CN110377774A - Carry out method, apparatus, server and the storage medium of personage's cluster - Google Patents

Carry out method, apparatus, server and the storage medium of personage's cluster Download PDF

Info

Publication number
CN110377774A
CN110377774A CN201910636600.XA CN201910636600A CN110377774A CN 110377774 A CN110377774 A CN 110377774A CN 201910636600 A CN201910636600 A CN 201910636600A CN 110377774 A CN110377774 A CN 110377774A
Authority
CN
China
Prior art keywords
neighbor
image
person
personal
images
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910636600.XA
Other languages
Chinese (zh)
Other versions
CN110377774B (en
Inventor
程雅慧
王吉
陈志博
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tencent Technology Shenzhen Co Ltd
Original Assignee
Tencent Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tencent Technology Shenzhen Co Ltd filed Critical Tencent Technology Shenzhen Co Ltd
Priority to CN201910636600.XA priority Critical patent/CN110377774B/en
Publication of CN110377774A publication Critical patent/CN110377774A/en
Application granted granted Critical
Publication of CN110377774B publication Critical patent/CN110377774B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/55—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The disclosure is directed to a kind of method, apparatus, server and storage mediums for carrying out personage's cluster, belong to technical field of image processing.The described method includes: obtaining target person image；In preset character image database, the highest first preset number level-one neighbours' character image of similarity between target person image, level-one neighbours' character image undetermined as target person image are determined；For each level-one neighbours character image, in character image database, the highest second preset number second level neighbours' character image of similarity between determining and level-one neighbours' character image, if any secondary neighbours' character image in second preset number second level neighbours' character image is target person image, level-one neighbours' character image is added in the corresponding fixed neighbours' character image set of target person image；Based on neighbours' character image set, personage's clustering processing is carried out.Using the disclosure, the accuracy of personage's clustering processing can be improved.

Description

Method, device, server and storage medium for clustering people
Technical Field
The present disclosure relates to the field of image processing technologies, and in particular, to a method, an apparatus, a server, and a storage medium for performing person clustering.
Background
In the related art, the distance between the first person image and the second person image may be calculated by a person clustering algorithm, and a smaller distance indicates that the actual person corresponding to the first person image is more likely to be the actual person corresponding to the second person image. And if the calculated distance between the first person image and the second person image is smaller than a preset distance threshold value, determining that the actual person corresponding to the first person image is the actual person corresponding to the second person image, and determining the first person image as another person image of the actual person corresponding to the second person image, thereby completing the person clustering process.
In the process of calculating the distance between the first person image and the second person image, firstly, a first neighbor person image set corresponding to the first person image and a second neighbor person image set corresponding to the second person image are respectively determined in a preset database containing a large number of person images based on a similarity algorithm, wherein the first neighbor person image set comprises K neighbor person images which are most similar to the first person image in the database, the second neighbor person image set comprises K neighbor person images which are most similar to the second person image in the database, and K is a preset value. It should be noted that the K neighbor personal images in each set of neighbor personal images are ranked according to the similarity to the personal images, and the ranked neighbor personal images with the greater similarity to the personal images are ranked further forward. Next, acquiring neighbor character images in the second neighbor character image set one by one, searching a target neighbor character image matched with the acquired neighbor character image in the first neighbor character image set every time acquiring a neighbor character image in the second neighbor character image set, determining a sorting position number of the target neighbor character image in the first neighbor character image set as a sub-distance corresponding to the acquired neighbor character image if the target neighbor character image matched with the acquired neighbor character image is found, and determining a preset distance (which is a larger numerical value) as a sub-distance corresponding to the acquired neighbor character image if the target neighbor character image matched with the acquired neighbor character image is not found. And finally, calculating the sum of the sub-distances corresponding to all the neighbor person images in the second neighbor person image set, wherein the sum is the distance from the first person image to the second person image.
In carrying out the present disclosure, the inventors found that at least the following problems exist:
in the above process, it is necessary to determine a first neighboring personal image set corresponding to the first personal image and a second neighboring personal image set corresponding to the second personal image based on a similarity algorithm. The more the number of the neighbor personal images in which the actual person corresponding to the first personal image and the actual person corresponding to the first personal image are the same actual person are, and the more the number of the neighbor personal images in which the actual person corresponding to the second personal image and the actual person corresponding to the second personal image are the same actual person are, the more accurate the distance between the first personal image and the second personal image is to be finally calculated. However, in practical applications, it is likely that the first neighboring personal image set determined based on the similarity algorithm includes neighboring personal images in which the corresponding actual persons and the actual persons corresponding to the first personal images are not the same actual persons, and it is likely that the second neighboring personal image set determined includes neighboring personal images in which the corresponding actual persons and the actual persons corresponding to the second personal images are not the same actual persons, which may reduce the accuracy of the calculated distance between the first personal image and the second personal image, and further may improve the possibility of generating an erroneous personal cluster in the personal clustering process.
Disclosure of Invention
In order to overcome the problems in the related art, the present disclosure provides the following technical solutions:
according to a first aspect of the embodiments of the present disclosure, there is provided a method for performing person clustering, the method including:
acquiring a target person image;
determining a first preset number of primary neighbor person images with the highest similarity to the target person image in a preset person image database as undetermined primary neighbor person images of the target person image;
for each primary neighbor person image, determining a second preset number of secondary neighbor person images with the highest similarity to the primary neighbor person image in the person image database, and if any one of the second preset number of secondary neighbor person images is the target person image, adding the primary neighbor person image to a determined neighbor person image set corresponding to the target person image;
and carrying out character clustering processing based on the neighbor character image set.
Optionally, before performing the person clustering process based on the set of neighbor person images, the method further includes:
for each secondary neighbor person image, determining a third preset number of tertiary neighbor person images with the highest similarity to the secondary neighbor person image in the person image database, and if the third preset number of tertiary neighbor person images meet a preset person image overlap ratio condition, adding the secondary neighbor person image to the neighbor person image set.
Optionally, the adding the second-level neighbor personal images to the neighbor personal image set if the third preset number of third-level neighbor personal images satisfy a preset personal image overlap ratio condition includes:
determining three-level neighbor personal images, which are the same as any one of the determined one-level neighbor personal images of the target personal image in the neighbor personal image set, in the third preset number of three-level neighbor personal images;
and if the ratio of the determined number of the three-level neighbor person images to the third preset number is greater than a preset ratio threshold, adding the two-level neighbor person images to the neighbor person image set.
Optionally, after determining a first preset number of primary neighbor personal images having the highest similarity with the target personal image as pending primary neighbor personal images of the target personal image, the method further includes:
determining a first type distance between each undetermined primary neighbor character image of the target character image and the target character image;
the character clustering processing based on the neighbor character image set comprises the following steps:
for each neighbor person image in the neighbor person image set, determining a first type distance corresponding to a primary neighbor person image to be determined of the target person image which is the same as the neighbor person image as a first type distance corresponding to the neighbor person image, and determining a target distance corresponding to the neighbor person image based on the first type distance;
sequencing all the neighbor character images in the neighbor character image set according to the sequence of the corresponding target distances from small to large to obtain a sequenced neighbor character image set;
and carrying out character clustering processing based on the sorted neighbor character image set.
Optionally, the determining, based on the first type of distance, a target distance corresponding to the neighboring person image includes:
determining a second type of distance between the neighbor personal image and the target personal image;
and carrying out weighted summation on the first type distance and the second type distance corresponding to the neighbor person image to obtain a target distance between the neighbor person image and the target person image.
Optionally, the first type of distance comprises a cosine distance, a first order norm L1 distance, or a second order norm L2 distance, and the second type of distance comprises a jacadre Jaccard distance.
According to a second aspect of the embodiments of the present disclosure, there is provided an apparatus for performing person clustering, the apparatus including:
the acquisition module is used for acquiring a target person image;
the determining module is used for determining a first preset number of primary neighbor person images with the highest similarity to the target person image in a preset person image database as undetermined primary neighbor person images of the target person image; for each primary neighbor person image, determining a second preset number of secondary neighbor person images with the highest similarity to the primary neighbor person image in the person image database, and if any one of the second preset number of secondary neighbor person images is the target person image, adding the primary neighbor person image to a determined neighbor person image set corresponding to the target person image;
and the clustering module is used for carrying out character clustering processing on the basis of the neighbor character image set.
Optionally, the determining module is further configured to:
for each secondary neighbor person image, determining a third preset number of tertiary neighbor person images with the highest similarity to the secondary neighbor person image in the person image database, and if the third preset number of tertiary neighbor person images meet a preset person image overlap ratio condition, adding the secondary neighbor person image to the neighbor person image set.
Optionally, the determining module is configured to:
determining three-level neighbor personal images, which are the same as any one of the determined one-level neighbor personal images of the target personal image in the neighbor personal image set, in the third preset number of three-level neighbor personal images;
and if the ratio of the determined number of the three-level neighbor person images to the third preset number is greater than a preset ratio threshold, adding the two-level neighbor person images to the neighbor person image set.
Optionally, the determining module is further configured to determine a first type distance between each undetermined primary neighbor personal image of the target personal image and the target personal image;
the clustering module is used for determining a first type distance corresponding to a primary neighbor personal image to be determined of the target personal image which is the same as the neighbor personal image as a first type distance corresponding to the neighbor personal image for each neighbor personal image in the neighbor personal image set, and determining a target distance corresponding to the neighbor personal image based on the first type distance; sequencing all the neighbor character images in the neighbor character image set according to the sequence of the corresponding target distances from small to large to obtain a sequenced neighbor character image set; and carrying out character clustering processing based on the sorted neighbor character image set.
Optionally, the clustering module is configured to:
determining a second type of distance between the neighbor personal image and the target personal image;
and carrying out weighted summation on the first type distance and the second type distance corresponding to the neighbor person image to obtain a target distance between the neighbor person image and the target person image.
Optionally, the first type of distance comprises a cosine distance, a first order norm L7 distance, or a second order norm L2 distance, and the second type of distance comprises a jacadre Jaccard distance.
According to a third aspect of embodiments of the present disclosure, there is provided a server comprising a processor, a communication interface, a memory, and a communication bus, wherein:
the processor, the communication interface and the memory complete mutual communication through the communication bus;
the memory is used for storing a computer program;
the processor is used for executing the program stored in the memory so as to realize the method for carrying out the character clustering.
According to a fourth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, in which a computer program is stored, and the computer program, when executed by a processor, implements the above method for clustering people.
The technical scheme provided by the embodiment of the disclosure can have the following beneficial effects:
in the embodiment of the present disclosure, after the undetermined primary neighbor personal image of the target person is determined, for each primary neighbor personal image in the undetermined primary neighbor personal image, it may be determined whether it is adjacent to the target personal image, that is, whether a personal image identical to the target personal image appears in the secondary neighbor personal image of the primary neighbor personal image. If the same person image as the target person image appears in the secondary neighbor person image of the primary neighbor person image, the primary neighbor person image and the target person image are in close proximity to each other, and the primary neighbor person image can be used as the determined primary neighbor person image of the target person image. The likelihood that the actual person corresponding to the determined primary neighbor personal image and the actual person corresponding to the target personal image are the same actual person is increased. And then, the accuracy of the distance between the figure images calculated based on the determined first-level neighbor figure images is improved subsequently, and the accuracy of figure clustering processing is also improved correspondingly.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and together with the description, serve to explain the principles of the disclosure. In the drawings:
FIG. 1 is a schematic flow diagram illustrating a method for people clustering according to an exemplary embodiment;
FIG. 2 is a diagram illustrating a neighbor image according to an exemplary embodiment;
FIG. 3 is a flow diagram illustrating a method of clustering people according to an exemplary embodiment;
FIG. 4 is a flowchart illustrating a method of clustering people according to an exemplary embodiment;
FIG. 5 is a flowchart illustrating a method of clustering people according to an exemplary embodiment;
FIG. 6 is a schematic diagram illustrating a distance calculation according to an exemplary embodiment;
FIG. 7 is a diagram illustrating an effect data according to an exemplary embodiment;
FIG. 8 is a block diagram of an apparatus for clustering persons according to an exemplary embodiment;
fig. 9 is a schematic diagram illustrating a configuration of a server according to an example embodiment.
With the foregoing drawings in mind, certain embodiments of the disclosure have been shown and described in more detail below. These drawings and written description are not intended to limit the scope of the disclosed concepts in any way, but rather to illustrate the concepts of the disclosure to those skilled in the art by reference to specific embodiments.
Detailed Description
Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the exemplary embodiments below are not intended to represent all implementations consistent with the present disclosure. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure, as detailed in the appended claims.
An exemplary embodiment of the present disclosure provides a method for clustering people, and as shown in fig. 1, a processing flow of the method may include the following steps:
in step S110, a target person image is acquired.
In implementation, the method provided by the embodiment of the present disclosure may be executed in a server. When the character clustering process is performed, a plurality of character images to be clustered may be input to the server, the server may cluster different character images, which are the same as the actual characters corresponding to the character images to be clustered, together, and the character images to be clustered may include the target character image. In order to perform the person clustering process on the target personal image, first, one-level neighbor personal images corresponding to the target personal image may be determined.
Step S120, determining a first preset number of primary neighbor person images with the highest similarity to the target person image in a preset person image database as undetermined primary neighbor person images of the target person.
In an implementation, a person image database may be provided in the server, and a large number of person images, which may be referred to as sample person images, are stored in the person image database. The neighbor images for any given person image may be determined in the person image database by the rank-order algorithm. Therefore, K primary neighbor images having the highest similarity with the target person image can be determined in the person image database, and K can be set as needed. The K first-level neighbor person images can be ranked from high to low according to the similarity with the target person image, and the more front the ranking, the higher the similarity between the first-level neighbor person image and the target person image. It should be noted that the level relationship between the neighboring person image and the target person may be divided according to whether the neighboring person image and the target person have a direct relationship or a near-far relationship having an indirect relationship. The neighbor personal image having a direct relationship with the target person may be referred to as a first-order neighbor personal image, a K-nearest neighbors (KNN) personal image. The neighbor personal image of the primary neighbor personal image has an indirect relationship with the target person, and is in contact with the target person through the primary neighbor personal image as a medium, so that the neighbor personal image of the primary neighbor personal image can be called a secondary neighbor personal image, and by analogy, a tertiary neighbor personal image and the like corresponding to the target personal image can be determined.
The primary neighbor personal image of the target personal image may be determined based on a similarity algorithm as the primary neighbor personal image to be determined of the target person. The pending primary neighbor personal image of the target personal image may be represented as N (p, k), then:
wherein,representing a pending primary neighbour figure. k is the number of all pending primary neighbor persona images in N (p, k). p is the target person image.
The more the number of the primary neighbor character images in which the actual character corresponding to the primary neighbor character image to be determined and the actual character corresponding to the target character image are the same actual character, the more accurate the result of the character clustering process based on the primary neighbor character image to be determined is. However, in practical applications, it is likely that the undetermined primary neighbor person images determined based on the similarity algorithm include primary neighbor person images (which may be referred to as hard negative sample person images) in which the corresponding actual persons and the actual persons corresponding to the target person images are not the same actual persons, which may introduce noise in calculating the rank-order distance and affect the result of the person clustering process.
As shown in fig. 2, the Probe is a target person image, and the person images P1-P4, N1-N6 are all pending primary neighbor person images of the Probe. The actual persons corresponding to P1-P4 and the actual persons corresponding to probes are the same actual persons, while the actual persons corresponding to N1-N6 and the actual persons corresponding to probes are not the same actual persons.
In the embodiment of the present disclosure, the primary neighbor person image of the target person image determined based on the similarity algorithm may be used as the undetermined primary neighbor person image of the target person, and the undetermined primary neighbor person image may be subsequently screened, so that the actual person that is more likely to correspond and the actual person that corresponds to the target person image are screened from the primary neighbor person image of the same actual person, and thus, the accuracy of the result of the person clustering process may be improved.
Step S130, for each primary neighbor personal image, determining a second preset number of secondary neighbor personal images with the highest similarity to the primary neighbor personal image in the personal image database, and if any one of the second preset number of secondary neighbor personal images is a target personal image, adding the primary neighbor personal image to the determined neighbor personal image set corresponding to the target personal image.
In implementation, after the undetermined primary neighbor character image of the target character is determined, the undetermined primary neighbor character images are K, whether each primary neighbor character image in the K undetermined primary neighbor character images is adjacent to the target character image (also called K-reciprocal neighbor characters), that is, whether a character image identical to the target character image appears in the secondary neighbor character image of the primary neighbor character image, if a character image identical to the target character image appears in the secondary neighbor character image of the primary neighbor character image, the primary neighbor character image is adjacent to the target character image, and the primary neighbor character image can be used as the determined primary neighbor character image of the target character image. The primary neighbor image that is close to the target person image can be denoted as R (p, k), and then:
R(p,k)＝{gi|(gi∈N(p,k)∧(p∈N(gik)) } (equation 2)
Wherein, giIs a primary neighbor image which is adjacent to the target person image. p is the target person image.
For each primary neighbor person image, a second preset number of secondary neighbor person images having the highest similarity to the primary neighbor person image may be searched in the person image database. The first predetermined number and the second predetermined number may be the same or different. After the second preset number of secondary neighbor personal images are found, it may be determined whether the second preset number of secondary neighbor personal images include the same secondary neighbor personal image as the target personal image. If the second preset number of secondary neighbor personal images include secondary neighbor personal images identical to the target personal image, it is determined that the primary neighbor personal image and the target personal image are mutually adjacent, and the primary neighbor personal image can be used as the determined primary neighbor personal image of the target personal image, and can be added to the neighbor personal image set corresponding to the target personal image. It should be noted that the neighbor personal image set corresponding to the target personal image may be an empty set at an initial time, and the neighbor personal images that are needed to be used in the process of performing the personal clustering process on the target personal image may be added to the neighbor personal image set subsequently.
As shown in fig. 2, the Probe is a target person image, and the person images P1-P4, N1-N6 are all pending primary neighbor person images of the Probe. Two level neighbor images corresponding to P1-P4, N1-N6 may be determined, respectively, as a series of small graphs pointed to by dashed arrows. It is possible to search for whether or not there is a thumbnail identical to the target person image in these thumbnails, and if a thumbnail identical to the target person image is found, one of the person images P1-P4, N1-N6 corresponding to this thumbnail is adjacent to the Probe and may be added to the set of neighbor person images corresponding to the Probe.
And step S140, carrying out character clustering processing based on the neighbor character image set.
In implementation, the neighbor personal images in the set of neighbor personal images may be ranked, and the neighbor personal images having a high similarity with the target personal image may be ranked in the front. And carrying out character clustering processing according to the sorted neighbor character image set.
For example, it is necessary to determine whether the actual person corresponding to the first person image is the actual person corresponding to the second person image, the first person image and the second person image may be used as the target person images in the embodiment of the present disclosure, and the sorted neighboring person image set corresponding to the first person image and the sorted neighboring person image set corresponding to the second person image are determined by the method provided in the embodiment of the present disclosure. And calculating the distance between the first person image and the second person image based on the two sorted neighbor person image sets, wherein the smaller the distance is, the greater the possibility that the actual person corresponding to the first person image is the actual person corresponding to the second person image is. And if the calculated distance between the first person image and the second person image is smaller than a preset distance threshold value, determining that the actual person corresponding to the first person image is the actual person corresponding to the second person image, and determining the first person image as another person image of the actual person corresponding to the second person image, so as to finish the person clustering processing.
In the embodiment of the present disclosure, after the undetermined primary neighbor personal image of the target person is determined, for each primary neighbor personal image in the undetermined primary neighbor personal image, it may be determined whether it is adjacent to the target personal image, that is, whether a personal image identical to the target personal image appears in the secondary neighbor personal image of the primary neighbor personal image. If the same person image as the target person image appears in the secondary neighbor person image of the primary neighbor person image, the primary neighbor person image and the target person image are in close proximity to each other, and the primary neighbor person image can be used as the determined primary neighbor person image of the target person image. The likelihood that the actual person corresponding to the determined primary neighbor personal image and the actual person corresponding to the target personal image are the same actual person is increased. And then, the accuracy of the distance between the figure images calculated based on the determined first-level neighbor figure images is improved subsequently, and the accuracy of figure clustering processing is also improved correspondingly.
Based on the same inventive concept, an exemplary embodiment of the present disclosure provides a method for clustering people, as shown in fig. 3, a processing flow of the method may include the following steps:
in step S310, a target person image is acquired.
In implementation, the method provided by the embodiment of the present disclosure may be executed in a server. When the character clustering process is performed, a plurality of character images to be clustered may be input to the server, the server may cluster different character images, which are the same as the actual characters corresponding to the character images to be clustered, together, and the character images to be clustered may include the target character image. In order to perform the person clustering process on the target personal image, first, one-level neighbor personal images corresponding to the target personal image may be determined.
Step S320 is to determine a first preset number of primary neighboring character images with the highest similarity to the target character image in a preset character image database as undetermined primary neighboring character images of the target character.
In implementation, the primary neighbor character image of the target character image determined based on the similarity algorithm can be used as the undetermined primary neighbor character image of the target character, the undetermined primary neighbor character image can be subsequently screened, and the primary neighbor character image of the actual character which is likely to correspond and the actual character which corresponds to the target character image are the same actual character is screened, so that the accuracy of the result of character clustering processing can be improved.
Step S330, for each primary neighbor image, determining a second preset number of secondary neighbor images with the highest similarity with the primary neighbor image in the image database, and if any one of the second preset number of secondary neighbor images is a target image, adding the primary neighbor image to a determined neighbor image set corresponding to the target image.
In an implementation, for each primary neighbor personal image, a second preset number of secondary neighbor personal images having the highest similarity to the primary neighbor personal image may be searched in the personal image database. The first predetermined number and the second predetermined number may be the same or different. After the second preset number of secondary neighbor personal images are found, it may be determined whether the second preset number of secondary neighbor personal images include the same secondary neighbor personal image as the target personal image. If the second preset number of secondary neighbor personal images include secondary neighbor personal images identical to the target personal image, it is determined that the primary neighbor personal image and the target personal image are mutually adjacent, and the primary neighbor personal image can be used as the determined primary neighbor personal image of the target personal image, and can be added to the neighbor personal image set corresponding to the target personal image. It should be noted that the neighbor personal image set corresponding to the target personal image may be an empty set at an initial time, and the neighbor personal images that are needed to be used in the process of performing the personal clustering process on the target personal image may be added to the neighbor personal image set subsequently.
Step S340, for each secondary neighbor character image, determining a third preset number of tertiary neighbor character images with the highest similarity with the secondary neighbor character image in the character image database, and if the third preset number of tertiary neighbor character images meet the preset character image contact ratio condition, adding the secondary neighbor character images into the neighbor character image set.
In implementation, because some actual people which are more likely to correspond to the image to be screened from the first neighbor people image to be determined and the actual people which correspond to the target people image are the primary neighbor people image of the same actual person, the number of the screened primary neighbor people images may be reduced, and thus the number of the neighbor people images which can be referred to is reduced. Further, a sample personal image in which the actual person corresponding to the target personal image is the same actual person may be excluded from the primary neighbor personal images due to factors such as illumination, posture, shooting angle of view, and the like. In the embodiment of the present disclosure, some other neighbor personal images that are screened out in other manners and are more likely to correspond to the actual person and the actual person corresponding to the target personal image are the same actual person may be added to the neighbor personal image set, so as to enrich the neighbor personal image set and possibly replace other available sample personal images into the neighbor personal image set again. The other neighbor personal images may be a secondary neighbor personal image, a tertiary neighbor personal image, etc. corresponding to the target personal image.
In one possible implementation manner, after the second-level neighbor personal image corresponding to the target personal image is determined, for each second-level neighbor personal image, a third preset number of third-level neighbor personal images having the highest similarity with the second-level neighbor personal image may be determined in the personal image database. The third predetermined number may be the same as or different from the first predetermined number, or the second predetermined number. In one possible implementation, both the first preset number and the second preset number may be set to K, and the third preset number may be set to K/2. After determining a third preset number of three-level neighbor character images, it may be determined whether the third preset number of three-level neighbor character images satisfy a preset character image overlap condition, and if the third preset number of three-level neighbor character images can satisfy the preset character image overlap condition, the second-level neighbor character images may be added to the neighbor character image set to supplement the number of neighbor character images in the neighbor character image set.
Alternatively, if a third preset number of the three-level neighbor personal images satisfy the preset personal image overlap ratio condition, the step of adding the two-level neighbor personal images to the set of neighbor personal images may include: determining three-level neighbor character images which are the same as any one determined one-level neighbor character image of the target character image in the neighbor character image set in a third preset number of three-level neighbor character images; and if the ratio of the determined number of the three-level neighbor character images to the third preset number is larger than the preset ratio threshold, adding the two-level neighbor character images into the neighbor character image set.
In implementation, after the third preset number of three-level neighbor personal images of each two-level neighbor personal image are determined, for the third preset number of three-level neighbor personal images of each two-level neighbor personal image, the three-level neighbor personal images identical to the determined one-level neighbor personal image of the target personal image in the neighbor personal image set can be found from the third preset number of three-level neighbor personal images, and the number of the found three-level neighbor personal images is counted. The set of neighbor personal images includes the determined primary neighbor personal image and the post-supplemented secondary neighbor personal image of the target personal image, and the neighbor personal images can be searched and matched from a third preset number of tertiary neighbor personal imagesThree-level neighbor personal images that are identical to the determined primary neighbor personal image of the target personal image in the collection. If the ratio of the counted number to the third preset number is greater than the preset ratio threshold, the secondary neighbor personal image may be added to the set of neighbor personal images. The preset proportional value threshold may be 2/3. The set of neighbor personal images can be denoted as R*(p, k), then:
wherein,the two-level neighbor figure image meeting the preset figure image contact ratio condition is obtained. The condition is recorded as s.t., specificallyConditions of, i.e.Satisfy the requirement ofMay be incorporated into R*(p, k).
And step S350, carrying out character clustering processing based on the neighbor character image set.
In implementation, the neighbor personal images in the set of neighbor personal images may be ranked, and the neighbor personal images having a high similarity with the target personal image may be ranked in the front. And carrying out character clustering processing according to the sorted neighbor character image set.
In the embodiment of the present disclosure, after the undetermined primary neighbor personal image of the target person is determined, for each primary neighbor personal image in the undetermined primary neighbor personal image, it may be determined whether it is adjacent to the target personal image, that is, whether a personal image identical to the target personal image appears in the secondary neighbor personal image of the primary neighbor personal image. If the same person image as the target person image appears in the secondary neighbor person image of the primary neighbor person image, the primary neighbor person image and the target person image are in close proximity to each other, and the primary neighbor person image can be used as the determined primary neighbor person image of the target person image. The likelihood that the actual person corresponding to the determined primary neighbor personal image and the actual person corresponding to the target personal image are the same actual person is increased. And then, the accuracy of the distance between the figure images calculated based on the determined first-level neighbor figure images is improved subsequently, and the accuracy of figure clustering processing is also improved correspondingly.
Based on the same inventive concept, an exemplary embodiment of the present disclosure provides a method for clustering people, as shown in fig. 4, a processing flow of the method may include the following steps:
in step S410, a target person image is acquired.
In implementation, the method provided by the embodiment of the present disclosure may be executed in a server. When the character clustering process is performed, a plurality of character images to be clustered may be input to the server, the server may cluster different character images, which are the same as the actual characters corresponding to the character images to be clustered, together, and the character images to be clustered may include the target character image. In order to perform the person clustering process on the target personal image, first, one-level neighbor personal images corresponding to the target personal image may be determined.
Step S420, in the preset character image database, determining a first preset number of primary neighbor character images with the highest similarity to the target character image as undetermined primary neighbor character images of the target character.
In implementation, the primary neighbor character image of the target character image determined based on the similarity algorithm can be used as the undetermined primary neighbor character image of the target character, the undetermined primary neighbor character image can be subsequently screened, and the primary neighbor character image of the actual character which is likely to correspond and the actual character which corresponds to the target character image are the same actual character is screened, so that the accuracy of the result of character clustering processing can be improved.
Step S430, determining a first type distance between each undetermined primary neighbor personal image of the target personal image and the target personal image.
In an implementation, after the pending primary neighbor images of the target person image are determined, a first type of distance between each pending primary neighbor image and the target person image may be calculated. Optionally, the first type of distance comprises a cosine distance, an L1 (first order norm) distance, or an L2 (second order norm) distance.
Step S440, for each primary neighbor personal image, determining a second preset number of secondary neighbor personal images with the highest similarity to the primary neighbor personal image in the personal image database, and if any one of the second preset number of secondary neighbor personal images is a target personal image, adding the primary neighbor personal image to the determined neighbor personal image set corresponding to the target personal image.
In an implementation, for each primary neighbor personal image, a second preset number of secondary neighbor personal images having the highest similarity to the primary neighbor personal image may be searched in the personal image database. The first predetermined number and the second predetermined number may be the same or different. After the second preset number of secondary neighbor personal images are found, it may be determined whether the second preset number of secondary neighbor personal images include the same secondary neighbor personal image as the target personal image. If the second preset number of secondary neighbor personal images include secondary neighbor personal images identical to the target personal image, it is determined that the primary neighbor personal image and the target personal image are mutually adjacent, and the primary neighbor personal image can be used as the determined primary neighbor personal image of the target personal image, and can be added to the neighbor personal image set corresponding to the target personal image. It should be noted that the neighbor personal image set corresponding to the target personal image may be an empty set at an initial time, and the neighbor personal images that are needed to be used in the process of performing the personal clustering process on the target personal image may be added to the neighbor personal image set subsequently.
Step S450, for each neighbor personal image in the neighbor personal image set, determining a first type distance corresponding to a primary neighbor personal image to be determined of a target personal image same as the neighbor personal image as the first type distance corresponding to the neighbor personal image, and determining a target distance corresponding to the neighbor personal image based on the first type distance.
In an implementation, the set of neighbor personal images may include a determined primary neighbor personal image of the target personal image and a post-supplemented secondary neighbor personal image, for each of all the neighbor personal images, a pending primary neighbor personal image of the target personal image that is the same as the neighbor personal image may be determined, the pending primary neighbor personal image may have a first type of distance, and the first type of distance may be determined as the first type of distance to which the neighbor personal image corresponds. And determining a target distance (also called Re-Rank-Order distance) corresponding to the neighbor person image based on the first type distance corresponding to the neighbor person image. The calculation formula of the target distance may be:
d*(p,gi)＝(1-λ)×dJ(p,gi)+λd(p,gi) (formula 4)
Wherein d is*(p,gi) For the neighbor person image g in the neighbor person image setiAnd the target distance between the target person image p. dJ(p,gi) For the neighbor person image g in the neighbor person image setiCorresponding second type of distance, d (p, g)i) For the neighbor person image g in the neighbor person image setiA corresponding first type of distance. And lambda is a weight coefficient and can be set according to requirements.
Although the number of the undetermined primary neighbor personal images of the target personal image is the first preset number, in practical applications, a fourth preset number of undetermined primary neighbor personal images that is far larger than the first preset number may be determined. However, because the similarity between the first-level neighbor character images, which are arranged behind the first preset number, in the fourth preset number of undetermined first-level neighbor character images and the target character image is much different, the first-level neighbor character images, which are arranged behind the first preset number, may not be used as a reference for calculating the distance in the process of performing character clustering processing on the target character image. In the fourth preset number of pending primary neighbor images, it is likely that all the neighbor images in the set of neighbor images can be found. Each undetermined primary neighbor figure image in the fourth preset number of undetermined primary neighbor figure images corresponds to a first type distance, and the first type distance corresponding to the primary neighbor figure image in the fourth preset number of undetermined primary neighbor figure images which is the same as the neighbor figure images in the neighbor figure image set can be determined as the first type distance corresponding to the neighbor figure image in the neighbor figure image set.
Optionally, the step of determining the target distance corresponding to the neighbor person image based on the first type distance may include: determining a second type of distance between the neighbor personal image and the target personal image; and carrying out weighted summation on the first type distance and the second type distance corresponding to the neighbor person image to obtain the target distance between the neighbor person image and the target person image.
In an implementation, after all the neighboring personal images in the set of neighboring personal images are determined, the second type of distance between each neighboring personal image and the target personal image may be calculated, and the target distance between each neighboring personal image and the target personal image may be determined based on the first type of distance and the second type of distance respectively corresponding to each neighboring personal image. The first type distance and the second type distance corresponding to the neighbor person image may be weighted and summed, and the result after weighted and summed is the target distance between the neighbor person image and the target person image. Alternatively, the second type of distance may comprise a Jaccard distance. The calculation formula of Jaccard distance is as follows:
wherein d isJ(p,gi) For the neighbor person image g in the neighbor person image setiAnd the Jaccard distance between the target person image p. R*(giK) is a neighbor image giK neighboring character images.
As shown in fig. 5, based on the Probe, a Gallery set corresponding to the Probe may be determined. In the Gallery set, the undetermined primary neighbor image corresponding to the Probe and the neighbor image set are determined. Respectively calculating the distances between the to-be-determined primary neighbor personal image corresponding to the Probe and each neighbor personal image and the Probe in the neighbor personal image set, and sequentially recording the distances as d and dJFinally, pair d and dJAnd performing weighted summation to output the final d.
Since different types of distances can describe the distance between the neighbor personal image and the target personal image from different angles, the different types of distances can be weighted and summed, the weighted and summed distances can better comprehensively describe the distance between the neighbor personal image and the target personal image, and the accuracy of the weighted and summed distance between the neighbor personal image and the target personal image is higher.
Step S460, sorting all the neighbor character images in the neighbor character image set according to the sequence of the corresponding target distances from small to large to obtain a sorted neighbor character image set.
In implementation, each of all the neighbor personal images in the neighbor personal image set corresponds to a target distance, and all the neighbor personal images in the neighbor personal image set can be sorted according to the sequence of the corresponding target distances from small to large to obtain the sorted neighbor personal image set.
And step S470, carrying out character clustering processing based on the sorted neighbor character image set.
In implementation, the distance between the first person image and the second person image can be calculated by a person clustering algorithm, and a smaller distance indicates that the actual person corresponding to the first person image is more likely to be the actual person corresponding to the second person image. And if the calculated distance between the first person image and the second person image is smaller than a preset distance threshold value, determining that the actual person corresponding to the first person image is the actual person corresponding to the second person image, and determining the first person image as another person image of the actual person corresponding to the second person image, thereby completing the person clustering process.
As shown in fig. 6, in the process of calculating the distance between the first person image and the second person image, the first person image O needs to be determined in a database containing a large number of person images and based on a similarity algorithmbCorresponding first set of neighbor personal images and second personal image OaAnd a corresponding second neighbor personal image set (each circle in the figure may represent a neighbor personal image), where the first neighbor personal image set includes K neighbor personal images that are most similar to the first personal image in the database, the second neighbor personal image set includes K neighbor personal images that are most similar to the second personal image in the database, and K is a preset numerical value. It should be noted that the K neighbor personal images in each set of neighbor personal images are ranked according to the similarity to the personal images, and the ranked neighbor personal images with the greater similarity to the personal images are ranked further forward. Next, acquiring neighbor character images in the second neighbor character image set one by one, searching a target neighbor character image matched with the acquired neighbor character image in the first neighbor character image set every time acquiring a neighbor character image in the second neighbor character image set, determining a sorting position number of the target neighbor character image in the first neighbor character image set as a sub-distance corresponding to the acquired neighbor character image if the target neighbor character image matched with the acquired neighbor character image is found, and determining a preset distance (which is a larger numerical value) as a sub-distance corresponding to the acquired neighbor character image if the target neighbor character image matched with the acquired neighbor character image is not found.
For example, in FIG. 6, a first human image O is calculatedbTo the second person image OaThe distance between the first neighbor image set and the second neighbor image set can be obtained one by one, the neighbor images in the second neighbor image set sequentially include a, c, d and b, the neighbor images respectively identical to a, b, c and d are sequentially searched in the neighbor images in the first neighbor image set, and the final distance is 5+2+4+0 if the neighbor images identical to a in the first neighbor image set are found in the position No. 5, the neighbor images identical to c in the position No. 2, the neighbor images identical to d in the position No. 4 and the neighbor images identical to b in the position No. 0, and the result is 11.
And finally, calculating the sum of the sub-distances corresponding to all the neighbor person images in the second neighbor person image set, wherein the sum is the distance from the first person image to the second person image.
The method provided by the embodiment of the disclosure is used for clustering people, and the obtained effect can be shown in fig. 7. In fig. 7, the effect data of the character clustering process by different methods is counted by comparing two data sets including mc _7days _ new _ probe.pb data set and wan _3days _ new _ probe.pb data set, where the filling of "none" indicates empty, and the "Rank-order" entry corresponds to the effect data of the character clustering process by one method, and the "Re-ranker" entry corresponds to the effect data of the character clustering process by another method.
In the embodiment of the present disclosure, after the undetermined primary neighbor personal image of the target person is determined, for each primary neighbor personal image in the undetermined primary neighbor personal image, it may be determined whether it is adjacent to the target personal image, that is, whether a personal image identical to the target personal image appears in the secondary neighbor personal image of the primary neighbor personal image. If the same person image as the target person image appears in the secondary neighbor person image of the primary neighbor person image, the primary neighbor person image and the target person image are in close proximity to each other, and the primary neighbor person image can be used as the determined primary neighbor person image of the target person image. The likelihood that the actual person corresponding to the determined primary neighbor personal image and the actual person corresponding to the target personal image are the same actual person is increased. And then, the accuracy of the distance between the figure images calculated based on the determined first-level neighbor figure images is improved subsequently, and the accuracy of figure clustering processing is also improved correspondingly.
Still another exemplary embodiment of the present disclosure provides an apparatus for person clustering, as shown in fig. 8, the apparatus including:
an obtaining module 710, configured to obtain a target person image;
a determining module 720, configured to determine, in a preset person image database, first preset number of primary neighbor person images with the highest similarity to the target person image as undetermined primary neighbor person images of the target person image; for each primary neighbor person image, determining a second preset number of secondary neighbor person images with the highest similarity to the primary neighbor person image in the person image database, and if any one of the second preset number of secondary neighbor person images is the target person image, adding the primary neighbor person image to a determined neighbor person image set corresponding to the target person image;
and the clustering module 730 is used for clustering people based on the neighbor people image set.
Optionally, the determining module 720 is further configured to:
for each secondary neighbor person image, determining a third preset number of tertiary neighbor person images with the highest similarity to the secondary neighbor person image in the person image database, and if the third preset number of tertiary neighbor person images meet a preset person image overlap ratio condition, adding the secondary neighbor person image to the neighbor person image set.
Optionally, the determining module 720 is configured to:
determining three-level neighbor personal images, which are the same as any one of the determined one-level neighbor personal images of the target personal image in the neighbor personal image set, in the third preset number of three-level neighbor personal images;
and if the ratio of the determined number of the three-level neighbor person images to the third preset number is greater than a preset ratio threshold, adding the two-level neighbor person images to the neighbor person image set.
Optionally, the determining module 720 is further configured to determine a first type distance between each pending primary neighbor personal image of the target personal image and the target personal image;
the clustering module 730 is configured to determine, for each neighboring personal image in the set of neighboring personal images, a first type distance corresponding to a primary neighboring personal image to be determined of the target personal image that is the same as the neighboring personal image as a first type distance corresponding to the neighboring personal image, and determine a target distance corresponding to the neighboring personal image based on the first type distance; sequencing all the neighbor character images in the neighbor character image set according to the sequence of the corresponding target distances from small to large to obtain a sequenced neighbor character image set; and carrying out character clustering processing based on the sorted neighbor character image set.
Optionally, the clustering module 730 is configured to:
determining a second type of distance between the neighbor personal image and the target personal image;
and carrying out weighted summation on the first type distance and the second type distance corresponding to the neighbor person image to obtain a target distance between the neighbor person image and the target person image.
Optionally, the first type of distance comprises a cosine distance, a first order norm L7 distance, or a second order norm L2 distance, and the second type of distance comprises a jacadre Jaccard distance.
With regard to the apparatus in the above-described embodiment, the specific manner in which each module performs the operation has been described in detail in the embodiment related to the method, and will not be elaborated here.
In the embodiment of the present disclosure, after the undetermined primary neighbor personal image of the target person is determined, for each primary neighbor personal image in the undetermined primary neighbor personal image, it may be determined whether it is adjacent to the target personal image, that is, whether a personal image identical to the target personal image appears in the secondary neighbor personal image of the primary neighbor personal image. If the same person image as the target person image appears in the secondary neighbor person image of the primary neighbor person image, the primary neighbor person image and the target person image are in close proximity to each other, and the primary neighbor person image can be used as the determined primary neighbor person image of the target person image. The likelihood that the actual person corresponding to the determined primary neighbor personal image and the actual person corresponding to the target personal image are the same actual person is increased. And then, the accuracy of the distance between the figure images calculated based on the determined first-level neighbor figure images is improved subsequently, and the accuracy of figure clustering processing is also improved correspondingly.
It should be noted that: in the embodiment, when people are clustered, the device for clustering people is only illustrated by the division of the functional modules, and in practical applications, the function distribution can be completed by different functional modules according to needs, that is, the internal structure of the server is divided into different functional modules, so as to complete all or part of the functions described above. In addition, the device for performing person clustering and the method for performing person clustering provided by the above embodiments belong to the same concept, and specific implementation processes thereof are described in the method embodiments in detail, and are not described herein again.
Fig. 9 shows a schematic structural diagram of a server 1900 provided in an exemplary embodiment of the present disclosure. The server 1900 may have a large difference due to different configurations or performances, and may include one or more processors (CPUs) 1910 and one or more memories 1920. The memory 1920 has stored therein at least one instruction, which is loaded and executed by the processor 1910 to implement the method for clustering people according to the above embodiments.
Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This application is intended to cover any variations, uses, or adaptations of the disclosure following, in general, the principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.
It will be understood that the present disclosure is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims (10)

1. A method for clustering people, the method comprising:
acquiring a target person image;
determining a first preset number of primary neighbor person images with the highest similarity to the target person image in a preset person image database as undetermined primary neighbor person images of the target person image;
for each primary neighbor person image, determining a second preset number of secondary neighbor person images with the highest similarity to the primary neighbor person image in the person image database, and if any one of the second preset number of secondary neighbor person images is the target person image, adding the primary neighbor person image to a determined neighbor person image set corresponding to the target person image;
and carrying out character clustering processing based on the neighbor character image set.
2. The method of claim 1, wherein prior to performing person clustering based on the set of neighbor person images, the method further comprises:
for each secondary neighbor person image, determining a third preset number of tertiary neighbor person images with the highest similarity to the secondary neighbor person image in the person image database, and if the third preset number of tertiary neighbor person images meet a preset person image overlap ratio condition, adding the secondary neighbor person image to the neighbor person image set.
3. The method of claim 2, wherein adding the secondary neighbor personal images to the set of neighbor personal images if the third preset number of tertiary neighbor personal images satisfies a preset personal image overlap condition comprises:
determining three-level neighbor personal images, which are the same as any one of the determined one-level neighbor personal images of the target personal image in the neighbor personal image set, in the third preset number of three-level neighbor personal images;
and if the ratio of the determined number of the three-level neighbor person images to the third preset number is greater than a preset ratio threshold, adding the two-level neighbor person images to the neighbor person image set.
4. The method of claim 1, wherein after determining a first preset number of primary neighbor personal images having the highest degree of similarity with the target personal image as the pending primary neighbor personal images of the target personal image, the method further comprises:
determining a first type distance between each undetermined primary neighbor character image of the target character image and the target character image;
the character clustering processing based on the neighbor character image set comprises the following steps:
for each neighbor person image in the neighbor person image set, determining a first type distance corresponding to a primary neighbor person image to be determined of the target person image which is the same as the neighbor person image as a first type distance corresponding to the neighbor person image, and determining a target distance corresponding to the neighbor person image based on the first type distance;
sequencing all the neighbor character images in the neighbor character image set according to the sequence of the corresponding target distances from small to large to obtain a sequenced neighbor character image set;
and carrying out character clustering processing based on the sorted neighbor character image set.
5. The method of claim 4, wherein determining the target distance corresponding to the neighboring person image based on the first type of distance comprises:
determining a second type of distance between the neighbor personal image and the target personal image;
and carrying out weighted summation on the first type distance and the second type distance corresponding to the neighbor person image to obtain a target distance between the neighbor person image and the target person image.
6. The method of claim 5, wherein the first type of distance comprises a cosine distance, a first order norm L1 distance, or a second order norm L2 distance, and wherein the second type of distance comprises a Jacard Jaccard distance.
7. An apparatus for clustering people, the apparatus comprising:
the acquisition module is used for acquiring a target person image;
the determining module is used for determining a first preset number of primary neighbor person images with the highest similarity to the target person image in a preset person image database as undetermined primary neighbor person images of the target person image; for each primary neighbor person image, determining a second preset number of secondary neighbor person images with the highest similarity to the primary neighbor person image in the person image database, and if any one of the second preset number of secondary neighbor person images is the target person image, adding the primary neighbor person image to a determined neighbor person image set corresponding to the target person image;
and the clustering module is used for carrying out character clustering processing on the basis of the neighbor character image set.
8. The apparatus of claim 7, wherein the determining module is further configured to:
for each secondary neighbor person image, determining a third preset number of tertiary neighbor person images with the highest similarity to the secondary neighbor person image in the person image database, and if the third preset number of tertiary neighbor person images meet a preset person image overlap ratio condition, adding the secondary neighbor person image to the neighbor person image set.
9. A server, comprising a processor, a communication interface, a memory, and a communication bus, wherein:
the processor, the communication interface and the memory complete mutual communication through the communication bus;
the memory is used for storing a computer program;
the processor is configured to execute the program stored in the memory to implement the method steps of any of claims 1-6.
10. A computer-readable storage medium, characterized in that a computer program is stored in the computer-readable storage medium, which computer program, when being executed by a processor, carries out the method steps of any one of claims 1 to 6.
CN201910636600.XA 2019-07-15 2019-07-15 Method, device, server and storage medium for person clustering Active CN110377774B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910636600.XA CN110377774B (en) 2019-07-15 2019-07-15 Method, device, server and storage medium for person clustering

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910636600.XA CN110377774B (en) 2019-07-15 2019-07-15 Method, device, server and storage medium for person clustering

Publications (2)

Publication Number Publication Date
CN110377774A true CN110377774A (en) 2019-10-25
CN110377774B CN110377774B (en) 2023-08-01

Family

ID=68253276

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910636600.XA Active CN110377774B (en) 2019-07-15 2019-07-15 Method, device, server and storage medium for person clustering

Country Status (1)

Country Link
CN (1) CN110377774B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112784086A (en) * 2021-01-28 2021-05-11 北京有竹居网络技术有限公司 Picture screening method and device, storage medium and electronic equipment

Citations (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030123731A1 (en) * 2001-11-30 2003-07-03 Sharp Laboratories Of America, Inc. Fast method of finding approximate nearest neighbor in high-dimensional space with a dynamically changing sample
KR101089504B1 (en) * 2010-06-22 2011-12-05 한국과학기술원 Method and apparatus of clustering image and method and apparatus of recognizing image using the same
CN103325122A (en) * 2013-07-03 2013-09-25 武汉大学 Pedestrian retrieval method based on bidirectional sequencing
CN103793721A (en) * 2014-03-04 2014-05-14 武汉大学 Pedestrian repeat recognition method and system based on area related feedback
CN104573681A (en) * 2015-02-11 2015-04-29 成都果豆数字娱乐有限公司 Face recognition method
US20150310309A1 (en) * 2014-04-29 2015-10-29 At&T Intellectual Property I, Lp Method and apparatus for organizing media content
CN105303150A (en) * 2014-06-26 2016-02-03 腾讯科技（深圳）有限公司 Method and system for implementing image processing
US20170177942A1 (en) * 2015-12-16 2017-06-22 Samsung Electronics Co., Ltd. Event-based image management using clustering
CN107818180A (en) * 2017-11-27 2018-03-20 北京小米移动软件有限公司 Video correlating method, image display method, device and storage medium
CN108228844A (en) * 2018-01-09 2018-06-29 美的集团股份有限公司 A kind of picture screening technique and device, storage medium, computer equipment
CN108733819A (en) * 2018-05-22 2018-11-02 深圳云天励飞技术有限公司 A kind of personnel's archives method for building up and device
CN108985232A (en) * 2018-07-18 2018-12-11 平安科技（深圳）有限公司 Facial image comparison method, device, computer equipment and storage medium
US20190042893A1 (en) * 2015-12-04 2019-02-07 Hewlett Packard Enterprise Development Lp Incremental clustering of a data stream via an orthogonal transform based indexing
US20190050672A1 (en) * 2015-12-03 2019-02-14 Hewlett Packard Enterprise Development Lp INCREMENTAL AUTOMATIC UPDATE OF RANKED NEIGHBOR LISTS BASED ON k-th NEAREST NEIGHBORS
CN109472208A (en) * 2018-10-12 2019-03-15 平安科技（深圳）有限公司 Certificates handling method, apparatus, computer equipment and storage medium based on recognition of face
CN109522937A (en) * 2018-10-23 2019-03-26 北京市商汤科技开发有限公司 Image processing method and device, electronic equipment and storage medium
US20190095463A1 (en) * 2017-09-27 2019-03-28 Shutterfly, Inc. System and method for reducing similar photos for display and product design
US20190146991A1 (en) * 2016-06-09 2019-05-16 Panasonic Intellectual Property Management Co., Ltd. Image search device, image search system, and image search method
CN109800664A (en) * 2018-12-28 2019-05-24 上海依图网络科技有限公司 A kind of method and device of determining passerby track

Patent Citations (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030123731A1 (en) * 2001-11-30 2003-07-03 Sharp Laboratories Of America, Inc. Fast method of finding approximate nearest neighbor in high-dimensional space with a dynamically changing sample
KR101089504B1 (en) * 2010-06-22 2011-12-05 한국과학기술원 Method and apparatus of clustering image and method and apparatus of recognizing image using the same
CN103325122A (en) * 2013-07-03 2013-09-25 武汉大学 Pedestrian retrieval method based on bidirectional sequencing
CN103793721A (en) * 2014-03-04 2014-05-14 武汉大学 Pedestrian repeat recognition method and system based on area related feedback
US20150310309A1 (en) * 2014-04-29 2015-10-29 At&T Intellectual Property I, Lp Method and apparatus for organizing media content
CN105303150A (en) * 2014-06-26 2016-02-03 腾讯科技（深圳）有限公司 Method and system for implementing image processing
CN104573681A (en) * 2015-02-11 2015-04-29 成都果豆数字娱乐有限公司 Face recognition method
US20190050672A1 (en) * 2015-12-03 2019-02-14 Hewlett Packard Enterprise Development Lp INCREMENTAL AUTOMATIC UPDATE OF RANKED NEIGHBOR LISTS BASED ON k-th NEAREST NEIGHBORS
US20190042893A1 (en) * 2015-12-04 2019-02-07 Hewlett Packard Enterprise Development Lp Incremental clustering of a data stream via an orthogonal transform based indexing
US20170177942A1 (en) * 2015-12-16 2017-06-22 Samsung Electronics Co., Ltd. Event-based image management using clustering
US20190146991A1 (en) * 2016-06-09 2019-05-16 Panasonic Intellectual Property Management Co., Ltd. Image search device, image search system, and image search method
US20190095463A1 (en) * 2017-09-27 2019-03-28 Shutterfly, Inc. System and method for reducing similar photos for display and product design
CN107818180A (en) * 2017-11-27 2018-03-20 北京小米移动软件有限公司 Video correlating method, image display method, device and storage medium
CN108228844A (en) * 2018-01-09 2018-06-29 美的集团股份有限公司 A kind of picture screening technique and device, storage medium, computer equipment
CN108733819A (en) * 2018-05-22 2018-11-02 深圳云天励飞技术有限公司 A kind of personnel's archives method for building up and device
CN108985232A (en) * 2018-07-18 2018-12-11 平安科技（深圳）有限公司 Facial image comparison method, device, computer equipment and storage medium
CN109472208A (en) * 2018-10-12 2019-03-15 平安科技（深圳）有限公司 Certificates handling method, apparatus, computer equipment and storage medium based on recognition of face
CN109522937A (en) * 2018-10-23 2019-03-26 北京市商汤科技开发有限公司 Image processing method and device, electronic equipment and storage medium
CN109800664A (en) * 2018-12-28 2019-05-24 上海依图网络科技有限公司 A kind of method and device of determining passerby track

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112784086A (en) * 2021-01-28 2021-05-11 北京有竹居网络技术有限公司 Picture screening method and device, storage medium and electronic equipment

Also Published As

Publication number Publication date
CN110377774B (en) 2023-08-01

Similar Documents

Publication Publication Date Title
Wu et al. Large scale incremental learning
Zhang et al. Detection of co-salient objects by looking deep and wide
Lei Ba et al. Predicting deep zero-shot convolutional neural networks using textual descriptions
Vicente et al. Leave-one-out kernel optimization for shadow detection
CN106203521B (en) The SAR image change detection learnt based on disparity map from step
CN113656596B (en) Multi-modal entity alignment method based on triple screening fusion
CN109871821B (en) Pedestrian re-identification method, device, equipment and storage medium of self-adaptive network
CN110689081A (en) Weak supervision target classification and positioning method based on bifurcation learning
CN102105901A (en) Annotating images
CN113407759A (en) Multi-modal entity alignment method based on adaptive feature fusion
US9842279B2 (en) Data processing method for learning discriminator, and data processing apparatus therefor
CN113129335B (en) Visual tracking algorithm and multi-template updating strategy based on twin network
CN112765362B (en) Knowledge-graph entity alignment method based on improved self-encoder and related equipment
US11860930B2 (en) Automatic image selection for visual consistency
Sumi et al. Improving classification accuracy using combined filter+ wrapper feature selection technique
Mazloom et al. Few-example video event retrieval using tag propagation
Yan et al. A novel clustering algorithm based on fitness proportionate sharing
CN110377774A (en) Carry out method, apparatus, server and the storage medium of personage's cluster
JP6017277B2 (en) Program, apparatus and method for calculating similarity between contents represented by set of feature vectors
Wang et al. An efficient method to detect mutual overlap of a large set of unordered images for structure-from-motion
Zheng et al. Learn++ for robust object tracking
Peng et al. A ranking based attention approach for visual tracking
Veeramuthu et al. A plug-in feature extraction and feature subset selection algorithm for classification of medicinal brain image data
JP2009032154A (en) Information processor and program
Dao et al. Signature-image-based event analysis for personal photo albums

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant