CN112001280A - Real-time online optimization face recognition system and method - Google Patents

Abstract

A real-time, online-optimizable face recognition method and system, wherein the method comprises the steps of: acquiring image data to be recognized, and analyzing the image data to obtain input data; The input data is used for face detection, wherein the application type of the input data is judged before each detection, and the corresponding reasoning branch is selected based on the application type to perform face detection to obtain the face area data in the input data; at least according to Face clarity, face brightness, face angle, and face visibility are reviewed for the face quality of the face area data; face feature extraction is performed on the face area data that has passed the face quality review and based on Face matching of HNSW to realize face recognition. The method realizes the optimization of the face detection inference route based on the scene type, and has the ability to expand the face database online to improve the recognition accuracy.

Description

A real-time, online-optimizable face recognition system and method

technical field

The invention belongs to the technical field, and in particular relates to a real-time, online-optimized face recognition system and method.

Background technique

Face recognition is a method of identification based on facial features. It has been widely used in various fields in real life due to its non-contact and intuitive advantages. However, according to different application scenarios, the quality of the face images obtained by the device will vary greatly, resulting in a decrease in the efficiency and accuracy of face recognition.

SUMMARY OF THE INVENTION

In view of the above-mentioned deficiencies of the prior art, one of the objectives of the present invention is to provide a real-time, online-optimizable face recognition system and method.

The embodiment of the present invention discloses a real-time, online-optimizable face recognition method, comprising the following steps: acquiring image data to be recognized, and analyzing the image data to obtain input data; The input data is used for face detection, wherein the application type of the input data is judged before each detection, and the corresponding reasoning branch is selected based on the application type to perform face detection to obtain the face area data in the input data; at least according to Face clarity, face brightness, face angle, and face visibility are reviewed for the face quality of the face area data; face feature extraction is performed on the face area data that has passed the face quality review and based on Face matching of HNSW to realize face recognition.

In a possible embodiment, the method further includes: adding the face features that meet the preset storage conditions to the face feature base database, and after the face features are obtained and confirmed to be stored in the database, according to the probability rule of expanding the face features Facial features are added to HNSW's multi-layer graph data structure.

In a possible embodiment, the preset storage conditions include: setting similarity as similarity, similarity threshold as t _sim , expanded similarity threshold as _{text , and text >t sim , and} the current feature has been expanded The number of records is cnt _ext . When similarity > _{text ext} and cnt _ext < 5, the facial features are stored in the library and marked as pending determination.

In a possible embodiment, the probability rule for extending face features is: suppose there are M layers in total, and the order from shallow to deep is: layer ₁ , layer ₂ ... layer _M , and the corresponding insertion probabilities are p ₁ , p ₂ …p _M , the insertion probability has the following relationship:

In a possible embodiment, the face detection includes: judging the application type of the input data; detecting faces of different sizes under feature maps of different scales, and dividing these scales into three sizes: small, medium, and large level, which corresponds to different reasoning branches under face detection; based on the application type, selectively executes the corresponding reasoning branches of the face detection network, and runs small and meduim in scenarios where the distance between faces is relatively long. Branch, run the small and large branches in the scene where the face distance is relatively close to obtain the first face area data.

In a possible embodiment, the method further includes: screening the first face region data based on the confidence to obtain second face region data.

In a possible embodiment, the data of the face area is divided into multiple blocks to count gray values respectively; the grayness and brightness of the face area are calculated according to the gray values, and the face is judged by these two indicators Brightness of the area.

In a possible embodiment, face key points are extracted through a face key point extraction algorithm, and the key point information includes the horizontal direction coordinates of the key points, the vertical direction coordinates of the key points, and the visibility of the key points; and the face angle in the vertical direction; or, count the number of key point visibility that meet the preset value, and review the face quality based on the number of key point visibility.

In a possible embodiment, at least 21 face key points are extracted.

In a possible embodiment, the data structure of HNSW graph search is constructed based on the face feature library; the HNSW vector matching algorithm is used to find the face feature with the greatest similarity.

A real-time, online-optimizable face recognition system, comprising: an input module for acquiring image data to be recognized, and analyzing the image data to obtain input data; a face detection module for The detection network performs face detection on the input data, wherein the application type of the input data is judged before each detection, and the corresponding reasoning branch is selected based on the application type to perform face detection, and the face area in the input data is obtained. data; a face quality review module, used for reviewing the face quality of the face area data based on at least face clarity, face brightness, face angle, and face visibility; a face recognition module, used for Face feature extraction and HNSW-based face matching are performed on the face region data reviewed by face quality to realize face recognition.

In a possible embodiment, it also includes an online expansion module of the face library, configured to add the face features that meet the preset storage conditions to the face feature base library, and after the face features are acquired and confirmed to be stored in the library, The face features are added to the multi-layer graph data structure of HNSW according to the probabilistic rules of augmenting face features.

In a possible embodiment, the preset storage conditions include: setting the similarity as _simjlarity , the similarity threshold as t _sim , the expanded similarity threshold as text , and _text >t _sim , and the current feature has been expanded The number of records is cnt _ext . When similarity > _{text ext} and cnt _ext < 5, the facial features are stored in the database and marked as pending.

In a possible embodiment, the probability rule for extending face features is: suppose there are M layers in total, and the order from shallow to deep is: layer ₁ , layer ₂ ... layer _M , and the corresponding insertion probabilities are p ₁ , p ₂ …p _M , the insertion probability has the following relationship:

In a possible embodiment, the face detection module is further used for: judging the application type of the input data; detecting faces of different sizes under the feature maps of different scales, and dividing these scales into small, meduim, There are three levels of large, and the levels correspond to different inference branches under the face detection based on the anchor box method; based on the application type, the corresponding branches of the face detection network are selectively executed, and the distance between the faces is farther. Run the small and meduim branches in the scene where the face distance is relatively close, and run the small and large branches in the scene where the face distance is relatively close to obtain the first face area data.

In a possible embodiment, the face detection module is further configured to: screen the first face region data based on the confidence to obtain second face region data.

In a possible embodiment, the face quality auditing module is further configured to: divide the face area data into multiple blocks to count gray values respectively; calculate the grayness and brightness of the face area according to the gray values , and use these two indicators to judge the brightness of the face area.

In a possible embodiment, the face quality review module is further configured to: extract face key points through a face key point extraction algorithm, and the key point information includes the horizontal direction coordinates of the key points, the vertical direction coordinates of the key points, and the key points. Visibility; obtain the face angle in the horizontal and vertical directions based on the key point information; or, count the number of key point visibility that meet the preset value, and review the face quality based on the number of key point visibility.

In a possible embodiment, the face recognition module is further used for: constructing a data structure for HNSW graph search based on a face feature library; and finding the face feature with the greatest similarity by using the HNSW vector matching algorithm.

A computer storage medium storing a computer program which, when executed, implements the aforementioned method.

Compared with the prior art, the present invention has the following beneficial effects:

This scheme uses a dedicated face key point detection network to extract face key points, not only outputting key point positions, but also including the visibility of key points; using 21 key points to calculate the face angle, the output angle is more accurate. The method realizes the optimization of the face detection inference route based on the scene type, and has the ability to expand the face database online to improve the recognition accuracy.

Description of drawings

1 is a flow chart of a method according to an embodiment of the present invention;

2 is a schematic diagram of a system structure according to an embodiment of the present invention;

FIG. 3 is a schematic structural diagram of a face detection according to an embodiment of the present invention;

4 is a schematic diagram of a structure of inference branch selection of a face detection model according to an embodiment of the present invention;

FIG. 5 is a schematic flowchart of a face brightness calculation process according to an embodiment of the present invention.

Detailed ways

In order to facilitate the understanding of those skilled in the art, the present invention will be further described below with reference to the embodiments and the accompanying drawings, and the contents mentioned in the embodiments are not intended to limit the present invention.

As shown in FIG. 1 , an embodiment of the present invention discloses a real-time, online-optimized face recognition method. By acquiring image data to be recognized, face detection, face quality review, and deep learning-based face feature recognition and other steps, it can realize the optimization of the face detection inference route based on the scene type, and have the ability to expand the face database online to improve the recognition accuracy. The present invention is suitable for multiple application scenarios such as access control, monitoring, and VIP identification. The method includes the following steps:

S100: Acquire image data to be identified, and analyze the image data to obtain input data. In a specific implementation, image data such as picture sequence or video stream input, offline video, etc. from the front-end application can be received.

S102, performing face detection on the input data based on a face detection network, wherein the application type of the input data is judged before each detection, and a corresponding reasoning branch is selected based on the application type to perform face detection to obtain the input data face area data in .

The application type of the input data is judged before face detection, and the application type judgment method may include but not limited to the following means: 1) the front-end application directly transmits the application identifier; 2) the type is judged based on the resolution of the picture or video; 3 ) runs online for a period of time, counts the occurrences and quality of faces of different sizes, and counts and discriminates the applicable detection branch for application input.

Detect faces of different sizes under feature maps of different scales, divide these scales into three levels of small, medium, and large, and map the levels to different inference branches under the face detection based on the anchor box method; According to the application type described above, the corresponding reasoning branch of the face detection network is selectively executed, and the small and meduim branches are run in the scene where the face is far away, and the small and large branches are run in the scene where the face is close. The first face area data. In order to achieve the purpose of reducing the amount of calculation and coarse screening based on the size of the face.

The first face region data is screened based on the confidence to obtain second face region data.

S103, at least review the face quality of the face region data according to the face definition, face brightness, face angle, and face visibility. The face detection area data here may be the second face area data.

Among them, the review of the face brightness may be to divide the face area data into multiple blocks to count the gray value respectively; calculate the grayness and brightness of the face area according to the gray value, and use these two indicators to judge the person The brightness of the face area.

Convert the image of the face area into a grayscale image and divide it into m×n areas of the same size. The size of the area x _ij in the i-th row and the j-th column is h×w, and the gray values are counted according to the following rules:

The two brightness statistics of the area x _ij are:

The brightness statistical indicators of the region are weighted and summed to obtain two face brightness indicators score _{bright_dark} , score _{bright_light} ,

score _{bright_dark} = sum(element_wise(W, summary _d ))

score _{bright_light} = sum(element_wise(W, summary _l )).

The review of face visibility can be to extract the key points of the face through the face key point extraction algorithm. The key point information includes the horizontal coordinates of the key points, the vertical coordinates of the key points, and the visibility of the key points; based on the key point information, the horizontal and The face angle in the vertical direction; or, count the number of key point visibility that meet the preset value, and review the face quality through the number of key point visibility.

S104 , perform face feature extraction and HNSW-based face matching on the face region data that has passed the face quality audit, so as to realize face recognition.

The method for extracting face features may be to align the face based on the aforementioned key points of the face, send it to a feature extraction network to extract face features, and normalize the obtained features.

The matching method of face features may be to construct the data structure of HNSW graph search based on the face feature database, and use the HNSW vector matching algorithm to find the face features with the greatest similarity.

In one embodiment, the method of the present invention further comprises:

S105, add the face features that meet the preset storage conditions to the face feature base database, and after the face features are acquired and confirmed to be stored, add the face features to the multi-layer HNSW according to the probability rule of expanding the face features Graph data structure.

Wherein, the preset storage conditions include: setting the similarity as similarity, the similarity threshold as t _sim , the expanded similarity threshold as _text , and _text >t _sim , and the number of current feature expanded records as cnt _ext , When similarity>t _ext and cnt _ext < 5, the face features are stored in the library and marked as to be determined.

Among them, the probability rule for expanding face features is: suppose there are M layers in total, and the order from shallow to deep is: layer ₁ , layer ₂ ... layer _M , the corresponding insertion probabilities are p ₁ , p ₂ ... p _M , and the insertion probability is The following relationship:

The method also includes: data output, that is, returning the detected and recognized face data to the front-end application in the form of a request or a message queue.

Corresponding to the aforementioned method, as shown in FIG. 2 , an embodiment of the present invention further discloses a real-time, online-optimizable face recognition system 10, including an input module 101, a face detection module 102, a face quality review module 103, a face The recognition module 104 , the face library online expansion module 105 and the data output module 106 .

Among them, the input module 101 is used for receiving the picture sequence or video stream input, offline video, etc. from the front-end application, and parsing it into a data structure that can be processed by the subsequent process.

The data output module 106 is used to integrate the face detection and face recognition results obtained by the face recognition module, and return them to the front-end application or message queue.

The face detection module 102 receives the data of the data input module 101, determines the application type of the input data, adopts the face detection network based on the anchor box, and selects a specific reasoning branch based on the application type to perform face detection, and obtains the face in the picture. area.

The face quality review module 103 measures whether the face quality is up to standard based on the four aspects of face area clarity, brightness, face angle and face occlusion, and discards unqualified faces.

The face recognition module 104 extracts the features of the qualified faces, and uses the HNSW vector matching algorithm based on the obtained feature vectors and the features of the face database to find the most similar faces with the query features. Among them, the face features are represented by high-dimensional feature vectors; before the system runs, the multi-layer graph data structure used by HNSW should be constructed based on the face database, or the data structure should be loaded from a cured file.

The face database online expansion module 105, based on the recognition result, similarity information, and face quality evaluation value of the face recognition module 104, automatically determines whether to store the queried face into the database.

The system corresponds to the foregoing method embodiments, and details are not described here.

The following will describe the video stream as the image data, and the specific method steps are as follows.

Input the video stream data of the multi-channel surveillance video into the input module 101, and the input module 101 parses the video stream data into frame data, and can input the frame image, frame time, and video stream unique identification ID at the frequency of every 5 frames. To the face detection module, the remaining frames are not processed for recognition.

The face detection module 102 uses the anchor box-based face detection model for face detection. This implementation method can use the Retinface face recognition algorithm, and set six anchor box sizes, namely small: 16x16, 32x32, medium: 64x64, 128x128, large: 256x256, 512x512, where the anchor box size represents the smallest detectable size. Refer to Figure 3 for the Retinaface face detection structure.

The results of face detection are screened based on the output confidence, and the remaining detection results are input to the face quality review module 103 . Filter the face detection results with confidence less than 0.9, that is, when the score _det ≥ 0.9, is the final detected face detection result.

This implementation method can adopt the method of automatic statistical discrimination of the system in the selection of inference branches of face detection, which is divided into the initial stage of system operation and the stage of enabling inference branch selection.

In the early stage of system operation, it is necessary to count the distribution of different face sizes. It is necessary to execute small, medium, and large face detection branches within a certain period of time after startup. In general, a period of peak flow of people is selected, and statistics detected by different branches are separately calculated. the number of faces received.

When the statistics reach a certain number, the inference branch selection of the face detection model is enabled. When there are many faces with small sizes in the statistical results, the small and medium branches are enabled. When there are many faces with large sizes in the statistical results, the medium and medium branches are enabled. large branch. Refer to Figure 4 for the selection of inference branches.

The detected face is sent to the face quality review module 103 . The face quality review includes four review links: clarity, brightness, face angle, and face occlusion. When all four conditions are met, the face picture can enter the recognition link.

The face sharpness is quantified by the edge detection operator. The Laplacian operator can be used to extract the face edge information, and its variance can be calculated as the face sharpness measurement value score _sharp . When the score _sharp > 360, it is considered that the face meets the sharpness requirements. .

Face brightness can use two indicators score _{bright_dark} and score _{bright_light} to measure the grayness and brightness of the face respectively. Before the calculation, the face image is converted into a grayscale image, and the image is divided into 3×3 areas of the same size, Let the size of the area x _ij in the i-th row and the j-th column be h×w, and each area is processed as follows:

The two brightness statistics of the area x _ij are:

That is, the proportion of the statistical gray value less than 50 is used as the grayness of the region, and the proportion of the statistical gray value greater than 200 is used as the brightness of the region. After obtaining the grayness and brightness of each region, a matrix is formed:

A 3×3 two-dimensional Gaussian kernel is used for weighted summation to obtain score _{bright_dark} and score _{bright_light} .

score _{bright_dark} = sum(element_wise(GuassKernel _3×3 , summary _d ))

score _{bright_light} = sum(element_wise(GuassKernel _3×3 , summary _l ))

When score _{bright_dark} < 0.4 and score _{bright_light} < 0.5 are satisfied, the face is considered to meet the brightness requirements. Refer to Figure 5 for the brightness calculation process.

To judge the face angle and face occlusion, first obtain the face key point information, and use the face key point extraction algorithm to extract 21 key points of the face. The key point information includes (x, y, score _vis ), which are the key points respectively. Horizontal coordinates, vertical coordinates of key points, and visibility of key points. Among them, score _vis ∈ [0, 1], 1 means that the key points are completely visible, and 21 key points include the left eye, right eye, nose, mouth corners and other parts.

Based on the 21 key points and their adjacent relationships, the topological structure map of the key points can be established, and the face angles angle _h and angle _v in the horizontal and vertical directions can be estimated based on the position information.

When |angle _h |<30, |angle _v |<45, the face is considered to meet the angle requirements.

Face occlusion is judged based on the visibility of key points. The number of key points with score _vis <0.3 cnt _unvis is counted. When cnt _unvis < 9, it is considered that there is no obvious occlusion of the face.

The face that meets the above four quality conditions can be recognized in the process, and the face that fails the quality review will output the position of the face and the reason for the failure to pass the quality review in the output module 106 .

The face that has passed the quality review is sent to the face recognition module 104, and the face is aligned based on the key point information. The face recognition is divided into two steps: face feature extraction and face feature matching.

The face feature extraction is based on deep learning extraction, and se-resnet50 is used as the feature extraction network. In the training stage, the face is classified and trained according to the id. In the feature extraction stage, the features of the previous layer of the classification layer are selected as the face features. The dimension is 256.

Face feature matching uses HNSW vector matching algorithm to find the face record with the greatest similarity, and cosine similarity is used for similarity calculation. The formula for calculating cosine similarity is:

a and b are the face feature vectors, that is, the closer to 0, the more similar the face features are.

Before using the HNSW vector matching algorithm, it is necessary to construct the multi-layer graph query construction in the HNSW algorithm based on the face feature database. In this implementation, a three-layer graph structure may be used, which is divided into layer ₁ , layer ₂ , and layer ₃ from shallow to deep.

Since the face library in the present invention includes the expanded face features added during the operation, they are different from the normal input face features, and the rules are different when constructing the three-layer graph structure.

When a face feature hits the face database, the online expansion of the face database can be performed if the following conditions are met: 1. similar _{a, b} < 0.1, 2. cnt _ext < 5, cnt _ext is the number of expanded records of the current feature. When these conditions are met, the query features are written into the face database and marked as pending, and are not updated to the multi-layer graph structure of HNSW for the time being. The identity is then confirmed by the front-end user, and when the identification is correct, the feature is marked as determined, and the feature is inserted into the multi-layer graph structure when the following probability rules are satisfied:

There are a total of 3 layers of graphs, sorted from shallow to deep as: layer ₁ , layer ₂ , layer ₃ , and the corresponding insertion probabilities are p ₁ , p ₂ , p ₃ , and the insertion probabilities are as follows:

Let p ₁ =0.1, p ₂ =0.3, p ₃ =0.6, and insert features into the multi-layer graph structure according to these three probabilities.

The data output module 106 returns the detected and recognized face data to the front-end application in the form of a request or a message queue.

In the several embodiments provided in this application, it should be understood that the disclosed apparatus and method may be implemented in other manners. For example, the above-described embodiments are only illustrative. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into Another system, or some features can be ignored, or not implemented.

Units described as separate components may or may not be physically separated, and components shown as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solutions in the embodiments of the present invention.

In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.

There are many specific application ways of the present invention, and the above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements can be made. These Improvements should also be considered as the protection scope of the present invention.

Claims (20)

1. A real-time online optimization-based face recognition method is characterized by comprising the following steps:

acquiring image data to be identified, and analyzing the image data to obtain input data;

carrying out face detection on the input data based on a face detection network, wherein the application type of the input data is judged before each detection, and a corresponding inference branch is selected based on the application type to carry out face detection so as to obtain face region data in the input data;

checking the face quality of the face region data at least according to face definition, face brightness, face angle and face visibility;

and performing face feature extraction and HNSW-based face matching on the face region data which passes face quality audit so as to realize face recognition.

2. The method of claim 1, further comprising adding the face features meeting the preset warehousing conditions to a face feature base, and after obtaining the face features and confirming warehousing, adding the face features to a multi-layer graph data structure of the HNSW according to a probability rule for expanding the face features.

3. The method of claim 2, wherein the pre-determined binning conditions comprise: let similarity be similarity, and similarity threshold be t_simWith the extended similarity threshold t_extAnd t is_ext＞t_simAnd the current feature extended record number is cnt_extWhen similarity > t_extAnd cnt_extAnd when the number is less than 5, the face features are stored in a storage and marked as to be determined.

4. A method as claimed in claim 2 or 3, wherein the probability rule for augmenting the face features is: let a total M-level graph, ordered from shallow to deep as: layer₁、layer₂…layer_MCorresponding insertion probability of p₁、p₂…p_MThe insertion probability has the following relationship:

5. the method of claim 1, wherein the face detection comprises:

judging the application type of the input data;

detecting human faces with different sizes under characteristic images with different scales, dividing the scales into three levels of small, meduim and large, and corresponding the levels to different inference branches under the human face detection; and selectively executing corresponding reasoning branches of the face detection network based on the application type, operating small and meduim branches in a scene with a longer face distance, and operating small and large branches in a scene with a shorter face distance to obtain first face region data.

6. The method of claim 5, further comprising: and screening the first face region data based on the confidence coefficient to obtain second face region data.

7. The method of claim 1, wherein the face region data is divided into a plurality of blocks of respective statistical gray values; and calculating the gray-dark degree and the brightness of the face area according to the gray value, and judging the brightness of the face area according to the two indexes.

8. The method of claim 1, wherein the face key points are extracted by a face key point extraction algorithm, and the key point information includes key point horizontal direction coordinates, key point vertical direction coordinates, and key point visibility; acquiring human face angles in the horizontal direction and the vertical direction based on the key point information; or counting the number of the visibility of the key points meeting the preset value, and auditing the face quality through the number of the visibility of the key points.

9. The method of claim 8, wherein at least 21 face key points are extracted.

10. The method of claim 1, wherein a data structure for HNSW graph search is constructed based on a face feature library; and searching the face features with the maximum similarity by adopting an HNSW vector matching algorithm.

11. A real-time, online-optimizable face recognition system, comprising:

the input module is used for acquiring image data to be identified and analyzing the image data to obtain input data;

the face detection module is used for carrying out face detection on the input data based on a face detection network, wherein the application type of the input data is judged before each detection, and a corresponding inference branch is selected based on the application type to carry out face detection so as to obtain face region data in the input data;

the face quality auditing module is used for auditing the face quality of the face area data at least according to face definition, face brightness, face angle and face visibility;

and the face recognition module is used for extracting the face features of the face region data which passes the face quality audit and carrying out face matching based on HNSW so as to realize face recognition.

12. The system of claim 11, further comprising a face library online expansion module, configured to add face features meeting preset warehousing conditions to the face feature base library, and after obtaining the face features and confirming warehousing, add the face features to the multi-layer graph data structure of HNSW according to probability rules of the expanded face features.

13. The system of claim 12, wherein the pre-defined binning conditions comprise: let similarity be similarity, and similarity threshold be t_simExpanding, expandingThe threshold value of the charging similarity is t_extAnd t is_ext＞t_simAnd the current feature extended record number is cnt_extWhen similarity > t_extAnd cnt_extAnd when the number is less than 5, the face features are stored in a storage and marked as to be determined.

14. The system of claim 12 or 13, wherein the probability rule for augmenting the face features is: let a total M-level graph, ordered from shallow to deep as: layer₁、layer₂…layer_MCorresponding insertion probability of p₁、p₂…p_MThe insertion probability has the following relationship:

15. the system of claim 11, wherein the face detection module is further to: judging the application type of the input data;

detecting human faces with different sizes under characteristic images with different scales, dividing the scales into three levels of small, meduim and large, and corresponding the levels to different inference branches under the human face detection based on an anchor box mode;

and based on the application type, selectively executing corresponding branches of the face detection network, operating small and meduim branches in a scene with a longer face distance, and operating small and large branches in a scene with a shorter face distance to obtain first face region data.

16. The system of claim 15, wherein the face detection module is further configured to: and screening the first face region data based on the confidence coefficient to obtain second face region data.

17. The system of claim 11, wherein the face quality audit module is further configured to: dividing the face region data into a plurality of blocks, and respectively counting gray values; and calculating the gray-dark degree and the brightness of the face area according to the gray value, and judging the brightness of the face area according to the two indexes.

18. The system of claim 11, wherein the face quality audit module is further to: extracting face key points by a face key point extraction algorithm, wherein key point information comprises key point horizontal direction coordinates, key point vertical direction coordinates and key point visibility; acquiring human face angles in the horizontal direction and the vertical direction based on the key point information; or counting the number of the visibility of the key points meeting the preset value, and auditing the face quality through the number of the visibility of the key points.

19. The system of claim 11, wherein the face recognition module is further configured to: constructing a data structure for HNSW graph search based on the face feature library; and searching the face features with the maximum similarity by adopting an HNSW vector matching algorithm.

20. A computer storage medium storing a computer program which, when executed, implements the method of claims 1-10.

Images