels based on the cluster root node feature, wherein the cluster root node feature and the plurality of layers of supervisor levels form a tree-shaped supervisor structure, and the supervisor levels include a plurality of candidate node features;

[0063] The number of supervisor levels is 1-4 layers.

[0064] Specifically, the step S20 includes:

[0065] S210: sorting the plurality of facial features based on the feature density to select a facial feature with the maximum feature density as a cluster root node feature;

[0066] S220: separating the remaining facial features into a plurality of candidate node features and a plurality of alternative node features based on the feature similarity, and associating the plurality of candidate node features to the cluster root node feature to complete a single clustering, wherein the plurality of candidate nodes form a supervisor level;

[0067] Still taking 10 facial features as an example, when the feature density of feature 1 is 3, it is the largest, so feature 1 is selected as the cluster root node feature, at this time, feature 4 and feature 5 are the candidate node features, and the other facial features are the alternative node features, at this time, feature 1 is associated with feature 4 and feature 5, and feature 4 and feature 5 form the first layer of the supervisor level.

[0068] S230: performing the single clustering on the candidate node features for a plurality of times to form a plurality of supervisor levels;

[0069] If the feature similarity between feature 4 and feature 2 is greater than 0.8, and the feature similarity between feature 5 and feature 9 is greater than 0.8, then feature 2 and feature 9 are selected as the candidate node features of the next layer, and form the supervisor level of the next layer, and the foregoing operation is performed again on feature 2 and feature 9. It should be noted that if feature 2 and feature 9 do not exist other facial features with feature similarity greater than 0.8, then a tree-shaped supervisor structure of 2-layer supervisor level starting from the cluster root node feature is formed.

[0070] S30: performing association relationship screening on the candidate node features to identify whether the candidate node features are reserved node features, and if the candidate node features are not reserved node features, eliminating the candidate node features to update the tree-shaped supervisor structure into a face cluster;

[0071] Specifically, the step 30 comprises:

[0072] S310: when the number of the supervisor levels is less than 4 layers, obtaining a first main similarity between the candidate node feature and the cluster root node feature;

[0073] S320: from the superior supervisor level of the candidate node feature, selecting other candidate node features having an association relationship with the candidate node feature as first comparison features, and obtaining a first time similarity between the candidate node feature and the first comparison features;

[0074] S330: determining a first final similarity based on the first main similarity and the first time similarity, comparing the first final similarity with a first similarity threshold, and if the first final similarity is greater than the first similarity threshold, determining that the candidate node feature is a reserved node feature;

[0075] On the basis of the foregoing, if the feature similarity between feature 2 and feature 10 is greater than 0.8, a structure of feature 1-feature 4 (the supervisor level of the first layer)-feature 2 (the supervisor level of the second layer)-feature 10 (the supervisor level of the third layer) is formed, and it is necessary to determine whether feature 10 is a reserved node feature, then the first main similarity between feature 10 and feature 1 is obtained, and the first time similarity between feature 10 and feature 2 and feature 4 is obtained, and then the first final similarity is formed.

[0076] The formula for obtaining the first final similarity is:

[0077]

[0078] Among them, the first final similarity, the first main similarity, the first main weight, the jth first time similarity, the jth first time weight, the number of the first time similarities. It should be noted that the sum of the first main weight and all the first time weights is 1, and the sum of all the first main similarities and all the first time similarities is processed to obtain a total similarity, and the first main similarity and all the first time similarities are divided by the total similarity to obtain the first main weight and all the first time weights.​

[0079] S340: When the number of the supervisor levels is 4 levels, and the candidate node feature is located in the supervisor level of the 4th level, a second main similarity between the candidate node feature and the cluster root node feature is obtained;

[0080] S350: From the superior supervisor level of the candidate node feature, other candidate node features having an association relationship with the candidate node feature are selected as second comparison features, and a second time similarity between the candidate node feature and the second comparison features is obtained;

[0081] S360: A second final similarity is determined based on the second main similarity and the second time similarity, the second main similarity and the second final similarity are compared with a third similarity threshold respectively, if the second final similarity is greater than the third similarity threshold, and the second main similarity is greater than the third similarity threshold, the candidate node feature is determined as a reserved node feature;

[0082] It should be noted that the first main similarity, the second main similarity, the first time similarity, the second time similarity and the feature similarity are the same, which can be obtained directly in the foregoing steps. The second similarity threshold and the third similarity threshold are the same as the first similarity threshold. The second final similarity and the first final similarity are obtained in the same way, which will not be described here.

[0083] Based on the foregoing, in the clustering chain of feature 1-feature 4 (the supervisor level of the 1st level)-feature 2 (the supervisor level of the 2nd level)-feature 10 (the supervisor level of the 3rd level), if the feature similarity between feature 10 and feature 15 is greater than 0.8, feature 1-feature 4 (the supervisor level of the 1st level)-feature 2 (the supervisor level of the 2nd level)-feature 10 (the supervisor level of the 3rd level)-feature 15 (the supervisor level of the 4th level) is formed. When determining whether feature 15 is a reserved node feature, the second main similarity between feature 15 and feature 1 needs to be determined to avoid over-clustering caused by the difference between feature 15 and feature 1 being too large due to the clustering chain being too long. It should be noted that the candidate node features of other supervisor levels except the supervisor level of the 4th level (feature 4, feature 2, feature 10) can execute the determination rules of steps S310-S330. Understandably, after completing the face clustering with feature 1 as the cluster root node feature, if there are still other face features, the foregoing steps are repeated to form several face clusters. Different face clusters correspond to different faces, and the face features in the face cluster correspond to the picture features of the same face at different angles.

[0084] By acquiring the feature density, selecting a starting point of clustering, i.e., the cluster root node feature, based on the feature density, dependence on the K value in a traditional clustering algorithm system is avoided, and only by the cluster root node feature, clusters of any shape can be autonomously discovered, i.e., classification of face clustering corresponding to different faces is autonomously completed; meanwhile, after construction of the tree-shaped supervisor structure is completed, by screening of the association relationship of the candidate node features, when a similarity transmission chain is too long, over-generalization misclustering may be caused, i.e., similarity between the candidate node features in the later supervisor level and the cluster root node feature is too low, leading to erroneous aggregation of face features of different faces.

[0085] Please refer to Figure 2 The second embodiment of the present application provides an employed density clustering system, which is applied to the employed density clustering method in the above-mentioned embodiments, and details thereof have been described above. As used below, the terms "module", "unit", "sub-unit", etc. can be a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, hardware, or a combination of software and hardware can also be implemented and conceived.

[0086] The system comprises:

[0087] The processing module 10 is configured to extract a plurality of face features corresponding to a plurality of face pictures from the plurality of face pictures, and acquire feature densities corresponding to the face features;

[0088] The processing module 10 comprises:

[0089] The first unit is configured to select one of the face features as a reference feature, and select the remaining face features as reference features;

[0090] The second unit is configured to acquire feature similarity between the reference feature and the reference features;

[0091] The third unit is configured to compare the feature similarity with a first similarity threshold, so as to select a plurality of preselected features corresponding to the reference feature from the plurality of reference features;

[0092] The fourth unit is configured to determine a feature density corresponding to the reference feature based on the reference feature and a number of the preselected features;

[0093] The first screening module 20 is configured to select a cluster root node feature from the plurality of face features based on the feature density, and construct a plurality of layers of supervisor levels connected in sequence based on the cluster root node feature, wherein the cluster root node feature and the plurality of layers of supervisor levels form a tree-shaped supervisor structure, and the supervisor levels comprise a plurality of candidate node features.

[0094] The first screening module 20 comprises:

[0095] The fifth unit is configured to sort the plurality of facial features based on the feature density, so as to select a facial feature with the maximum feature density as a cluster root node feature;

[0096] The sixth unit is configured to separate the remaining facial features into a plurality of candidate node features and a plurality of alternative node features based on the feature similarity, and associate the plurality of candidate node features to the cluster root node feature, so as to complete a single clustering, and the plurality of candidate nodes constitute a supervisor level;

[0097] The seventh unit is configured to perform the single clustering on the candidate node features for a plurality of times, so as to form a plurality of supervisor levels;

[0098] The second screening module 30 is configured to screen the association relationship of the candidate node features, so as to identify whether the candidate node feature is a reserved node feature, and if the candidate node feature is not a reserved node feature, the candidate node feature is removed, so as to update the tree-shaped supervisor structure to a facial clustering;

[0099] The second screening module 30 comprises:

[0100] The eighth unit is configured to obtain a first main similarity between the candidate node feature and the cluster root node feature when the number of supervisor levels is less than 4 layers;

[0101] The ninth unit is configured to select, from a superior supervisor level of the candidate node feature, other candidate node features having an association relationship with the candidate node feature as first comparison features, and obtain a first similarity between the candidate node feature and the first comparison features;

[0102] The tenth unit is configured to determine a first final similarity based on the first main similarity and the first similarity, compare the first final similarity with a second similarity threshold, and if the first final similarity is greater than the second similarity threshold, determine that the candidate node feature is a reserved node feature;

[0103] The eleventh unit is configured to obtain a second main similarity between the candidate node feature and the cluster root node feature when the number of supervisor levels is 4 layers and the candidate node feature is located in a fourth supervisor level;

[0104] The twelfth unit is configured to select, from a superior supervisor level of the candidate node feature, other candidate node features having an association relationship with the candidate node feature as second comparison features, and obtain a second similarity between the candidate node feature and the second comparison features.

[0105] The thirteenth unit is configured to determine a second final similarity based on the second primary similarity and the second secondary similarity, compare the second primary similarity and the second final similarity with a third similarity threshold respectively, and determine the candidate node feature as a reserved node feature if the second final similarity is greater than the third similarity threshold and the second primary similarity is greater than the third similarity threshold.

[0106] The application further provides a computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the employment density clustering method when executing the computer program.

[0107] The application further provides a storage medium having a computer program stored thereon, and the computer program is executable on a processor to implement the employment density clustering method.

[0108] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily mean the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0109] The above embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the scope of the patent of the present application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the scope of protection of the present application. Therefore, the scope of protection of the patent of the present application should be subject to the appended claims.

Claims

1. An employed density clustering method, characterized by, The method comprises the following steps: extracting a plurality of face features corresponding to a plurality of face pictures from the plurality of face pictures, and obtaining a feature density corresponding to the face features; the step of obtaining the feature density corresponding to the face features comprises: selecting one of the face features as a reference feature and selecting the remaining face features as reference features; obtaining a feature similarity between the reference feature and the reference features; comparing the feature similarity with a first similarity threshold to select a plurality of preselected features corresponding to the reference feature from a plurality of reference features; determining a feature density corresponding to the reference feature based on the reference feature and the number of preselected features; based on the feature density, selecting a cluster root node feature from a plurality of face features, and constructing a plurality of layers of sequentially connected supervisor levels based on the cluster root node feature, wherein the number of supervisor levels is 1-4 layers, the cluster root node feature and a plurality of layers of supervisor levels form a tree-shaped supervisor structure, and the supervisor level comprises a plurality of candidate node features; screening the association relationship of the candidate node features to identify whether the candidate node features are reserved node features, and if the candidate node features are not reserved node features, eliminating the candidate node features to update the tree-shaped supervisor structure to a face clustering; the step of screening the association relationship of the candidate node features to identify whether the candidate node features are reserved node features comprises: when the number of supervisor levels is less than 4 layers, obtaining a first main similarity between the candidate node feature and the cluster root node feature; from the upper supervisor level of the candidate node feature, selecting other candidate node features having an association relationship with the candidate node feature as first comparison features, and obtaining a first similarity between the candidate node feature and the first comparison features; determining a first final similarity based on the first main similarity and the first similarity, comparing the first final similarity with a second similarity threshold, and if the first final similarity is greater than the second similarity threshold, determining the candidate node feature as a reserved node feature.

2. The employed density clustering method of claim 1, wherein, The calculation formula of the feature similarity is: ， wherein, denotes a feature similarity between the reference feature and the i-th reference feature, denotes the reference feature, denotes the i-th reference feature, denotes a dot product operation, denotes a vector length.

3. The employed density clustering method of claim 1, wherein, the step of selecting a cluster root node feature from a plurality of face features based on the feature density comprises: sorting a plurality of face features based on the feature density to select a face feature with the maximum feature density as a cluster root node feature.

4. The engaged density clustering method of claim 1, wherein, the step of constructing a plurality of layers of sequentially connected supervisor levels based on the cluster root node feature comprises: based on the feature similarity, separating the remaining face features into a plurality of candidate node features and a plurality of alternative node features, associating a plurality of candidate node features to the cluster root node feature to complete a single clustering, and a plurality of candidate node features form a supervisor level; performing a plurality of single clusterings on the candidate node features to form a plurality of supervisor levels.

5. The engaged density clustering method of claim 1, wherein, the step of screening the association relationship of the candidate node features to identify whether the candidate node features are reserved node features further comprises: When the number of the supervisor levels is 4 layers, and the candidate node feature is located in the supervisor level of the 4th layer, a second main similarity between the candidate node feature and the cluster root node feature is obtained; From the superior supervisor level of the candidate node feature, other candidate node features having an association relationship with the candidate node feature are selected as second comparison features, and a second time similarity between the candidate node feature and the second comparison features is obtained; A second final similarity is determined based on the second main similarity and the second time similarity, the second main similarity and the second final similarity are compared with a third similarity threshold respectively, if the second final similarity is greater than the third similarity threshold, and the second main similarity is greater than the third similarity threshold, the candidate node feature is determined as a reserved node feature.

6. The engaged density clustering method of claim 1, wherein, The formula for obtaining the first final similarity is: ， wherein, denotes the first final similarity, denotes the first main similarity, denotes the first main weight, denotes the j-th first similarity, denotes the j-th first weight, denotes the number of first similarities.

7. A hired density clustering system for use in the hired density clustering method according to any one of claims 1 to 6, characterized by The system comprises: A processing module is configured to extract a plurality of face features corresponding to a plurality of face pictures from the plurality of face pictures, and obtain a feature density corresponding to the face features; A first screening module is configured to select a cluster root node feature from the plurality of face features based on the feature density, and construct a plurality of layers of supervisor levels connected in sequence based on the cluster root node feature, wherein the cluster root node feature and the plurality of layers of supervisor levels form a tree-shaped supervisor structure, and the supervisor levels comprise a plurality of candidate node features; A second screening module is configured to perform association relationship screening on the candidate node features to identify whether the candidate node features are reserved node features, and if the candidate node features are not reserved node features, the candidate node features are removed, and the tree-shaped supervisor structure is updated to a face clustering.

Images