WO2023125839A1

WO2023125839A1 - Facial data archiving method and related device

Info

Publication number: WO2023125839A1
Application number: PCT/CN2022/143530
Authority: WO
Inventors: 章跃; 谢友平; 刘国伟
Original assignee: 深圳云天励飞技术股份有限公司; 成都云天励飞技术有限公司
Priority date: 2021-12-31
Filing date: 2022-12-29
Publication date: 2023-07-06
Also published as: CN114443893A

Abstract

The embodiments of the present invention provide a facial data archiving method. The method comprises: acquiring batch facial data to be archived, wherein each piece of facial data comprises a facial feature and a facial pose feature; calculating the similarity between every two facial features among all facial features, so as to obtain the global similarity of each piece of facial data; performing threshold screening on the global similarity of each piece of facial data according to the facial pose feature and a preset similarity threshold matrix, so as to obtain a first facial data set, wherein the preset similarity threshold matrix is obtained according to facial features and facial pose features included in archived facial data in a preset archive library; and archiving the first facial data set by means of the archived facial data in the preset archive library. A similarity threshold adaptive to a facial pose feature can be provided by means of a preset threshold matrix, such that the problem in the prior art of the archiving rate being low due to the usage of a unified threshold evaluation mode is solved.

Description

Method for archiving face data and related equipment

This application claims the priority of the Chinese patent application with the application number 202111678101.0 and the title of the invention "Face data archiving method and related equipment" submitted to the China Patent Office on December 31, 2021, the entire contents of which are incorporated herein by reference Applying.

technical field

The invention relates to the field of artificial intelligence, in particular to a face data archiving method and related equipment.

Background technique

With the continuous development of artificial intelligence and the continuous improvement of image recognition related technologies, more and more face recognition applications have been implemented. The archiving of face data is to attribute the face data of the same person to the same face file for face data query and processing. In the archiving of face data, the usual archiving method is to perform clustering and archiving of face data. However, factors such as face posture, brightness, and clarity have a great impact on clustering and archiving, resulting in a low archiving rate. Among them, when the face pose changes very little, the similarity difference is not big, but after the pose change reaches a certain level, the similarity changes greatly, which in turn affects the archiving rate. Therefore, due to the change of face posture, there is a problem of low archiving rate in the archiving of existing face data.

Contents of the invention

An embodiment of the present invention provides a method for archiving face data. After obtaining batches of face data from video streams, the corresponding face features are extracted for archiving. During the archiving process, a preset threshold matrix is used for similarity judgment. Since the preset threshold matrix is obtained according to the facial pose characteristics of the archived face data, a similarity threshold suitable for the facial pose characteristics can be provided through the preset threshold matrix to solve the problem of using unified The threshold evaluation method makes the problem of low filing rate.

In a first aspect, an embodiment of the present invention provides a method for archiving face data, the method comprising:

Obtain batches of face data to be archived, the face data including face features and face gesture features;

Calculate the pairwise similarity between all facial features to obtain the global similarity of each face data;

Threshold screening is performed on the global similarity of each face data according to the facial pose feature and a preset similarity threshold matrix to obtain a first human face data set, and the preset similarity threshold matrix is based on Obtain the facial features and facial posture features contained in the archived face data in the preset archive library;

The first human face data set is archived through the archived human face data in the preset archive library.

Optionally, the method also includes:

Obtain the archived face data in the preset archive library, and extract the face features and facial pose features of the archived face data;

The preset threshold value matrix is constructed and obtained according to the facial features and facial pose features of the archived facial data.

Optionally, the multiple archived face datasets include a frontal face dataset and multiple deflected face datasets, the facial pose features include face angle feature values, and the The face feature and the face pose feature of the face data are constructed to obtain the preset threshold matrix, including:

For each archived face file in the preset archive library, the face data contained in the archived face file is divided according to the face angle feature value to obtain a plurality of different face angle intervals. The archived face dataset;

Calculate the average similarity between the archived face data sets according to the face features of the archived face data, or calculate the difference between the frontal face data set and the archived face data according to the face features of the archived face data. The average similarity between the deflected face datasets;

According to the average similarity between the archived face data sets, the average similarity matrix of each archived human face file is constructed, wherein, each of the archived human face files corresponds to a described average similarity matrix, so The dimension of the average similarity matrix is related to the number of the divided intervals;

The preset threshold matrix is obtained by calculating according to the average similarity matrix.

Optionally, one average similarity matrix unit in the average similarity matrix corresponds to an average similarity value, and the calculation according to the average similarity matrix to obtain the preset threshold matrix includes:

For all the average similarity matrices, extracting the average similarity values in the corresponding average similarity matrix units is added and averaged to obtain the final similarity matrix;

The final similarity matrix is used as the preset threshold matrix.

Optionally, archiving the first face data set through the archived face data in the preset archive library includes:

Filtering the first human face data set according to the human face spatiotemporal characteristics to obtain a second human face data set;

The archiving process is performed on the second human face data set through the archived human face data in the preset archiving library.

Optionally, the second face relationship graph includes a first face relationship cluster, and the second face data set is archived through the archived face data in the preset archive library, include:

Based on the second human face data set, generate and form a first human face relational graph;

Performing strong connectivity calculations on the first human face relational graph to obtain a second human face relational graph;

The batch of face data is archived based on the second face relationship graph.

Optionally, the second face relation graph includes a first face relation cluster, and archiving the batch of face data based on the second face relation graph includes:

Optimizing the first face relationship cluster according to a preset community discovery algorithm to obtain a second face relationship cluster;

The batch of face data is archived based on the second face relation cluster.

In a second aspect, an embodiment of the present invention provides a face data archiving device, the device comprising:

Obtaining module, is used for obtaining batch face data to be archived, and described face data comprises face feature and face gesture feature;

Calculation module, used to calculate the pairwise similarity between all facial features, and obtain the global similarity of each facial data;

The first screening module is configured to perform threshold screening on the global similarity of each face data according to the face pose feature and a preset similarity threshold matrix to obtain a first face data set, the preset The similarity threshold matrix is obtained according to the facial features and facial posture features contained in the archived face data in the preset archive library;

The archiving module is configured to archive the first human face data set through the archived human face data in the preset archiving library.

In a third aspect, an embodiment of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and operable on the processor. When the processor executes the computer program, The steps in the archiving method of face data provided by the embodiment of the present invention are realized.

In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the archiving of face data provided by the embodiment of the invention is realized steps in the method.

In the embodiment of the present invention, batches of face data to be archived are obtained, and the face data includes face features and face pose features; the pairwise similarity between all face features is calculated to obtain the global profile of each face data Similarity; according to the face pose feature and the preset similarity threshold matrix, threshold screening is performed on the global similarity of each face data to obtain the first face data set, and the preset similarity threshold The value matrix is obtained according to the facial features and facial posture features contained in the archived face data in the preset archive library; through the archived face data in the preset archive library, the first face data Sets are archived. By obtaining batches of face data from the video stream and extracting the corresponding face features for archiving, in the archiving process, the preset threshold matrix is used for similarity judgment, because the preset threshold matrix is based on the archived face data Therefore, a preset threshold matrix can be used to provide a similarity threshold that is compatible with the facial posture features, so as to solve the problem of low filing rate caused by the unified threshold evaluation method in the prior art.

Description of drawings

The drawings that need to be used in the embodiments of the present application will be introduced below.

Fig. 1 is the flow chart of a kind of archiving method of face data provided by the embodiment of the present invention;

Fig. 2 is a flowchart of a method for constructing a threshold matrix provided by an embodiment of the present invention;

Fig. 3 is a schematic structural diagram of a face data archiving device provided by an embodiment of the present invention;

4 is a schematic structural diagram of another archiving device for face data provided by an embodiment of the present invention;

Fig. 5 is a schematic structural diagram of a building block provided by an embodiment of the present invention;

Fig. 6 is a schematic structural diagram of a second calculation sub-module provided by an embodiment of the present invention;

Fig. 7 is a schematic structural diagram of an archiving module provided by an embodiment of the present invention;

Fig. 8 is a schematic structural diagram of an archiving sub-module provided by an embodiment of the present invention;

Fig. 9 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention.

Detailed ways

Embodiments of the present application are described below in conjunction with the accompanying drawings.

Please refer to Fig. 1, Fig. 1 is the flow chart of a kind of archiving method of human face data provided by the embodiment of the present invention, as shown in Fig. 1, the archiving method of this human face data comprises the following steps:

101. Obtain batches of face data to be archived.

In the embodiment of the present invention, the above-mentioned video stream may be a video stream obtained in real time by a camera, or a video stream uploaded by a user, and the above-mentioned camera may be a camera installed in each scene, such as a camera installed indoors or installed outdoors camera.

Further, the foregoing video stream may be a video stream collected by one camera, or may be a video stream collected by multiple cameras.

The batch face data can be obtained from the video stream through the face detection algorithm, and the batch face data can be obtained from the video through the face detection algorithm. The face data includes face features and face pose features. Each face data also includes face image information, face confidence information, and face detection frame ID. In the face detection algorithm, the same face detection frame ID will be assigned to the same target identified by the algorithm.

Each face data also includes time information and position information. When the camera captures the video stream, the image frame in the video stream includes time information, and the video stream includes position information. The above time information can be understood as the time point when the image frame is collected. The location information can be understood as the location where the video stream is collected, and the above location information can also be understood as the installation location of the camera, and each camera will be provided with corresponding latitude and longitude information as the location information when it is installed.

The time information corresponding to the face data can be obtained through the time information corresponding to the image frame, and the position information corresponding to the face data can be obtained through the position information of the video stream.

The face features corresponding to the face data can be extracted from the face image information through the face feature extraction algorithm, and the face pose features corresponding to the face data can be extracted from the face image information through the face state estimation algorithm , the spatiotemporal features of the face corresponding to the face data can be extracted from the above time information and position information.

The above-mentioned face feature can be a face feature vector, and the face feature vector is to map the high-dimensional space image data of the face image to a low-dimensional, forming a preset dimension feature vector (such as a 512-dimensional feature vector) that can represent the feature of the picture itself , different face images can be judged by comparing the distance between the face feature vectors to determine whether they are similar. The closer the face feature vectors are, the greater the probability that the face images are faces of the same person. The degree of discretization of the face feature vector depends on the position of the selected face key points and the degree of distinction of each feature.

The facial posture feature mentioned above may be an angle value of a face orientation deflection, and the face orientation deflection may include a pitch angle deflection, a yaw angle deflection, a roll angle deflection, and the like.

102. Calculate the pairwise similarity between all face features, and obtain the global similarity of each face data.

In the embodiment of the present invention, one face data corresponds to one face feature, and a batch of face data corresponds to a batch of face features, and for the batch of face features, the similarity between two pairs is calculated. It is equivalent to performing a similarity calculation on a facial feature with other facial features except itself to obtain the global similarity of the facial feature.

For example, the batch of face data is M face data, and the corresponding M face features are extracted, and each face feature is calculated with M-1 face features except itself to obtain the corresponding global similarity , at this time, the global similarity corresponding to each face feature has M-1 similarities.

In a possible embodiment, for each face data, the highest TopN similarities among the M-1 similarities can be selected as the global similarity of the face data.

103. Perform threshold screening on the global similarity of each face data according to the facial posture features and the preset similarity threshold matrix.

In the embodiment of the present invention, the preset threshold value matrix is obtained according to the face pose characteristics of the archived face data. Specifically, thresholds corresponding to different facial posture features are set in the above-mentioned preset threshold matrix, and the corresponding thresholds can be matched in the preset threshold matrix according to different facial posture features, and the global similarity can be compared with the corresponding thresholds. In comparison, if there is a similarity greater than a corresponding threshold in the global similarity, the corresponding face data may be added to the first face data set, wherein one face data corresponds to one first face data set.

For example, for a face data A, the face pose feature of M-1 face data is A1, and in the remaining M-1 face data, there is any one face data B and the face data The similarity of A is S1, and the face pose feature of face data B is B1, then the corresponding threshold C1 can be found in the preset threshold matrix according to (A1, B1), and the similarity S1 is compared with the threshold C1, If the similarity S1 is greater than or equal to the threshold C1, the face data B is added to the first face data set of the face data A. If the similarity S1 is smaller than the threshold C1, no processing is performed. The first face data set of the face data A includes the face data A.

104. Perform archiving processing on the first human face data set by presetting the archived human face data in the archiving repository.

In the embodiment of the present invention, the first face data set retains the face data with a higher similarity with the master face data, and the first face data set corresponding to the master face data with the same face detection frame can be grouped together. into the same file.

Optionally, the first human face data set is screened according to the human face spatio-temporal characteristics to obtain a second human face data set; Data sets are archived.

The above-mentioned spatio-temporal feature of the face includes time information and position information, and the first face data set is screened according to the spatio-temporal feature of the face.

Further, for the first face data set of a face data, the face data is the main face data of the first face data set, if two faces with the same time information appear in the first face data set or two or more face data, then keep the face data with the highest similarity with the master face data among the two or more face data, and delete the rest of the face data in the first face data set. If two or more face data with the same time information and different position information appear in the first face data set, then keep the one with the highest similarity with the main face data among the two or more face data For the face data, delete the rest of the face data in the first face data set. Judging whether the position and time relationship between the master face data and each face data conforms to the first preset relationship, if the position and time relationship between the master face data and certain face data does not conform to the first preset relationship, then the The certain face data is deleted from the first face data set. Judging whether the position and time relationship between each two face data conforms to the second preset relationship except the main face data, if the position and time relationship between certain two face data do not conform to the second preset relationship, then judge Whether the two face data match the trajectory data corresponding to the master face data, delete the face data that does not match the trajectory data corresponding to the master face data. The trajectory data corresponding to the master face data can be determined according to the face detection frame ID corresponding to the master face data, specifically, the face data having the same face detection frame ID as the master face data and the master face data can form corresponding trajectory data.

For example, the location information of two face data is LA and LB, and the time information is TA and TB respectively, then the expression of the relationship between the above location and time is: ε=|LA-LB|/|TA-TB|. The above-mentioned first preset relationship is ε1, if the position and time relationship ε between the master’s face data and a certain face data is greater than ε1, it means that the position and time relationship between the master’s face data and a certain face data does not conform to the first default relationship. The above second preset relationship is ε2, if the position and time relationship ε between certain two face data is greater than ε2, it means that the position and time relationship between certain two face data does not conform to the second preset relationship.

The first human face data set is screened by the above method to obtain the second human face data set, wherein one second human face data set corresponds to one human face data.

In the embodiment of the present invention, the face data with higher similarity with the main face data is retained in the second face data set, and the facial spatiotemporal data is used to filter out the face data that does not conform to the spatiotemporal relationship, and the second face data set is added. The probability that the face data belongs to the face of the same person.

Further, the second face data sets corresponding to the main face data with the same face detection frame can be grouped into the same file.

In a possible embodiment, the second face data set is clustered and archived through a part of the archived face data, and the archiving of the face data is to archive the unarchived face data into this part of the archived face data in the corresponding file. Specifically, the archived face data also includes file tags, and the archived face data is used as the clustering center to cluster the main face data of all the second face data sets. The second face dataset is deduplicated and archived.

In a possible embodiment, the second face data set is formed into a face relation graph, and archived according to the face relation clusters in the face relation graph. It should be noted that, in the face relationship graph, each node represents a face data, and the edge between two nodes represents the similarity between two face data.

Optionally, the above-mentioned second human face data set can be formed into a first human face relational map; strong connectivity calculation is performed on the above-mentioned first human face relational map to obtain a second human face relational map; based on the above-mentioned second human face relational map Archive the above-mentioned batches of face data. Wherein, both the first human face relationship graph and the second human face relationship graph are directed graphs.

It should be noted that in computer graph theory, Strongly Connected (Strongly Connected) means that there is a path from v1 to v2 between any two points v1 and v2 in the directed graph G (Directed Graph). Points and edges are not repeated, it is called a path) and the path from v2 to v1. The method of bidirectional traversal to obtain the intersection can be used to find the strongly connected components of the first face relationship map, and the time complexity is O(N^2+M). The Kosaraju algorithm or the Tarjan algorithm can also be used to find the strongly connected components of the first face relation graph, both of which have a time complexity of O(N+M).

By performing strong connectivity calculations on the first face relation graph, some erroneous face data can be eliminated, making the face relation clusters more accurate.

Optionally, the second face relation map includes the first face relation cluster, and the first face relation cluster can be optimized according to a preset community discovery algorithm to obtain the second face relation cluster; based on the above second face relation cluster The face relation cluster archives the aforementioned batches of face data.

It should be noted that the second face relationship graph represents the face relationship network. In the face relationship network, some faces are closely connected, and some faces are relatively sparsely connected. In the face relational network, the closely connected part can be regarded as a community, and there is a relatively close connection between the internal nodes, while the relatively sparse connection between the two communities. The community discovery algorithm algorithm can realize the community division of large-scale networks with different granularities in a short period of time, and there is no need to specify the number of communities. When the modularity no longer increases, the iteration will automatically stop.

In the embodiment of the present invention, when the distribution of cameras in the outdoor scene is not very dense, and the differences in camera resolution and definition are large, the facial features captured in different scenes are greatly different. Clustering division will carry some wrong data, so the community discovery algorithm can be used to perform secondary optimization on the first face relationship cluster of the cluster group to improve the accuracy of face clustering.

Optionally, please continue to refer to FIG. 2. FIG. 2 is a flowchart of a threshold matrix construction method provided by an embodiment of the present invention. As shown in FIG. 2, before step 103 of the embodiment in FIG. 1, a preset threshold matrix is also included. The construction method specifically includes the following steps:

201. Acquire archived face data in a preset archive library, and extract face features and facial posture features of the archived face data.

In the embodiment of the present invention, multiple archived face files can be obtained, and each archived face file corresponds to the archived face data of a person, and the archived face data in the archived face files can include the cover person Face data and archived face data, wherein the cover face data can be front face data, and the face orientation deflection of the front face data can include a pitch angle deflection of 0°, a yaw angle deflection of 0°, and a roll angle deflection of 0°. Deflection is 0°.

For the same archived face file, there are multiple archived face data, each archived face data includes face image information, face confidence information, face detection frame ID, time information and location information, The face feature extraction algorithm can be used to extract the face features of the face image information of the archived face data, and the face features of the archived face data can be extracted, and the archived face data can be analyzed by the face pose estimation algorithm. The face image information is used to estimate the face state, and the face pose features of the archived face data are obtained.

202. Construct a preset threshold value matrix according to the facial features and facial posture features of the archived facial data.

In the embodiment of the present invention, for an archived face file, each of the archived face data therein has a corresponding face feature and a face pose feature extracted. It can calculate the face pose feature difference of two archived face data, and calculate the similarity of the two archived face data, and compare the similarity of the two archived face data with the two archived face data The facial pose features are poorly correlated.

Optionally, the above-mentioned face pose feature includes a face angle feature value, and for each archived face file in the preset archive library, according to the face angle feature value, the person included in the archived face file is The face data is divided to obtain multiple archived face data sets under different face angle intervals; the average similarity between the archived face data sets is calculated according to the facial features of the archived face data, or the average similarity between the archived face data sets is calculated according to the archived face data. The face feature of the face data calculates the average similarity between the frontal face data set and the deflected face data set; according to the average similarity between the archived face data sets, the construction of each archived face file is obtained. An average similarity matrix, wherein each archived face file corresponds to an average similarity matrix, and the dimension of the average similarity matrix is related to the number of divided intervals; a preset threshold matrix is obtained by calculating the average similarity matrix.

By dividing the face angle feature value and setting the average similarity corresponding to the division interval as the similarity threshold, it is possible to avoid the problem that a single similarity cannot judge whether the face is similar when the face angle is deflected.

For example, the face pose feature is the angle feature value of the face yaw angle, the angle feature value of the face yaw angle can be between 0° and 60° degrees, and the angle feature value of the face yaw angle can be expressed as Divide the number of preset angles. For example, divide the angle feature value of the face yaw angle by every 10°, then get 0°, (0, 10°], (10°, 20°], (20°, 30° °], (30°, 40°], (40°, 50°], (50°, 60°] and other intervals, if the above intervals are formed into a matrix, the following table 1 is obtained:

Table 1

The two archived face data are archived face data A and archived face data B respectively, and the angle feature value of the face yaw angle of archived face data A is 35°, falling into (30°, 40 °] interval, the angle eigenvalue of the face yaw angle of the archived face data B is 20°, which falls within the (10°, 20°] interval, calculate the archived face data A and the archived face data The similarity SAB between B, the obtained similarity SAB corresponds to (30°, 40°] and (10°, 20°], and Table 2 is obtained as follows:

Table 2

It should be noted that, in Table 1 and Table 2, after removing the first row and the first column, it is in the corresponding matrix form. When there are multiple similarities corresponding to (30°, 40°] and (10°, 20°], the average similarity is calculated and filled in. For example, there are two archived face data respectively archived person Face data C and archived face data D, the angle feature value of the face yaw angle of archived face data C is 35°, falling in the interval of (30°, 40°], the archived face data D’s The angular eigenvalue of the face yaw angle is 20°, which falls within the interval (10°, 20°), calculate the similarity SCD between the archived face data C and the archived face data D, and obtain the similarity SCD corresponds to (30°, 40°] and (10°, 20°], find the average S of SAB and SCD as the average similarity corresponding to (30°, 40°] and (10°, 20°], get Table 3 is as follows:

table 3

Optionally, the average similarity between the front face data set and the deflected face data set may be calculated according to the face features of the archived face data.

For an archived face file, the average similarity table as shown in Table 4 can be obtained, and Table 4 is as follows:

Table 4

Only a single average similarity can be retained, as shown in Table 5:

table 5

Wherein, the above-mentioned a0, a1, a2, a3, a4, a5, a6 are the average similarity between the frontal face data set and the deflected face data set calculated according to the face features of the archived face data.

It should be noted that, in Table 4 and Table 5 above, if the first row and the first column are removed, the corresponding average similarity matrix is obtained.

For multiple archived face files, there are multiple average similarity matrices correspondingly. For all the above average similarity matrices, the average similarity values in the corresponding average similarity matrix units can be extracted and averaged to obtain the final A similarity matrix; the above-mentioned final similarity matrix is used as the above-mentioned preset threshold matrix.

As the face deflection angle is small, the difference in similarity is not large, but after the deflection reaches a certain level, the similarity changes greatly. Using a unified threshold evaluation method cannot solve this large deflection problem well, and it is relatively difficult to evaluate. complex. In the embodiment of the present invention, the similarity threshold matrix of the smooth moving average similarity can be used to fit the similarity in various postures as the threshold matrix, which can further improve the clustering effect and further improve the archiving rate of face data.

It should be noted that the face data archiving method provided by the embodiment of the present invention can be applied to smart phones, computers, servers and other devices that can perform face data archiving.

Optionally, please refer to FIG. 3. FIG. 3 is a schematic structural diagram of an archiving device for face data provided by an embodiment of the present invention. As shown in FIG. 3, the device includes:

Acquisition module 301, is used for obtaining batches of human face data to be archived, and described human face data comprises human face feature and human face posture feature;

Calculation module 302, is used for calculating the pairwise similarity between all face features, obtains the global similarity of each face data;

The first screening module 303 is configured to perform threshold screening on the global similarity of each face data according to the facial pose feature and a preset similarity threshold matrix to obtain a first human face data set, the preset The set similarity threshold matrix is obtained according to the facial features and facial posture features contained in the archived face data in the preset archive library;

The archiving module 304 is configured to archive the first face data set through the archived face data in the preset archiving library.

Optionally, as shown in Figure 4, the device further includes:

The extraction module 306 is used to obtain the archived face data in the preset archive library, and extract the facial features and facial gesture features of the archived face data;

The construction module 307 is configured to construct the preset threshold matrix according to the facial features and facial pose features of the archived facial data.

Optionally, as shown in FIG. 5 , the multiple archived face datasets include a frontal face dataset and multiple deflected face datasets, the facial pose features include face angle feature values, and the Building blocks 307, including:

The division sub-module 3071 is used to divide the face data contained in the archived face files according to the face angle feature value for each archived face file in the preset archive library, to obtain multiple An archived face dataset under different face angle intervals;

The first calculation sub-module 3072 is used to calculate the average similarity between the archived face data sets according to the face features of the archived face data, or to calculate the average similarity according to the face features of the archived face data the average similarity between the frontal face dataset and the deflected face dataset;

Construction sub-module 3073, for constructing the average similarity matrix of each archived face file according to the average similarity between the archived face data sets, wherein each of the archived face files corresponds to one of the The average similarity matrix, the dimension of the average similarity matrix is related to the quantity of the divided intervals;

The second calculation sub-module 3074 is configured to calculate the preset threshold matrix according to the average similarity matrix.

Optionally, as shown in FIG. 6, an average similarity matrix unit in the average similarity matrix corresponds to an average similarity value, and the archiving module 304 includes:

The extracting unit 30741 is used for extracting the average similarity values in the corresponding average similarity matrix units for all the average similarity matrices, and performing addition and averaging to obtain a final similarity matrix;

The acquiring unit 30742 is configured to use the final similarity matrix as the preset threshold matrix.

Optionally, the archiving module 305 is also used to screen the first face data set according to the spatiotemporal features of the face to obtain the second face data set; through the archived face data in the preset archiving library, Perform archiving processing on the second human face data set.

Optionally, as shown in FIG. 7, the archiving module 305 includes:

An integration submodule 3051, configured to form the second face dataset into a first face relational graph;

The third calculation sub-module 3052 is used to perform strong connectivity calculation on the first human face relationship graph to obtain a second human face relationship graph;

The archiving submodule 3053 is configured to archive the batch of face data based on the second face relational graph.

Optionally, as shown in FIG. 8, the second face relation map includes the first face relation cluster, and the archiving submodule 3053 includes:

An optimization unit 50531, configured to optimize the first face relationship cluster according to a preset community discovery algorithm to obtain a second face relationship cluster;

The archiving unit 30532 is configured to archive the batch of face data based on the second face relation cluster.

It should be noted that the face data archiving device provided by the embodiment of the present invention can be applied to smart phones, computers, servers and other devices that can perform face data archiving.

The face data archiving device provided by the embodiment of the present invention can realize each process realized by the face data archiving method in the above method embodiment, and can achieve the same beneficial effect. To avoid repetition, details are not repeated here.

Referring to FIG. 9, FIG. 9 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. As shown in FIG. 9, it includes: a memory 902, a processor 901 and a A computer program for the archiving method of face data running on 901, wherein:

The processor 901 is used to call the computer program stored in the memory 902, and perform the following steps:

Perform archiving processing on the first human face data set through the archived human face data in the preset archiving library.

Optionally, the method executed by the processor 901 further includes:

Optionally, the multiple archived face datasets include a frontal face dataset and multiple deflected face datasets, the facial pose features include face angle feature values, and the processor 901 executes the according to The facial features and facial posture features of the archived facial data are constructed to obtain the preset threshold matrix, including:

Optionally, one average similarity matrix unit in the average similarity matrix executed by the processor 901 corresponds to an average similarity value, and the calculation according to the average similarity matrix to obtain the preset threshold matrix includes:

The final similarity matrix is used as the preset threshold matrix.

Optionally, the archiving processing of the first face data set by using the archived face data in the preset archiving library performed by the processor 901 includes:

Optionally, the second face relation graph includes a first face relation cluster, and the process executed by the processor 901 is to process the second face data through the archived face data in the preset archive repository. collections for archiving, including:

The batch of face data is archived based on the second face relationship graph.

Optionally, the second face relation graph includes a first face relation cluster, and the archiving of the batch of face data based on the second face relation graph executed by the processor 901 includes:

The batch of face data is archived based on the second face relation cluster.

The electronic device provided by the embodiment of the present invention can realize each process realized by the face data archiving method in the above method embodiment, and can achieve the same beneficial effect. To avoid repetition, it is not repeated here.

The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method for archiving face data or the application terminal provided by the embodiment of the present invention is realized. The various processes of the face data archiving method can achieve the same technical effect, so in order to avoid repetition, details will not be repeated here.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented through computer programs to instruct related hardware, and the programs can be stored in a computer-readable storage medium. During execution, it may include the processes of the embodiments of the above-mentioned methods. Wherein, the storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM) or a random access memory (Random Access Memory, RAM for short).

Claims

A method for archiving face data, comprising the following steps:

Obtain batches of face data to be archived, the face data including face features and face gesture features;

Calculate the pairwise similarity between all facial features to obtain the global similarity of each face data;

Threshold screening is performed on the global similarity of each face data according to the facial pose feature and a preset similarity threshold matrix to obtain a first human face data set, and the preset similarity threshold matrix is based on Obtain the facial features and facial posture features contained in the archived face data in the preset archive library;

Perform archiving processing on the first human face data set through the archived human face data in the preset archiving library.
The archiving method of face data as claimed in claim 1, is characterized in that, described method also comprises:

Obtain the archived face data in the preset archive library, and extract the face features and facial pose features of the archived face data;

The preset threshold value matrix is constructed and obtained according to the facial features and facial pose features of the archived facial data.
The archiving method of human face data as claimed in claim 2, is characterized in that, described human face posture feature comprises human face angle feature value, and described human face feature and human face posture feature according to described archived human face data , to construct the preset threshold matrix, including:

For each archived face file in the preset archive library, the multiple archived face data sets include a frontal face data set and multiple deflected face data sets, according to the face angle feature value Dividing the face data contained in the archived face file to obtain a plurality of archived face data sets under different face angle intervals;

Calculate the average similarity between the archived face data sets according to the face features of the archived face data, or calculate the difference between the frontal face data set and the archived face data according to the face features of the archived face data. The average similarity between the deflected face datasets;

According to the average similarity between the archived face data sets, the average similarity matrix of each archived human face file is constructed, wherein, each of the archived human face files corresponds to a described average similarity matrix, so The dimension of the average similarity matrix is related to the number of the divided intervals;

The preset threshold matrix is obtained by calculating according to the average similarity matrix.
The archiving method of face data as claimed in claim 3, is characterized in that, in described average similarity matrix, one average similarity matrix unit corresponds to an average similarity value, and described according to described average similarity matrix calculation obtains The preset threshold matrix described above includes:

For all the average similarity matrices, extracting the average similarity values in the corresponding average similarity matrix units is added and averaged to obtain the final similarity matrix;

The final similarity matrix is used as the preset threshold matrix.
The archiving method of face data as claimed in claim 1, wherein the archiving of the first face data set through the archived face data in the preset archiving library includes:

Filtering the first human face data set according to the human face spatiotemporal characteristics to obtain a second human face data set;

The archiving process is performed on the second human face data set through the archived human face data in the preset archiving library.
The archiving method of human face data according to claim 5, wherein the second human face relation graph includes a first human face relation cluster, and the archived face data in the preset archiving library , archiving the second face data set, including:

Based on the second human face data set, generate and form a first human face relational graph;

Performing strong connectivity calculations on the first human face relational graph to obtain a second human face relational graph;

The batch of face data is archived based on the second face relationship graph.
The archiving method of face data according to claim 6, characterized in that, the second face relation graph includes a first face relation cluster, and the batch of people is classified based on the second face relation graph. Face data is archived, including:

Optimizing the first face relationship cluster according to a preset community discovery algorithm to obtain a second face relationship cluster;

The batch of face data is archived based on the second face relation cluster.
A kind of archiving device of face data, it is characterized in that, described device comprises:

Obtaining module, is used for obtaining batch face data to be archived, and described face data comprises face feature and face gesture feature;

Calculation module, used to calculate the pairwise similarity between all facial features, and obtain the global similarity of each facial data;

The first screening module is configured to perform threshold screening on the global similarity of each face data according to the face pose feature and a preset similarity threshold matrix to obtain a first face data set, the preset The similarity threshold matrix is obtained according to the facial features and facial posture features contained in the archived face data in the preset archive library;

The archiving module is configured to archive the first human face data set through the archived human face data in the preset archiving library.
An electronic device, characterized in that it comprises: a memory, a processor, and a computer program stored on the memory and operable on the processor, when the processor executes the computer program, the computer program according to claim 1 is realized. Steps in the archiving method of face data described in any one of to 7.
A computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the human face according to any one of claims 1 to 7 is realized. The steps in the archiving method of the data.