WO2022193232A1 - 人脸聚类方法及装置、分类存储方法、介质、电子设备 - Google Patents

人脸聚类方法及装置、分类存储方法、介质、电子设备 Download PDF

Info

Publication number
WO2022193232A1
WO2022193232A1 PCT/CN2021/081539 CN2021081539W WO2022193232A1 WO 2022193232 A1 WO2022193232 A1 WO 2022193232A1 CN 2021081539 W CN2021081539 W CN 2021081539W WO 2022193232 A1 WO2022193232 A1 WO 2022193232A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
face
category
images
clustered
Prior art date
Application number
PCT/CN2021/081539
Other languages
English (en)
French (fr)
Inventor
张文浩
刘瀚文
许景涛
Original Assignee
京东方科技集团股份有限公司
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 京东方科技集团股份有限公司 filed Critical 京东方科技集团股份有限公司
Priority to CN202180000510.7A priority Critical patent/CN115398485A/zh
Priority to PCT/CN2021/081539 priority patent/WO2022193232A1/zh
Priority to US17/773,123 priority patent/US20230245421A1/en
Publication of WO2022193232A1 publication Critical patent/WO2022193232A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/762Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/762Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
    • G06V10/763Non-hierarchical techniques, e.g. based on statistics of modelling distributions
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/761Proximity, similarity or dissimilarity measures
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/776Validation; Performance evaluation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/98Detection or correction of errors, e.g. by rescanning the pattern or by human intervention; Evaluation of the quality of the acquired patterns
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/30Scenes; Scene-specific elements in albums, collections or shared content, e.g. social network photos or video
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172Classification, e.g. identification

Definitions

  • the present disclosure relates to the technical field of image processing, and in particular, to a face clustering method and device, an image classification storage method, a computer-readable storage medium, and an electronic device.
  • the face clustering method can be used in many scenarios. For example, in the mobile phone album, the images of the same person can be grouped into a group through the image clustering method. The clustering accuracy of current face clustering methods is low.
  • the present disclosure aims to solve at least one of the technical problems existing in the prior art, and proposes a face clustering method and apparatus, an image classification storage method, a computer-readable storage medium, and an electronic device.
  • the present disclosure provides a face clustering method, which includes: acquiring a face image to be clustered; and performing clustering processing on the face image to be clustered, the clustering processing includes:
  • the current similarity threshold and the similarity it is judged whether there is an image of the same category of the face image to be clustered in the image category library; when there is the face image to be clustered in the image category library When the images of the same category are obtained, determine the category tags of the face images to be clustered according to the category tags of the images of the same category of the face images to be clustered; when there is no person to be clustered in the image category library When the image of the same category is the face image, a category label is assigned to the face image to be clustered according to the first preset rule.
  • the similarity threshold is positively related to the number level of image categories currently in the image category library.
  • the clustering process when there is no image of the same category of the face image to be clustered in the image category library, the clustering process further includes:
  • the face image to be clustered is added to the image category library.
  • acquiring the similarity threshold corresponding to the quantity level of the current image category in the image category library includes:
  • each category includes an image
  • the ith comparison step includes: calculating the similarity between a class of images whose comparison order is i in the image category library and the face image to be clustered When the similarity of a class of images whose comparison order is i in the image category library and the face image to be clustered is greater than or equal to the similarity threshold, determine the comparison order in the image category library A class of image i is the image of the same class of the face image to be clustered; when the similarity of a class of images whose comparison order is i in the image class library and the face image to be clustered is less than When the similarity threshold is set, the next comparison step is performed; wherein, i is an integer greater than zero, and i is less than or equal to the number of current image categories in the image category library.
  • calculating the similarity between a class of images whose comparison order is i in the image class library and the face image to be clustered includes:
  • the similarity between a class of images whose comparison order is i in the image class library and the face image to be clustered is determined.
  • the number of categories corresponding to the first quantity level is (0, 10000], and the similarity threshold corresponding to the first quantity level is between 0.3 and 0.6;
  • the number of categories corresponding to the second quantity level is (10000, 20000], and the similarity threshold corresponding to the second quantity level is between 0.61 and 0.64;
  • the number of categories corresponding to the third quantity level is (20000, 30000], and the similarity threshold corresponding to the third quantity level is between 0.65 and 0.664;
  • the number of categories corresponding to the fourth quantity level is (30000, 50000], and the similarity threshold corresponding to the fourth quantity level is between 0.665 and 0.70;
  • the number of categories corresponding to the fifth number level is (50000, ⁇ ), and the similarity threshold corresponding to the fifth number level is between 0.705 and 0.9.
  • the category label includes a category serial number
  • Assign category labels to the face images to be clustered according to the first preset rule including:
  • the face clustering method further includes: before acquiring the first face image to be clustered, establishing a correspondence between different quantity levels and similarity thresholds;
  • the process of establishing the correspondence between different quantity levels and similarity thresholds includes:
  • a plurality of test sets are set, and each of the test sets includes multiple types of test images; wherein, in different test sets, the number of categories of the test images is in different quantity levels;
  • the test images in the test set are clustered to obtain the test clustering results of the test images; and according to the test clustering results and the theoretical clustering results of each of the test images, determining the clustering error rate corresponding to the test threshold;
  • test threshold corresponding to the clustering error rate satisfying the preset test conditions is acquired, and the test threshold is used as the similarity threshold corresponding to the number level of the test image categories in the test set.
  • the clustering error rate includes: a false acceptance rate and a false rejection rate.
  • the preset test conditions include: the false rejection rate and the false recognition rate are equal.
  • the number of categories of the test images is the upper limit value of the number in the corresponding number level.
  • An embodiment of the present disclosure further provides an image classification storage method applied to an electronic device, the electronic device includes an image acquisition unit, and the image classification storage method includes:
  • the first face image is used as the face image to be clustered, and the first face image is processed by the above-mentioned face clustering method.
  • the clustering process is performed to obtain the category label of the first face image.
  • the image classification storage method further includes:
  • a class label of the first face image is output.
  • a batch of image sets are stored in the electronic device
  • the image classification storage method also includes:
  • the face images in the batch image set are sequentially used as the face images to be clustered, and the clustering process is performed.
  • the image classification storage method further includes:
  • the class labels are updated one by one for the face images that have completed the clustering process based on the current similarity threshold.
  • the step of updating the category label of any one of the face images that have completed the clustering process includes:
  • the current similarity threshold and the similarity between the face image with the label to be updated and at least one type of face image with the updated label determine whether the face image with the label to be updated exists in the face image with the updated label
  • determine the Describe the new category label of the face image of the label to be updated when there is no image of the same category of the face image of the label to be updated in the face image of the updated label, according to the second preset rule
  • the face images with updated labels are assigned new class labels;
  • the similarity threshold remains unchanged.
  • the category label includes a category serial number
  • the face images that have completed the clustering process are updated one by one according to the first order, and the first order is: the order that the face images that have completed the clustering process obtain the class labels;
  • An embodiment of the present disclosure further provides a face clustering apparatus, including: a memory and a processor, where a computer program is stored on the memory, wherein the computer program implements the above-mentioned face clustering when executed by the processor method.
  • Embodiments of the present disclosure further provide a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements the above-mentioned face clustering method or image classification storage method.
  • Embodiments of the present disclosure also provide an electronic device, including:
  • a memory and a processor where a computer program is stored on the memory, wherein the computer program implements the above-mentioned image classification storage method when executed by the processor.
  • FIG. 1 is a schematic diagram of clustering processing of face images to be clustered according to some embodiments of the present disclosure.
  • FIG. 2 is a schematic diagram of clustering processing performed on face images to be clustered according to other embodiments of the present disclosure.
  • FIG. 3 is a flowchart of establishing a correspondence between different quantity levels and similarity thresholds provided in an embodiment of the present disclosure.
  • FIG. 4 is a schematic diagram of an image classification storage method provided in some embodiments of the present disclosure.
  • FIG. 5 is a schematic diagram of an image classification storage method provided in other embodiments of the present disclosure.
  • FIG. 6 is a schematic diagram of an image classification storage method provided in other embodiments of the present disclosure.
  • FIG. 7 is a schematic diagram of an image clustering apparatus provided in some embodiments of the present disclosure.
  • FIG. 8 is a schematic diagram of an electronic device provided in some embodiments of the present disclosure.
  • the embodiment of the present disclosure provides a face clustering method, and the face clustering method can be executed based on a CPU.
  • the face clustering method includes performing clustering processing on the face images to be clustered, and the process of the clustering processing is the process of determining the category of the face images to be clustered.
  • FIG. 1 is a schematic diagram of clustering processing of face images to be clustered provided in some embodiments of the present disclosure. As shown in FIG. 1 , the clustering processing includes:
  • the similarity thresholds corresponding to at least two different quantity levels are positively correlated; exemplarily, the similarity thresholds corresponding to at least two adjacent quantity levels are positively correlated; exemplarily, the similarity thresholds It is positively related to the number rank of the current image categories in the image category library.
  • the number level is used to indicate the degree of the total number of image categories in the image category library. The higher the level, the greater the total number, or the total number of image categories corresponding to a higher level is greater than that of a lower level.
  • the total number of image categories For example, when the total number of image categories is between (0, 10000], the number level is the first level; when the total number of image categories is (10000, 20000], the number level is the second level; the total number of image categories is (20000, 30000], the quantity level is the third level; and so on.
  • one quantity level can correspond to the total number of one image category, and the total number of image categories can be a specific value (for example, 10000), It can also be a numerical range (for example, 10000 to 20000).
  • mapping relationship table may be obtained, and the mapping relationship table records: the quantity levels of the categories and their corresponding similarity thresholds.
  • the images in the image category library are all face images
  • the category of the image can represent the identity of the person
  • the images of different persons belong to different categories
  • the images of the same person belong to the same category.
  • step S02 obtain the similarity between the face image to be clustered and at least one type of image in the image category library; and determine whether there is a face image to be clustered in the image category library according to the current similarity threshold and the similarity
  • step S021 determines the person to be clustered according to the category label of the image of the same category of the face image to be clustered Class labels for face images.
  • step S022 is performed, that is, a category label is assigned to the face image to be clustered according to the first preset rule.
  • each type of image in the image category library may include one image or multiple images.
  • the similarity between each image in this type of image and the face image to be clustered can be acquired.
  • each category includes an image; when the similarity between the face image to be clustered and a certain category of images in the image category library is greater than the similarity threshold, the image of this category is the person to be clustered.
  • each category includes multiple images; when the similarity between the face image to be clustered and at least n images in the mth category in the image category library is greater than the similarity threshold, the Class images are images of the same class of face images to be clustered, for example, n is an integer, and is approximately half of the total number of images in the mth class.
  • the face images to be clustered may be images in a batch image set, for example, the batch image set includes batch images stored in the device; or, the batch image set includes batch images downloaded from the network.
  • the face image to be clustered may also be an image collected by an image collection unit such as a camera.
  • each image in the image category library has a category label, and the category label is used to indicate which category the image belongs to; the category label can be a category serial number or a category name.
  • the category label includes a category serial number.
  • the category labels of images of various categories in the image category library are: "00001", “00002", "00003", and so on.
  • the first preset rule is, for example, adding 1 to the number of current image categories in the image category library to obtain the category serial number (ie, category label) of the face image to be clustered.
  • the category label of the face images to be clustered is determined as "00010".
  • the similarity threshold is determined according to the quantity level of the current image categories in the image category library, which is beneficial to improve the image clustering efficiency. Accuracy. For example, if there are few current image categories in the image category library, set the similarity threshold to a smaller value, thereby reducing or preventing images of the same category from being identified as different categories; if the current image category in the image category library If there are more, the similarity threshold is set to a larger value, thereby reducing or preventing images of different categories from being identified as the same category. Therefore, adjusting the similarity threshold according to the number of categories of the current images in the image category library can improve the accuracy of image clustering.
  • FIG. 2 is a schematic diagram of clustering processing of face images to be clustered according to other embodiments of the present disclosure, and FIG. 2 is a specific implementation scheme of FIG. 1 .
  • each category in the image category library includes an image as an example for description.
  • the clustering process performed on each face image to be clustered includes:
  • this step S01 specifically includes:
  • step S01a determine whether the quantity level of the number of image categories in the image category library has changed; when the number level of the number of image categories in the image category library has not changed, go to step S01b, when the image category library Step S01c is performed when the number level of the number of image categories in the image category is changed.
  • the number of categories corresponding to the first quantity level is (0, 10000], and the similarity threshold corresponding to the first quantity level is between 0.3 and 0.6, for example, 0.36 or 0.37 or 0.40.
  • the first The quantity level is divided into two sub-levels, denoted as the first sub-level and the second sub-level, the number of categories corresponding to the first sub-level is (0, 5000], and the similarity threshold corresponding to the first sub-level is, for example, 0.36;
  • the number of categories corresponding to the sublevel is (5000, 10000]; the similarity threshold corresponding to the second sublevel is, for example, 0.59.
  • the number of categories corresponding to the second quantity level is (10000, 20000], and the similarity threshold corresponding to the second quantity level is between 0.61 and 0.64.
  • the similarity threshold corresponding to the second quantity level is Between 0.61 and 0.63, or between 0.61 and 0.62; for example, the similarity threshold corresponding to the second quantity level is 0.62.
  • the number of categories corresponding to the third number level is (20000, 30000], and the similarity threshold corresponding to the third number level is between 0.65 and 0.664.
  • the similarity threshold corresponding to the third number level is Between 0.655 and 0.659, or between 0.66 and 0.662.
  • the similarity threshold corresponding to the third quantity level is 0.66.
  • the number of categories corresponding to the fourth number level is (30000, 50000], and the similarity threshold corresponding to the fourth number level is between 0.665 and 0.70.
  • the similarity threshold corresponding to the fourth number level is Between 0.665 and 0.68, for example, the similarity threshold corresponding to the fourth quantity level is 0.67.
  • the number of categories corresponding to the fifth number level is (50000, ⁇ ), and the similarity threshold corresponding to the fifth number level is between 0.705 and 0.9.
  • the similarity threshold corresponding to the fifth number level is between 0.71 and 0.75, for example, the similarity threshold corresponding to the fifth number level is 0.72.
  • step S02 obtain the similarity between the face image to be clustered and at least one type of image in the image category library; and according to the current similarity threshold and the similarity, determine whether there is a similarity between the face image to be clustered in the image category library Category images; when there are images of the same category of the face images to be clustered in the image category library, go to step S021, that is, determine the person to be clustered according to the category labels of the images of the same category of the face images to be clustered Class labels for face images.
  • step S022 is performed, that is, a category label is assigned to the face image to be clustered according to the first preset rule.
  • step S02 specifically includes:
  • S02a Determine the comparison order of various types of images in the image category library and the face images to be clustered.
  • the comparison order can be determined according to the category label of each category of images.
  • the category label includes a category serial number.
  • the order of comparison between each type of image and the face image to be clustered can be determined according to the category serial number of each type of image.
  • the ith comparison step comprises: calculating the similarity of a class of images whose comparison order is i in the image category library and the face image to be clustered, when the image
  • the similarity between a class of images whose comparison order is i in the category library and the face image to be clustered is greater than or equal to the similarity threshold, it is determined that a class of images whose comparison order is i in the image category library is to be clustered.
  • i is an integer greater than zero, and i is less than or equal to the number of current image categories in the image category library.
  • Category image the similarity between the face image to be clustered and a class of images with category serial number "00002" is less than the similarity threshold, and the similarity between the face image to be clustered and a class of images with category serial number "00003" continues to be calculated , and so on, until it is determined that an image of the same category of the face image to be clustered is determined, or it is determined that there is no image of the same category of the face image to be clustered in the image category library.
  • calculating the similarity between a class of images whose comparison order is i in the image category library and the face image to be clustered specifically includes: obtaining the comparison order in the image category library is i The feature vector of a class of images, as the first feature vector. The feature vector of the face image to be clustered is acquired as the second feature vector. Then, according to the similarity between the first feature vector and the second feature vector, determine the similarity between a class of images whose comparison order is i in the image class library and the face image to be clustered.
  • the similarity between the first feature vector and the second feature vector may specifically be cosine similarity.
  • step S021 specifically includes: setting the category label of the face image to be clustered to be the same as the category label of the image of the same category.
  • step S022 specifically includes: adding 1 to the number of current image categories in the image category library to obtain a category serial number (ie, category label) of the face image to be clustered.
  • the clustering processing of the images to be clustered further includes: when there is no image of the same category of the face image to be clustered in the image category library, performing step S023: adding the face image to be clustered to the image in the category library.
  • the sequence of steps S022 and S023 is not particularly limited.
  • the face images to be clustered are not added to the image category library. In this way, when the image categories in the image category library gradually increase , there is always only one face image in each category.
  • the face clustering method further includes: before acquiring the first face image to be clustered, establishing a correspondence between different quantity levels and similarity thresholds.
  • FIG. 3 is a flowchart of establishing the correspondence between different quantity levels and similarity thresholds provided in the embodiment of the present disclosure. As shown in FIG. 3 , the process of establishing the correspondence between different quantity levels and similarity thresholds includes:
  • each test set includes multiple types of test images; wherein, in different test sets, the number of categories of test images is at different levels.
  • the number of categories of test images is the upper limit value of the number in the corresponding number level.
  • the number of categories corresponding to the first quantity level is (0, 10000]; the number of categories corresponding to the second level of quantity is (10000, 20000]; the number of categories corresponding to the third level of quantity is (20000, 30000]; The number of categories corresponding to the fourth level of quantity is (30000, 50000]; the number of categories corresponding to the fifth level of number is (50000, ⁇ ).
  • the rest of the The number of categories of test images in each test set are: 10,000, 20,000, 30,000, and 50,000; in the test set corresponding to the fifth level, the number of categories of test images can be between 50,000 and 60,000.
  • FIG. 4 is a schematic diagram of obtaining the similarity threshold corresponding to the test set provided in some embodiments of the present disclosure. As shown in FIG. 4 , the step of obtaining the similarity threshold corresponding to the test set includes:
  • S002a set at least one test threshold.
  • the step of clustering the a-th test image in the test set includes: calculating the similarity between the a-th test image and each of the remaining test images, when the a-th test image and the b-th test image are similar
  • the similarity of the images is greater than or equal to the current test threshold, it is determined that the a-th test image is successfully matched with the b-th test image (that is, it belongs to the same class as the b-th test image);
  • the similarity of the b test images is less than the current test threshold, it is determined that the a-th test image fails to match with the b-th test image (ie, it does not belong to the same class as the b-th test image).
  • the clustering error rate is a parameter used to express the degree of clustering error.
  • test clustering result of the test image refers to whether the test image is successfully matched with each of the other test images after the similarity between the test image and each of the other test images is calculated.
  • the theoretical clustering result of the test image refers to whether the test image and each other test image should match successfully in theory.
  • the theoretical class label of each test image can be preset to determine the theoretical clustering of each test image. class result. Understandably, in theory, test images with the same class label should match each other successfully.
  • the clustering error rates include: False Acceptance Rate (FAR) and False Rejection Rate (FRR).
  • False rejection rate The proportion of test images that should be of the same class are mistakenly considered to be of a different class.
  • NFA is the sum of the number of "false acceptances" for all test images.
  • test images A to F there are three categories of test images, which are respectively denoted as test images A to F.
  • the category labels of test images A to B are all "01"
  • the category labels of test images C to D are all "02”.
  • the class labels of test images E to F are all "03".
  • test image A and test images B and C are all greater than the test threshold, and the similarity thresholds of test image A and test images D to F are less than test threshold, it indicates that test image A has a "false acceptance" (because test images A and C are mistakenly considered to be one class).
  • NFR is the sum of the number of "false rejections" for all test images.
  • the category labels of test images A to B are all "01", and the category labels of test images C to D are all "02".
  • the class labels of test images E to F are all "03".
  • NGRA is the total number of matches within a class.
  • each class includes 8 test images.
  • each class includes 8 test images.
  • S002c Obtain a test threshold corresponding to a clustering error rate that satisfies a preset test condition, and use the test threshold as a similarity threshold corresponding to the number level of the image test sample categories in the test set.
  • the preset test conditions include equal false rejection rates and false acceptance rates.
  • Embodiments of the present disclosure also provide an image classification and storage method applied to an electronic device, where the electronic device includes an image acquisition unit.
  • the electronic device may be a terminal such as a mobile phone or a computer; for another example, the electronic device is an access control device with an image acquisition unit.
  • FIG. 5 is a schematic diagram of an image classification storage method provided in some embodiments of the present disclosure. As shown in FIG. 5 , the image classification storage method includes:
  • the image acquisition unit may perform video acquisition in real time, and the first image is a face image acquired by the image acquisition unit from the video. Each time a first image is acquired, the first image is used as the face image to be clustered, and the above steps S01 and S02 are performed.
  • FIG. 5 is a schematic diagram of an image classification storage method provided in some embodiments of the present disclosure. As shown in FIG. 5 , the image classification storage method includes:
  • FIG. 6 is a schematic diagram of a method for classifying and storing images provided in other embodiments of the present disclosure. As shown in FIG. 6 , before the first first image is collected by the image collection unit, a batch of image sets are stored in the electronic device .
  • the image classification storage method includes:
  • the images in the batch image set are sequentially used as the face images to be clustered, and clustering processing is performed.
  • the initialized image category library includes images of multiple categories.
  • the first image in the batch image set is clustered, there is no image in the image category library, and there is no image of the same category as the image to be clustered. Therefore, it is the first image in the batch image set.
  • the first image in the batch image set is added to the image class library.
  • images of the same category can also be saved in the same folder, and images of different categories can be saved in different folders.
  • the label update step includes: updating the category labels one by one for the face images that have completed the clustering process based on the current similarity threshold.
  • the face image that has completed the clustering process refers to the face image for which the category label has been obtained.
  • the face images that have been clustered include: face images in the image category library, and also include: face images that have been clustered but not added to the image category library.
  • the tag update signal may be a signal generated by the electronic device under set conditions. For example, when the number of first face images collected by the image acquisition unit reaches a certain number, the electronic device generates a tag update signal; for another example, when the physical time reaches a preset time (for example, 2:00 am every Monday), the electronic device Generates a label update signal.
  • step S13 and step S11 in the embodiment of the present disclosure is not limited.
  • step S11 is continued.
  • the category label is updated based on the current similarity threshold and the similarity between the face image whose label is to be updated and at least one type of face image whose label has been updated.
  • the current similarity threshold is the similarity threshold corresponding to the quantity level of the current image categories in the image category library.
  • the electronic device may generate a label update signal at multiple times, and each time a label update signal is generated, a label update step is performed. During the same label update step, the similarity threshold remains unchanged.
  • the step of updating the category label of any face image that has completed the clustering process includes: according to the current similarity threshold, and at least one category between the face image to be updated with the label and the updated label
  • the similarity of the face images determines whether there is an image of the same category of the face image whose label is to be updated in the face image whose label has been updated.
  • Category label when there is no image of the same category as the face image with the label to be updated in the face image whose label has been updated, assign a new category label to the face image with the label to be updated according to the second preset rule.
  • the face image is the "face image whose label is to be updated”.
  • the face image of the updated label when the similarity between the face image of the updated label and the face image of the label to be updated is greater than or equal to the current similarity threshold, the face image of the updated label can be used as the face of the label to be updated. Image of the same category.
  • the face images that have completed the clustering process are updated with the category labels one by one according to a first order, wherein the first order is: the order in which the face images that have completed the clustering process obtain the category labels.
  • Category labels include category numbers.
  • assigning a new category label to the face image whose label is to be updated according to the second preset rule which specifically includes: adding 1 to the total number of categories of the face image whose label has been updated, to obtain the face whose label is to be updated. New category labels for images.
  • the label update step performed in response to the label update signal can be viewed as the following process:
  • An updated image library is established, wherein at the initial moment of receiving the label update signal, the updated image library is an empty set; after that, the category labels are updated one by one for the face images that have completed the clustering process according to the above first order.
  • the process of updating the class label of any one of the face images that have completed the clustering process includes: according to the current similarity threshold and the similarity between the face image of the label to be updated and at least one class of images in the updated image library, Determine whether there is an image of the same category of the face image whose label is to be updated in the updated image library; if there is, use the category label of the same category of image as the new category label of the face image to be updated with the label; otherwise, update the image library Add 1 to the total number of categories of face images in , to get the new category label of the face image whose label is to be updated.
  • the updated image library is used to replace the image category library.
  • the clustering accuracy of each image can be improved by updating the category labels of the face images that have been clustered.
  • FIG. 7 is a schematic diagram of an image clustering apparatus provided in some embodiments of the present disclosure.
  • the image clustering apparatus includes a memory 11 and a processor 12, and a computer program is stored on the memory 11, wherein the computer When the program is executed by the processor 12, the face clustering method in the above-mentioned embodiment.
  • the present disclosure also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements the above-mentioned face clustering method or image classification storage method.
  • FIG. 8 is a schematic diagram of an electronic device provided in some embodiments of the present disclosure.
  • the electronic device includes: an image capture unit 30 , a memory 40 and a processor 50 , and the image capture unit 30 includes a camera and the like for image capturing collected device.
  • a computer program is stored in the memory 40 , wherein the computer program implements the above-mentioned image classification storage method when executed by the processor 50 .
  • the above-mentioned memory and the computer-readable storage medium include, but are not limited to, the following readable media: such as random access memory (RAM), read only memory (ROM), non-volatile random access memory (NVRAM), programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable PROM (EEPROM), Flash Memory, Magnetic or Optical Data Storage, Registers, Disk or Tape, such as Compact Disc (CD) or DVD (Digital Universal Disc) and other non-transitory media.
  • processors include, but are not limited to, general purpose processors, central processing units (CPUs), microprocessors, digital signal processors (DSPs), controllers, microcontrollers, state machines, and the like.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Computing Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Human Computer Interaction (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

公开一种人脸聚类方法及装置、图像分类存储方法、计算机可读存储介质、电子设备,人脸聚类方法包括:对待聚类的图像进行聚类处理,所述聚类处理包括:获取与图像类别库中当前的图像类别的数量等级所对应的相似度阈值;至少两个数量等级所对应的相似度阈值不同;根据当前的相似度阈值以及待聚类人脸图像与所述图像类别库中至少一类图像的相似度,判断图像类别库中是否存在待聚类人脸图像的同类别图像;当存在待聚类人脸图像的同类别图像时,根据待聚类人脸图像的同类别图像的类别标签确定待聚类人脸图像的类别标签;当不存在待聚类人脸图像的同类别图像时,按照第一预设规则为待聚类人脸图像分配类别标签。

Description

人脸聚类方法及装置、分类存储方法、介质、电子设备 技术领域
本公开涉及图像处理技术领域,具体涉及一种人脸聚类方法及装置、图像分类存储方法、计算机可读存储介质、电子设备。
背景技术
人脸聚类方法可以用到很多场景中,例如,在手机相册中,通过图像聚类的方法可以把同一人的图像聚到一个组里面。目前人脸聚类方法的聚类准确率较低。
发明内容
本公开旨在至少解决现有技术中存在的技术问题之一,提出了一种人脸聚类方法及装置、图像分类存储方法、计算机可读存储介质、电子设备。
为了实现上述目的,本公开提供一种人脸聚类方法,包括:获取待聚类人脸图像;对所述待聚类人脸图像进行聚类处理,所述聚类处理包括:
获取与图像类别库中当前的图像类别的数量等级所对应的相似度阈值;其中,至少两个不同的数量等级所对应的相似度阈值不同;
获取待聚类人脸图像与所述图像类别库中至少一类图像的相似度;
根据当前的相似度阈值以及所述相似度,判断所述图像类别库中是否存在所述待聚类人脸图像的同类别图像;当所述图像类别库中存在所述待聚类人脸图像的同类别图像时,根据所述待聚类人脸图像的同类别图像的类别标签确定所述待聚类人脸图像的类别标签;当所述图像类别库中不存在所述待聚类人脸图像的同类别图像时,按照第一预设规则为所述待聚类人脸图像分配类别标签。
在一些实施例中,所述相似度阈值与所述图像类别库中当前的图 像类别的数量等级正相关。
在一些实施例中,当所述图像类别库中不存在所述待聚类人脸图像的同类别图像时,所述聚类处理还包括:
将所述待聚类人脸图像加入所述图像类别库中。
在一些实施例中,获取与图像类别库中当前的图像类别的数量等级所对应的相似度阈值,包括:
判断所述图像类别库中的图像类别的数量所处的数量等级是否发生变化;当所述图像类别库中的图像类别的数量所处的数量等级未发生变化时,采用上一张待聚类人脸图像进行聚类处理时所用的相似度阈值;当所述图像类别库中的图像类别的数量所处的数量等级发生变化时,根据变化后的数量等级更新所述相似度阈值。
在一些实施例中,在所述图像类别库中,每个类别包括一张图像,
获取待聚类人脸图像与所述图像类别库中至少一类图像的相似度;根据当前的相似度阈值以及所述相似度,判断所述图像类别库中是否存在所述待聚类人脸图像的同类别图像,具体包括:
确定所述图像类别库中的各类图像与所述待聚类人脸图像的比对次序;
根据所述比对次序进行至少一次比对步骤,其中,第i次比对步骤包括:计算所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度,当所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度大于或等于所述相似度阈值时,确定所述图像类别库中比对次序为i的一类图像为所述待聚类人脸图像的同类别图像;当所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度小于所述相似度阈值时,进行下一次比对步骤;其中,i为大于零的整数,且i小于或等于图像类别库中当前的图像类别的数量。
在一些实施例中,计算所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度,包括:
获取所述图像类别库中比对次序为i的一类图像的特征向量,作为第一特征向量;
获取所述待聚类人脸图像的特征向量,作为第二特征向量;
根据所述第一特征向量与所述第二特征向量的相似度,确定所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度。
在一些实施例中,第1数量等级所对应的类别数量为(0,10000],第1数量等级所对应的相似度阈值在0.3~0.6之间;
第2数量等级所对应的类别数量为(10000,20000],第2数量等级所对应的相似度阈值在0.61~0.64之间;
第3数量等级所对应的类别数量为(20000,30000],第3数量等级所对应的相似度阈值在0.65~0.664之间;
第4数量等级所对应的类别数量为(30000,50000],第4数量等级所对应的相似度阈值在0.665~0.70之间;
第5数量等级所对应的类别数量为(50000,∞),第5数量等级所对应的相似度阈值在0.705~0.9之间。
在一些实施例中,所述类别标签包括类别序号;
按照第一预设规则为所述待聚类人脸图像分配类别标签,包括:
将所述图像类别库中当前的图像类别的数量加1,得到所述待聚类人脸图像的类别序号。
在一些实施例中,所述人脸聚类方法还包括:在获取第一张待聚类人脸图像之前,建立不同数量等级与相似度阈值的对应关系;
其中,建立不同数量等级与相似度阈值的对应关系的过程包括:
设置多个测试集,每个所述测试集包括多类测试图像;其中,在 不同的测试集中,所述测试图像的类别数量处于不同的数量等级;
对于每个所述测试集,执行以下步骤:
设置至少一个测试阈值;
基于每个所述测试阈值,对所述测试集中的测试图像进行聚类,得到所述测试图像的测试聚类结果;并根据每个所述测试图像的测试聚类结果和理论聚类结果,确定所述测试阈值对应的聚类错误率;
获取满足预设测试条件的聚类错误率所对应的测试阈值,并将该测试阈值作为所述测试集中测试图像的类别的数量等级所对应的相似度阈值。
在一些实施例中,所述聚类错误率包括:错误接受率和错误拒绝率。
在一些实施例中,所述预设测试条件包括:所述错误拒绝率和所述错误识别率相等。
在一些实施例中,除了最高数量等级所对应的测试集之外,其余每个所述测试集中,所述测试图像的类别数量为相应的数量等级中的数量上限值。
本公开实施例还提供一种应用于电子设备的图像分类存储方法,所述电子设备包括图像采集单元,所述图像分类存储方法包括:
响应于所述图像采集单元采集到的第一人脸图像,将所述第一人脸图像作为待聚类人脸图像,并利用上述的人脸聚类方法对所述第一人脸图像进行所述聚类处理,以获得所述第一人脸图像的类别标签。
在一些实施例中,所述图像分类存储方法还包括:
输出所述第一人脸图像的类别标签。
在一些实施例中,在所述图像采集单元采集到第一张第一人脸图像之前,所述电子设备中存储有批量图像集;
所述图像分类存储方法还包括:
在所述图像采集单元采集到第一张所述第一人脸图像之前,将所述批量图像集中的人脸图像依次作为所述待聚类人脸图像,进行所述聚类处理。
在一些实施例中,所述图像分类存储方法还包括:
响应于标签更新信号,基于当前的相似度阈值对已完成聚类处理的人脸图像逐个进行类别标签的更新。
在一些实施例中,对任意一个已完成聚类处理的人脸图像的类别标签进行更新的步骤包括:
根据当前的相似度阈值、以及待更新标签的人脸图像与已更新标签的至少一类人脸图像的相似度,判断已更新标签的人脸图像中是否存在所述待更新标签的人脸图像的同类别图像;当已更新标签的人脸图像中存在所述待更新标签的人脸图像的同类别图像时,根据所述待更新标签的人脸图像的同类别图像的类别标签,确定所述待更新标签的人脸图像的新的类别标签;当已更新标签的人脸图像中不存在所述待更新标签的人脸图像的同类别图像时,按照第二预设规则为所述待更新标签的人脸图像分配新的类别标签;
其中,对已完成聚类处理的人脸图像逐个进行类别标签的更新的过程中,所述相似度阈值保持不变。
在一些实施例中,所述类别标签包括类别序号;
已完成聚类处理的人脸图像按照第一次序逐个进行类别标签的更新,所述第一次序为:已完成聚类处理的人脸图像获得类别标签的次序;
按照第二预设规则为所述待更新标签的人脸图像分配新的类别标签,包括:
将当前已更新标签的人脸图像的类别总数加1,得到所述待更新标签的人脸图像的新的类别标签。
本公开实施例还提供一种人脸聚类装置,包括:存储器和处理器,所述存储器上存储有计算机程序,其中,所述计算机程序被所述处理器执行时实现上述的人脸聚类方法。
本公开实施例还提供一种计算机可读存储介质,其上存储有计算机程序,该计算机程序被处理器执行时实现上述的人脸聚类方法或图像分类存储方法。
本公开实施例还提供一种电子设备,包括:
图像采集单元;
存储器和处理器,所述存储器上存储有计算机程序,其中,所述计算机程序被所述处理器执行时实现上述的图像分类存储方法。
附图说明
附图是用来提供对本公开的进一步理解,并且构成说明书的一部分,与下面的具体实施方式一起用于解释本公开,但并不构成对本公开的限制。在附图中:
图1为本公开的一些实施例中提供的对待聚类人脸图像进行的聚类处理的示意图。
图2为本公开的另一些实施例中提供的对待聚类人脸图像进行的聚类处理的示意图。
图3为本公开实施例中提供的建立不同数量等级与相似度阈值的对应关系的流程图。
图4为本公开的一些实施例中提供的图像分类存储方法的示意图。
图5为本公开的另一些实施例中提供的图像分类存储方法的示意图。
图6为本公开的另一些实施例中提供的图像分类存储方法的示意图。
图7为本公开的一些实施例中提供的图像聚类装置的示意图。
图8为本公开的一些实施例中提供的电子设备的示意图。
具体实施方式
为使本公开实施例的目的、技术方案和优点更加清楚,下面将结合本公开实施例的附图,对本公开实施例的技术方案进行清楚、完整地描述。显然,所描述的实施例是本公开的一部分实施例,而不是全部的实施例。基于所描述的本公开的实施例,本领域普通技术人员在无需创造性劳动的前提下所获得的所有其他实施例,都属于本公开保护的范围。
本公开实施例提供一种人脸聚类方法,该人脸聚类方法可以基于CPU执行。所述人脸聚类方法包括对待聚类人脸图像进行聚类处理,该聚类处理的过程即为确定待聚类人脸图像的类别的过程。图1为本公开的一些实施例中提供的对待聚类人脸图像进行的聚类处理的示意图,如图1所示,该聚类处理包括:
S01、获取与图像类别库中当前的图像类别的数量等级所对应的相似度阈值;其中,至少两个不同的数量等级所对应的相似度阈值是不同的。
示例性地,至少两个不同的数量等级所对应的相似度阈值是正相关的;示例性地,至少两个相邻的数量等级所对应的相似度阈值是正相关的;示例性地,相似度阈值与图像类别库中当前的图像类别的数量等级正相关。
需要说明的是,数量等级用于表示图像类别库中的图像类别总数量的多少程度,等级越高,表示总数量越多,或者说较高等级对应的图像类别总数量大于较低等级对应的图像类别总数量。例如,图像类别的总数量在(0,10000]之间时,数量等级为第1等级;图像类别的总数量在(10000,20000]时,数量等级为第2等级;图像类别的总数量在(20000,30000]时,数量等级为第3等级;依次类推。需要说明 的是,一个数量等级可以对应1个图像类别的总数量,图像类别的总数量可以是一个具体数值(例如10000),也可以是一个数值范围(例如10000~20000)。
还需要说明的是,在步骤S01之前,可以获取映射关系表,该映射关系表中记载有:类别的数量等级与其对应的相似度阈值。
在一些实施例中,图像类别库中的图像均为人脸图像,图像的类别可以表示人的身份,不同人的图像属于不同的类别,同一个人的图像属于同一类别。
S02、获取待聚类人脸图像与所述图像类别库中至少一类图像的相似度;并根据当前的相似度阈值以及所述相似度,判断图像类别库中是否存在待聚类人脸图像的同类别图像;当图像类别库中存在待聚类人脸图像的同类别图像时,进行步骤S021,即,根据待聚类人脸图像的同类别图像的类别标签确定所述待聚类人脸图像的类别标签。当图像类别库中不存在待聚类人脸图像的同类别图像时,进行步骤S022,即,按照第一预设规则为待聚类人脸图像分配类别标签。
其中,图像类别库中的每一类图像中可以包括一张图像,也可以包括多张图像。在获取待聚类人脸图像与所述图像类别库中某一类图像的相似度时,可以获取该类图像中的每一张图像与待聚类人脸图像的相似度。
例如,图像类别库中,每个类别包括一张图像;当待聚类人脸图像与图像类别库中的某一类图像的相似度大于相似度阈值时,则该类图像为待聚类人脸图像的同类别图像。又例如,图像类别库中,每个类别包括多张图像;当待聚类人脸图像与图像类别库中的第m类中的至少n张图像的相似度均大于相似度阈值时,则该类图像为待聚类人脸图像的同类别图像,例如,n为整数,且近似为第m类中图像总数的一半。
其中,待聚类人脸图像可以是批量图像集中的图像,例如,批量图像集包括设备中所存储的批量图像;或者,批量图像集包括从网络上下载的批量图像。当然,待聚类人脸图像也可以为摄像头等图像采集单元所采集的图像。
其中,图像类别库中每张图像具有类别标签,该类别标签用于表示图像属于哪一类;类别标签可以为类别序号,也可以为类别的名称。在一个示例中,类别标签包括类别序号,例如,图像类别库中多种类别的图像的类别标签分别为:“00001”、“00002”、“00003”等。此时,第一预设规则例如为:将所述图像类别库中当前的图像类别的数量加1,得到待聚类人脸图像的类别序号(即,类别标签)。示例性地,对待聚类人脸图像进行聚类处理时,若图像类别库中的已存在9个图像类别,则将待聚类人脸图像的类别标签确定为“00010”。
在本公开实施例中,在对每个待聚类人脸图像进行聚类时,是根据图像类别库中当前的图像类别的数量等级来确定相似度阈值的,从而有利于提高图像聚类的准确率。例如,若图像类别库中当前的图像类别较少,则将相似度阈值设置为较小值,从而减少或防止相同类别的图像被识别成不同类别的情况;若图像类别库中当前的图像类别较多,则将相似度阈值设置为较大值,从而减少或防止不同类别的图像被识别为相同类别的情况。因此,根据图像类别库中当前的图像的类别数量来调整相似度阈值,可以提高图像聚类的准确率。
图2为本公开的另一些实施例中提供的对待聚类人脸图像进行的聚类处理的示意图,图2为图1的一种具体化实现方案。其中,以图像类别库中的每个类别包括一张图像为例进行说明。如图2所示,对每张待聚类人脸图像进行的聚类处理包括:
S01、获取与图像类别库中当前的图像类别的数量等级所对应的相似度阈值;其中,相似度阈值与图像类别库中当前的图像类别的数量 等级正相关。
其中,该步骤S01具体包括:
S01a、判断图像类别库中的图像类别的数量所处的数量等级是否发生变化;当图像类别库中的图像类别的数量所处的数量等级未发生变化时,进行步骤S01b,当图像类别库中的图像类别的数量所处的数量等级发生变化时,进行步骤S01c。
S01b、采用上一张待聚类人脸图像进行聚类处理时所用的相似度阈值。
S01c、根据变化后的数量等级更新相似度阈值。
示例性地,第1数量等级所对应的类别数量为(0,10000],第1数量等级所对应的相似度阈值在0.3~0.6之间,例如为0.36或0.37或0.40。又例如,第1数量等级分为两个子等级,记作第一子等级和第二子等级,第一子等级对应的类别数量为(0,5000],第一子等级对应的相似度阈值例如为0.36;第二子等级对应的类别数量为(5000,10000];第二子等级对应的相似度阈值例如为0.59。
示例性地,第2数量等级所对应的类别数量为(10000,20000],第2数量等级所对应的相似度阈值在0.61~0.64之间,例如,第2数量等级所对应的相似度阈值在0.61~0.63之间,或者在0.61~0.62之间;例如,第2数量等级所对应的相似度阈值为0.62。
示例性地,第3数量等级所对应的类别数量为(20000,30000],第3数量等级所对应的相似度阈值在0.65~0.664之间。例如,第3数量等级所对应的相似度阈值在0.655~0.659之间,或者在0.66~0.662之间。例如,第3数量等级所对应的相似度阈值为0.66。
示例性地,第4数量等级所对应的类别数量为(30000,50000],第4数量等级所对应的相似度阈值在0.665~0.70之间。例如,第4数量等级所对应的相似度阈值在0.665~0.68之间,例如,第4数量等级 所对应的相似度阈值为0.67。
示例性地,第5数量等级所对应的类别数量为(50000,∞),第5数量等级所对应的相似度阈值在0.705~0.9之间。例如,第5数量等级所对应的相似度阈值在0.71~0.75之间,例如,第5数量等级所对应的相似度阈值为0.72。
S02、获取待聚类人脸图像与图像类别库中至少一类图像的相似度;并根据当前的相似度阈值以及所述相似度,判断图像类别库中是否存在待聚类人脸图像的同类别图像;当所述图像类别库中存在所述待聚类人脸图像的同类别图像时,进行步骤S021,即,根据待聚类人脸图像的同类别图像的类别标签确定待聚类人脸图像的类别标签。当所述图像类别库中不存在所述待聚类人脸图像的同类别图像时,进行步骤S022,即,按照第一预设规则为所述待聚类人脸图像分配类别标签。
可选地,步骤S02具体包括:
S02a、确定图像类别库中的各类图像与待聚类人脸图像的比对次序。
可选地,对比次序可以根据每类图像的类别标签确定。如上文所示,类别标签包括类别序号,这种情况下,每类图像与待聚类人脸图像的对比次序可以根据每类图像的类别序号确定。
S02b、根据比对次序进行至少一次比对步骤,其中,第i次比对步骤包括:计算图像类别库中比对次序为i的一类图像与待聚类人脸图像的相似度,当图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度大于或等于相似度阈值时,确定图像类别库中比对次序为i的一类图像为待聚类人脸图像的同类别图像;当图像类别库中比对次序为i的一类图像与待聚类人脸图像的相似度小于所述相似度阈值时,进行下一次比对步骤。
i为大于零的整数,且i小于或等于图像类别库中当前的图像类别 的数量。
例如,计算待聚类人脸图像与类别序号为“00001”的一类图像的相似度,若待聚类人脸图像与类别序号为“00001”的一类图像的相似度大于或等于相似度阈值,则确定类别序号为“00001”的一类图像为待聚类人脸图像的同类别图像;否则,继续计算待聚类人脸图像与类别序号为“00002”的一类图像的相似度,若待聚类人脸图像与类别序号为“00002”的一类图像的相似度大于或等于相似度阈值,则确定类别序号为“00002”的一类图像为待聚类人脸图像的同类别图像;待聚类人脸图像与类别序号为“00002”的一类图像的相似度小于相似度阈值,继续计算待聚类人脸图像与类别序号为“00003”的一类图像的相似度,以此类推,直至确定出待聚类人脸图像的同类别图像,或者确定出图像类别库中不存在待聚类人脸图像的同类别图像。
其中,第i次比对步骤中,计算图像类别库中比对次序为i的一类图像与待聚类人脸图像的相似度,具体包括:获取所述图像类别库中比对次序为i的一类图像的特征向量,作为第一特征向量。获取所述待聚类人脸图像的特征向量,作为第二特征向量。之后,根据所述第一特征向量与所述第二特征向量的相似度,确定所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度。
其中,第一特征向量与第二特征向量的相似度具体可以为余弦相似度。
在一些实施例中,步骤S021具体包括:将待聚类人脸图像的类别标签设置为与其同类别图像的类别标签相同。
在一些实施例中,步骤S022具体包括:将所述图像类别库中当前的图像类别的数量加1,得到待聚类人脸图像的类别序号(即,类别标签)。
在一些实施例中,对待聚类图像进行的聚类处理还包括:当图像 类别库中不存在待聚类人脸图像的同类别图像时,执行步骤S023:将待聚类人脸图像加入图像类别库中。其中,步骤S022与S023的先后顺序不作特别限定。
在一些示例中,当图像类别库中存在待聚类人脸图像的同类别图像时,不将待聚类人脸图像加入图像类别库,这样,当图像类别库中的图像类别逐渐增加的过程中,每个类别中始终只有一张人脸图像。
在一些实施例中,所述人脸聚类方法还包括:在获取第一张待聚类人脸图像之前,建立不同数量等级与相似度阈值的对应关系。图3为本公开实施例中提供的建立不同数量等级与相似度阈值的对应关系的流程图,如图3所示,建立不同数量等级与相似度阈值的对应关系的过程包括:
S001、设置多个测试集,每个测试集包括多类测试图像;其中,在不同的测试集中,测试图像的类别数量处于不同的数量等级。
在一些实施例中,除了最高数量等级所对应的测试集之外,其余每个测试集中,测试图像的类别数量为相应的数量等级中的数量上限值。
例如,第1数量等级所对应的类别数量为(0,10000];第2数量等级所对应的类别数量为(10000,20000];第3数量等级所对应的类别数量为(20000,30000];第4数量等级所对应的类别数量为(30000,50000];第5数量等级所对应的类别数量为(50000,∞)。相应地,除了第5数量等级所对应的测试集之外,其余每个测试集中的测试图像的类别数量分别为:10000、20000、30000、50000;第5数量等级所对应的测试集中,测试图像的类别数量可以在50000~60000之间。
S002、对于每个测试集,获取与该测试集中的测试图像的类别的数量等级相对应的相似度阈值,从而得到每个数量等级对应的相似度阈值。
图4为本公开的一些实施例中提供的获取与测试集相对应的相似度阈值的示意图,如图4所示,获取与测试集相对应的相似度阈值的步骤包括:
S002a、设置至少一个测试阈值。
S002b、基于每个测试阈值,对测试集中的测试图像进行聚类,得到测试图像的测试聚类结果;并根据每个测试图像的测试聚类结果和理论聚类结果,确定测试阈值对应的聚类错误率。
在一些实施例中,对测试集中的第a个测试图像进行聚类的步骤包括:将第a个测试图像与其余每个测试图像进行相似度计算,当第a个测试图像与第b个测试图像的相似度大于或等于当前的测试阈值时,则确定第a个测试图像与第b个测试图像匹配成功(即,与第b个测试图像属于同一类);当第a个测试图像与第b个测试图像的相似度小于当前的测试阈值时,则确定第a个测试图像与第b个测试图像匹配失败(即,与第b个测试图像不属于同一类)。
其中,聚类错误率是用于表示聚类错误程度的参数。
可以理解的是,测试图像的测试聚类结果是指,将该测试图像与其余每个测试图像进行相似度计算后,该测试图像与其余每个测试图像是否匹配成功。测试图像的理论聚类结果是指,该测试图像与其余的每张测试图像在理论上是否应当匹配成功,可以预先设置每个测试图像理论上的类别标签,从而确定每个测试图像的理论聚类结果。可以理解的是,理论上,具有相同类别标签的测试图像应当彼此匹配成功。
在一些实施例中,聚类错误率包括:错误接受率(FAR)和错误拒绝率(FRR)。
其中,错误接受率为,将本应该是不同类别的测试图像错误地认为是相同类别的比例。错误拒绝率为:将本应该是相同类别的测试图 像错误地认为是不同类别的比例。
其中,错误接受率=(NFA/NIRA)×100%。NFA为所有测试图像发生“错误接受”的次数之和。例如,有三个类别的测试图像,分别记作测试图像A~测试图像F,其中,测试图像A~B的类别标签均为“01”,测试图像C~D的类别标签均为“02”,测试图像E~F的类别标签均为“03”。假设,将测试图像A与其他测试图像进行相似度计算时,得出测试图像A与测试图像B、C的相似度阈值均大于测试阈值,测试图像A与测试图像D~F的相似度阈值小于测试阈值,则表明测试图像A发生一次“错误接受”(因为错误地认为测试图像A与C为一类)。以此类推,每个测试图像进行聚类后,都可以得到该测试图像发生“错误接受”的次数,所有测试图像发生“错误接受”的次数的总和即为上述NFA。NIRA为类间匹配的总次数,例如,测试图像分为M1个类别,每个类别包括N1个测试图像,那么NIRA=M1×N1×(M1-1)×N1。
错误拒绝率=(NFR/NGRA)×100%。NFR为所有测试图像发生“错误拒绝”的次数之和。例如,有三个类别的测试图像,分别记作测试图像A~测试图像F,其中,测试图像A~B的类别标签均为“01”,测试图像C~D的类别标签均为“02”,测试图像E~F的类别标签均为“03”。假设,将测试图像B与其他测试图像进行相似度计算时,得出测试图像B与其余所有测试图像的相似度均小于测试阈值,则表明测试图像B发生一次“错误拒绝”(因为错误地认为测试图像B与测试图像A为不同类别)。以此类推,每个测试图像进行聚类后,均可以得到该测试图像发生“错误拒绝”的次数,所有测试图像发生“错误拒绝”的次数的总和即为上述NFR。NGRA为类内匹配的总次数,例如,测试图像分为M1个类别,每个类别包括N1个测试图像,那么NGRA=M1×N1×(N1-1)。
例如,测试集中有110类测试图像,每类包括8张测试图像。假定在对每个测试图像进行聚类的过程中,共出现了1000次“错误接受”的情况以及160次“错误拒绝”的情况。那么,类间匹配的总次数为110×8×109×8=767360次,类内匹配的总次数为110×8×7=6160次,错误接受率为1000/767360*100%=0.13%,错误拒绝率为160/6160=2.6%。
S002c、获取满足预设测试条件的聚类错误率所对应的测试阈值,并将该测试阈值作为测试集中的图像测试样本的类别的数量等级对应的相似度阈值。
在一些实施例中,预设测试条件包括:错误拒绝率和错误接受率相等。
本公开实施例还提供一种应用于电子设备的图像分类存储方法,电子设备包括图像采集单元。例如,电子设备可以为手机、电脑等终端;又例如,电子设备为具有图像采集单元的门禁设备。
图5为本公开的一些实施例中提供的图像分类存储方法的示意图,如图5所示,图像分类存储方法包括:
S11、响应于图像采集单元采集到的第一图像,将第一图像作为待聚类人脸图像,并利用上述人脸聚类方法对第一图像进行聚类处理。
在一个示例中,图像采集单元可以实时进行视频采集,第一图像为图像采集单元从视频中获取到的人脸图像。每获取一张第一图像,均对该第一图像进行作为待聚类人脸图像,进行上述步骤S01和步骤S02。图5为本公开的一些实施例中提供的图像分类存储方法的示意图,如图5所示,图像分类存储方法包括:
图6为本公开的另一些实施例中提供的图像分类存储方法的示意图,如图6所示,其中,在图像采集单元采集到第一张第一图像之前,电子设备中存储有批量图像集。
如图6所示,图像分类存储方法包括:
S10、在图像采集单元采集到第一张第一图像之前,将批量图像集中的图像依次作为待聚类人脸图像,进行聚类处理。
在一个示例中,在对批量图像集中的图像进行聚类处理之前,图像类别库中可以不存在任何图像;在对批量图像集中的图像依次进行上述步骤S01和步骤S02的聚类处理后,从而得到初始化的图像类别库,初始化后的图像类别库中包括多个类别的图像。这种情况下,对批量图像集中的第一个图像进行聚类处理时,图像类别库中没有图像,也就不存在待聚类的图像的同类别图像,因此,为批量图像集中的第一个图像分配类别标签后,将批量图像集中的第一个图像加入图像类别库中。
S11、响应于图像采集单元采集到的第一人脸图像,将第一人脸图像作为待聚类人脸图像,并利用上述步骤S01和S02对第一人脸图像进行聚类处理,以获得第一人脸图像的类别标签。
S12、输出第一人脸图像的类别标签。
另外,在实际应用中,还可以将同一类别的图像可以保存在同一文件夹中,将不同类别的图像保存在不同的文件夹中。
S13、响应于标签更新信号,进行标签更新步骤,该标签更新步骤包括:基于当前的相似度阈值对已完成聚类处理的人脸图像逐个进行类别标签的更新。
其中,已完成聚类处理的人脸图像是指,已获得类别标签的人脸图像。例如,已完成聚类处理的人脸图像包括:图像类别库中的人脸图像,还包括:已完成聚类处理、但未加入图像类别库中的人脸图像。
在一些实施例中,标签更新信号可以为电子设备在设定条件下所产生的信号。例如,当图像采集单元采集到的第一人脸图像达到一定数量后,电子设备产生标签更新信号;又例如,当物理时间达到预设 时间(例如,每周一的凌晨2点)时,电子设备产生标签更新信号。
因此,本公开实施例中的步骤S13与步骤S11的先后顺序不作限定。例如,在完成对图像类别库中的每个图像的类别更新后,当图像采集单元采集到新的第一人脸图像时,继续执行步骤S11。
其中,在进行标签更新步骤时,是基于当前的相似度阈值以及待更新标签的人脸图像与已更新标签的至少一类人脸图像的相似度,来进行类别标签更新的。
可以理解的是,当前的相似度阈值即为,图像类别库中当前的图像类别的数量等级所对应的相似度阈值。另外,电子设备可以在多个时刻产生标签更新信号,每产生一次标签更新信号,均执行一次标签更新步骤。在同一次标签更新步骤中,相似度阈值保持不变。
在一些实施例中,对任意一个已完成聚类处理的人脸图像的类别标签进行更新的步骤包括:根据当前的相似度阈值、以及待更新标签的人脸图像与已更新标签的至少一类人脸图像的相似度,判断已更新标签的人脸图像中是否存在待更新标签的人脸图像的同类别图像。当已更新标签的人脸图像中存在待更新标签的人脸图像的同类别图像时,根据待更新标签的人脸图像的同类别图像的类别标签,确定待更新标签的人脸图像的新的类别标签;当已更新标签的人脸图像中不存在待更新标签的人脸图像的同类别图像时,按照第二预设规则为待更新标签的人脸图像分配新的类别标签。
应当理解的是,对某一个已完成聚类处理的人脸图像进行类别标签的更新时,该人脸图像即为“待更新标签的人脸图像”。
其中,当某一个已更新标签的人脸图像与待更新标签的人脸图像的相似度大于或等于当前的相似度阈值时,则该已更新标签的人脸图像可以作为待更新标签的人脸图像的同类别图像。
在一些实施例中,已完成聚类处理的人脸图像按照第一次序逐个 进行类别标签的更新,其中,第一次序为:已完成聚类处理的人脸图像获得类别标签的次序。类别标签包括类别序号。这种情况下,按照第二预设规则为待更新标签的人脸图像分配新的类别标签,具体包括:将当前已更新标签的人脸图像的类别总数加1,得到待更新标签的人脸图像的新的类别标签。
示例性地,响应于标签更新信号而执行的标签更新步骤可以看做以下过程:
建立一个更新图像库,其中,在接收到标签更新信号的初始时刻,更新图像库为空集;之后,按照上述第一次序对已完成聚类处理的人脸图像逐个进行类别标签的更新。对任意一个已完成聚类处理的人脸图像的类别标签进行更新的过程包括:根据当前的相似度阈值、以及待更新标签的人脸图像与更新图像库中的至少一类图像的相似度,判断更新图像库中是否存在待更新标签的人脸图像的同类别图像;若存在,则将该同类别图像的类别标签作为待更新标签的人脸图像的新类别标签;否则,将更新图像库中的人脸图像的类别总数加1,得到待更新标签的人脸图像的新类别标签。当所有已完成聚类处理的人脸图像全部完成类别标签的更新后,以更新图像库代替图像类别库。
通过对已完成聚类处理的人脸图像进行类别标签的更新,可以提高各图像的聚类的准确性。
图7为本公开的一些实施例中提供的图像聚类装置的示意图,如图7所示,图像聚类装置包括存储器11和处理器12,存储器11上存储有计算机程序,其中,所述计算机程序被处理器12执行时上述实施例中的人脸聚类方法。
本公开还提供一种计算机可读存储介质,其上存储有计算机程序,该计算机程序被处理器执行时实现上述人脸聚类方法或者图像分类存储方法。
图8为本公开的一些实施例中提供的电子设备的示意图,如图8所示,电子设备包括:图像采集单元30、存储器40和处理器50,图像采集单元30包括摄像头等用于进行图像采集的器件。存储器40上存储有计算机程序,其中,计算机程序被处理器50执行时实现权上述的图像分类存储方法。
上述存储器和所述计算机可读存储介质包括但不限于以下可读介质:诸如随机存取存储器(RAM)、只读存储器(ROM)、非易失性随机存取存储器(NVRAM)、可编程只读存储器(PROM)、可擦除可编程只读存储器(EPROM)、电可擦除PROM(EEPROM)、闪存、磁或光数据存储、寄存器、磁盘或磁带、诸如光盘(CD)或DVD(数字通用盘)的光存储介质以及其它非暂时性介质。处理器的示例包括但不限于通用处理器、中央处理单元(CPU)、微处理器、数字信号处理器(DSP)、控制器、微控制器、状态机等。
可以理解的是,以上实施方式仅仅是为了说明本公开的原理而采用的示例性实施方式,然而本公开并不局限于此。对于本领域内的普通技术人员而言,在不脱离本公开的精神和实质的情况下,可以做出各种变型和改进,这些变型和改进也视为本公开的保护范围。

Claims (21)

  1. 一种人脸聚类方法,包括:获取待聚类人脸图像;对所述待聚类人脸图像进行聚类处理,所述聚类处理包括:
    获取与图像类别库中当前的图像类别的数量等级所对应的相似度阈值;其中,至少两个不同的数量等级所对应的相似度阈值不同;
    获取待聚类人脸图像与所述图像类别库中至少一类图像的相似度;
    根据当前的相似度阈值以及所述相似度,判断所述图像类别库中是否存在所述待聚类人脸图像的同类别图像;当所述图像类别库中存在所述待聚类人脸图像的同类别图像时,根据所述待聚类人脸图像的同类别图像的类别标签确定所述待聚类人脸图像的类别标签;当所述图像类别库中不存在所述待聚类人脸图像的同类别图像时,按照第一预设规则为所述待聚类人脸图像分配类别标签。
  2. 根据权利要求1所述的人脸聚类方法,其中,所述相似度阈值与所述图像类别库中当前的图像类别的数量等级正相关。
  3. 根据权利要求1所述的人脸聚类方法,其中,当所述图像类别库中不存在所述待聚类人脸图像的同类别图像时,所述聚类处理还包括:
    将所述待聚类人脸图像加入所述图像类别库中。
  4. 根据权利要求1所述的人脸聚类方法,其中,获取与图像类别库中当前的图像类别的数量等级所对应的相似度阈值,包括:
    判断所述图像类别库中的图像类别的数量所处的数量等级是否发 生变化;当所述图像类别库中的图像类别的数量所处的数量等级未发生变化时,采用上一张待聚类人脸图像进行聚类处理时所用的相似度阈值;当所述图像类别库中的图像类别的数量所处的数量等级发生变化时,根据变化后的数量等级更新所述相似度阈值。
  5. 根据权利要求1所述的人脸聚类方法,其中,在所述图像类别库中,每个类别包括一张图像,
    获取待聚类人脸图像与所述图像类别库中至少一类图像的相似度;根据当前的相似度阈值以及所述相似度,判断所述图像类别库中是否存在所述待聚类人脸图像的同类别图像,具体包括:
    确定所述图像类别库中的各类图像与所述待聚类人脸图像的比对次序;
    根据所述比对次序进行至少一次比对步骤,其中,第i次比对步骤包括:计算所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度,当所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度大于或等于所述相似度阈值时,确定所述图像类别库中比对次序为i的一类图像为所述待聚类人脸图像的同类别图像;当所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度小于所述相似度阈值时,进行下一次比对步骤;
    其中,i为大于零的整数,且i小于或等于图像类别库中当前的图像类别的数量。
  6. 根据权利要求5所述的人脸聚类方法,其中,计算所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度,包括:
    获取所述图像类别库中比对次序为i的一类图像的特征向量,作为第一特征向量;
    获取所述待聚类人脸图像的特征向量,作为第二特征向量;
    根据所述第一特征向量与所述第二特征向量的相似度,确定所述图像类别库中比对次序为i的一类图像与所述待聚类人脸图像的相似度。
  7. 根据权利要求1至6中任意一项所述的人脸聚类方法,其中,
    第1数量等级所对应的类别数量为(0,10000],第1数量等级所对应的相似度阈值在0.3~0.6之间;
    第2数量等级所对应的类别数量为(10000,20000],第2数量等级所对应的相似度阈值在0.61~0.64之间;
    第3数量等级所对应的类别数量为(20000,30000],第3数量等级所对应的相似度阈值在0.65~0.664之间;
    第4数量等级所对应的类别数量为(30000,50000],第4数量等级所对应的相似度阈值在0.665~0.70之间;
    第5数量等级所对应的类别数量为(50000,∞),第5数量等级所对应的相似度阈值在0.705~0.9之间。
  8. 根据权利要求1至6中任意一项所述的人脸聚类方法,其中,所述类别标签包括类别序号;
    按照第一预设规则为所述待聚类人脸图像分配类别标签,包括:
    将所述图像类别库中当前的图像类别的数量加1,得到所述待聚类人脸图像的类别序号。
  9. 根据权利要求1至6中任意一项所述的人脸聚类方法,其中,所述人脸聚类方法还包括:在获取第一张待聚类人脸图像之前,建立不同数量等级与相似度阈值的对应关系;
    其中,建立不同数量等级与相似度阈值的对应关系的过程包括:
    设置多个测试集,每个所述测试集包括多类测试图像;其中,在不同的测试集中,所述测试图像的类别数量处于不同的数量等级;
    对于每个所述测试集,执行以下步骤:
    设置至少一个测试阈值;
    基于每个所述测试阈值,对所述测试集中的测试图像进行聚类,得到所述测试图像的测试聚类结果;并根据每个所述测试图像的测试聚类结果和理论聚类结果,确定所述测试阈值对应的聚类错误率;
    获取满足预设测试条件的聚类错误率所对应的测试阈值,并将该测试阈值作为所述测试集中测试图像的类别的数量等级所对应的相似度阈值。
  10. 根据权利要求9所述的人脸聚类方法,其中,所述聚类错误率包括:错误接受率和错误拒绝率。
  11. 根据权利要求10所述的人脸聚类方法,其中,所述预设测试条件包括:所述错误拒绝率和所述错误识别率相等。
  12. 根据权利要求9所述的人脸聚类方法,其中,除了最高数量等级所对应的测试集之外,其余每个所述测试集中,所述测试图像的 类别数量为相应的数量等级中的数量上限值。
  13. 一种应用于电子设备的图像分类存储方法,所述电子设备包括图像采集单元,所述图像分类存储方法包括:
    响应于所述图像采集单元采集到的第一人脸图像,将所述第一人脸图像作为待聚类人脸图像,并利用权利要求1至12中任意一项所述的人脸聚类方法对所述第一人脸图像进行所述聚类处理,以获得所述第一人脸图像的类别标签。
  14. 根据权利要求13所述的图像分类存储方法,其中,所述图像分类存储方法还包括:
    输出所述第一人脸图像的类别标签。
  15. 根据权利要求13所述的图像分类存储方法,其中,在所述图像采集单元采集到第一张第一人脸图像之前,所述电子设备中存储有批量图像集;
    所述图像分类存储方法还包括:
    在所述图像采集单元采集到第一张所述第一人脸图像之前,将所述批量图像集中的人脸图像依次作为所述待聚类人脸图像,进行所述聚类处理。
  16. 根据权利要求13至15中任意一项所述的图像分类存储方法,其中,所述图像分类存储方法还包括:
    响应于标签更新信号,基于当前的相似度阈值对已完成聚类处理 的人脸图像逐个进行类别标签的更新。
  17. 根据权利要求16所述的图像分类存储方法,其中,对任意一个已完成聚类处理的人脸图像的类别标签进行更新的步骤包括:
    根据当前的相似度阈值、以及待更新标签的人脸图像与已更新标签的至少一类人脸图像的相似度,判断已更新标签的人脸图像中是否存在所述待更新标签的人脸图像的同类别图像;当已更新标签的人脸图像中存在所述待更新标签的人脸图像的同类别图像时,根据所述待更新标签的人脸图像的同类别图像的类别标签,确定所述待更新标签的人脸图像的新的类别标签;当已更新标签的人脸图像中不存在所述待更新标签的人脸图像的同类别图像时,按照第二预设规则为所述待更新标签的人脸图像分配新的类别标签;
    其中,对已完成聚类处理的人脸图像逐个进行类别标签的更新的过程中,所述相似度阈值保持不变。
  18. 根据权利要求17所述的图像分类存储方法,其中,所述类别标签包括类别序号;
    已完成聚类处理的人脸图像按照第一次序逐个进行类别标签的更新,所述第一次序为:已完成聚类处理的人脸图像获得类别标签的次序;
    按照第二预设规则为所述待更新标签的人脸图像分配新的类别标签,包括:
    将当前已更新标签的人脸图像的类别总数加1,得到所述待更新标签的人脸图像的新的类别标签。
  19. 一种人脸聚类装置,包括:存储器和处理器,所述存储器上存储有计算机程序,其中,所述计算机程序被所述处理器执行时实现权利要求1至12中任一所述的人脸聚类方法。
  20. 一种计算机可读存储介质,其上存储有计算机程序,该计算机程序被处理器执行时实现权利要求1至12中任一所述的人脸聚类方法或权利要求13至18中任一所述的图像分类存储方法。
  21. 一种电子设备,包括:
    图像采集单元;
    存储器和处理器,所述存储器上存储有计算机程序,其中,所述计算机程序被所述处理器执行时实现权利要求13至18中任一所述的图像分类存储方法。
PCT/CN2021/081539 2021-03-18 2021-03-18 人脸聚类方法及装置、分类存储方法、介质、电子设备 WO2022193232A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
CN202180000510.7A CN115398485A (zh) 2021-03-18 2021-03-18 人脸聚类方法及装置、分类存储方法、介质、电子设备
PCT/CN2021/081539 WO2022193232A1 (zh) 2021-03-18 2021-03-18 人脸聚类方法及装置、分类存储方法、介质、电子设备
US17/773,123 US20230245421A1 (en) 2021-03-18 2021-03-18 Face clustering method and apparatus, classification storage method, medium and electronic device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/081539 WO2022193232A1 (zh) 2021-03-18 2021-03-18 人脸聚类方法及装置、分类存储方法、介质、电子设备

Publications (1)

Publication Number Publication Date
WO2022193232A1 true WO2022193232A1 (zh) 2022-09-22

Family

ID=83321335

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/081539 WO2022193232A1 (zh) 2021-03-18 2021-03-18 人脸聚类方法及装置、分类存储方法、介质、电子设备

Country Status (3)

Country Link
US (1) US20230245421A1 (zh)
CN (1) CN115398485A (zh)
WO (1) WO2022193232A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117853770B (zh) * 2024-02-01 2024-07-16 杭州海康威视系统技术有限公司 人体聚类方法、装置及设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108073948A (zh) * 2012-01-17 2018-05-25 华为技术有限公司 一种照片分类管理方法、服务器、装置及系统
CN108875834A (zh) * 2018-06-22 2018-11-23 北京达佳互联信息技术有限公司 图像聚类方法、装置、计算机设备及存储介质
CN109426781A (zh) * 2017-08-29 2019-03-05 阿里巴巴集团控股有限公司 人脸识别数据库的构建方法、人脸识别方法、装置及设备
US20190146991A1 (en) * 2016-06-09 2019-05-16 Panasonic Intellectual Property Management Co., Ltd. Image search device, image search system, and image search method
CN111079653A (zh) * 2019-12-18 2020-04-28 中国工商银行股份有限公司 数据库自动分库方法及装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108073948A (zh) * 2012-01-17 2018-05-25 华为技术有限公司 一种照片分类管理方法、服务器、装置及系统
US20190146991A1 (en) * 2016-06-09 2019-05-16 Panasonic Intellectual Property Management Co., Ltd. Image search device, image search system, and image search method
CN109426781A (zh) * 2017-08-29 2019-03-05 阿里巴巴集团控股有限公司 人脸识别数据库的构建方法、人脸识别方法、装置及设备
CN108875834A (zh) * 2018-06-22 2018-11-23 北京达佳互联信息技术有限公司 图像聚类方法、装置、计算机设备及存储介质
CN111079653A (zh) * 2019-12-18 2020-04-28 中国工商银行股份有限公司 数据库自动分库方法及装置

Also Published As

Publication number Publication date
US20230245421A1 (en) 2023-08-03
CN115398485A (zh) 2022-11-25

Similar Documents

Publication Publication Date Title
US11348249B2 (en) Training method for image semantic segmentation model and server
CN110837410B (zh) 任务调度方法、装置、电子设备及计算机可读存储介质
CN111476256A (zh) 基于半监督学习的模型训练方法、装置及电子设备
CN111104825A (zh) 人脸注册库更新方法、装置、设备及介质
CN113379301A (zh) 通过决策树模型对用户进行分类的方法、装置和设备
CN110705489B (zh) 目标识别网络的训练方法、装置、计算机设备和存储介质
CN112101156B (zh) 一种目标识别的方法、装置和电子设备
CN113111063A (zh) 一种应用于多数据源的医疗患者主索引发现方法
US11455500B2 (en) Automatic classifier profiles from training set metadata
WO2022193232A1 (zh) 人脸聚类方法及装置、分类存储方法、介质、电子设备
CN111507400A (zh) 应用分类方法、装置、电子设备以及存储介质
CN115174131A (zh) 基于异常流量识别的信息拦截方法、系统及云平台
CN113159133A (zh) 数据标注方法及装置、电子设备、存储介质
CN116363416A (zh) 一种图像去重方法、装置、电子设备和存储介质
CN111783869B (zh) 训练数据筛选方法、装置、电子设备及存储介质
EP4030316A1 (en) Face search method and device, apparatus, and computer-readable storage medium
CN112699908B (zh) 标注图片的方法、电子终端、计算机可读存储介质及设备
CN112783843A (zh) 数据读取方法、装置及电子设备
CN112989869B (zh) 人脸质量检测模型的优化方法、装置、设备及存储介质
CN112395119B (zh) 异常数据处理方法、装置、服务器及存储介质
CN112861962A (zh) 样本处理方法、装置、电子设备和存储介质
CN111143626A (zh) 团伙识别方法、装置、设备及计算机可读存储介质
CN112927004A (zh) 用于大数据画像的信息云计算分析方法及信息推送服务器
CN112882888A (zh) 基于智能在线通信和大数据的信息生成方法及系统
CN112949362B (zh) 人员信息标注方法、装置及电子设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21930824

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21930824

Country of ref document: EP

Kind code of ref document: A1