A face recognition method, a terminal device, and a storage medium

Abstract

The application relates to a face recognition method, a terminal device and a storage medium, the method comprising the following steps: collecting a public face image dataset and a face image dataset of a department to be checked for attendance and extracting face features; copying the face features of the department to be checked for attendance c times, and then combining the face features with the corresponding face features of the public face image dataset to form a face feature matrix F; performing dimension reduction processing on the face feature matrix F to obtain a dimension reduction matrix M and a dimension-reduced feature matrix F ld , and calculating the mean value of the dimension-reduced feature matrix F ld , thereby obtaining a dimension-reduced feature matrix F y ld ; performing dimension reduction processing on the face features of the staff to be checked for attendance to obtain a dimension-reduced feature matrix F y ld , and calculating the matching degrees of the features in F y ld . The application can improve the face recognition efficiency while reducing the storage cost of large-scale face data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of face recognition, and in particular to a face recognition method, a terminal device and a storage medium. BACKGROUND

[0002] Face recognition is the most widely used and most classic biometric application.

[0003] Face feature extraction and storage is the core step of face recognition. Face features contain individual naturalness and are an important factor in distinguishing differences between different individuals. With the advent of the Web3.0 information explosion era, data volume is growing exponentially. Units and departments hope to reduce data storage costs and improve data utilization while maintaining recognition accuracy.

[0004] Therefore, how to improve the efficiency of face recognition and reduce the storage cost of large-scale face features has become a concern for relevant companies and departments. The most direct method to improve face recognition efficiency and reduce storage cost is to improve the processing power of the processor and increase the storage capacity of the hardware. By improving the performance of the processor, the face feature matching efficiency is improved, and by distributed storage, a large amount of face feature data is stored. However, the hardware development method has technical bottlenecks and cannot keep up with the speed of today's data explosion. Moreover, it is expensive and difficult to popularize and apply. Therefore, how to improve the efficiency of face recognition and reduce the storage capacity from the software perspective has become a hot topic of concern for relevant companies and departments. By reducing the dimensionality of the data itself through software methods, valuable features are retained and redundant features are removed, which can not only reduce the feature storage cost, but also increase the inter-class difference, thereby improving the efficiency of face recognition. SUMMARY

[0005] To solve the above problems, the present application provides a face recognition method, a terminal device and a storage medium.

[0006] The specific scheme is as follows:

[0007] A face recognition method, comprising the following steps:

[0008] S1: collecting a public face image dataset and a face image dataset of a department to be checked;

[0009] S2: extracting face features of each face image data in the public face image dataset and the face image dataset of the department to be checked by a face feature extractor, respectively;

[0010] S3: copying all face features corresponding to the face image dataset of the department to be checked c times, and merging the face features corresponding to the public face image dataset to form a face feature matrix F;

[0011] S4: Dimensionality reduction is performed on the face feature matrix F to obtain a dimensionality reduction matrix M and a dimensionality-reduced feature matrix F ld , and the mean value of the dimensionality-reduced feature matrix F ld is calculated

[0012] S5: Based on the mean value and the dimensionality reduction matrix M, dimensionality reduction is performed on the face features corresponding to the face image data set of the department to be checked and a dimensionality-reduced feature matrix F corresponding to the face image data set of the department to be checked is formed y ld ;

[0013] S6: Based on the mean value and the dimensionality reduction matrix M, dimensionality reduction is performed on the face features of the personnel to be checked to obtain dimensionality-reduced features corresponding to each personnel to be checked

[0014] S7: For the dimensionality-reduced features corresponding to the personnel to be checked , a matching degree calculation is performed on each dimensionality-reduced feature of the dimensionality-reduced feature matrix F y ld , and it is determined whether the maximum matching degree is greater than a matching degree threshold value. If so, the dimensionality-reduced feature corresponding to the face image data of the department to be checked in the dimensionality-reduced feature matrix F y ld at the time of the maximum matching degree is extracted as the identity information of the personnel to be checked. Otherwise, it is determined that the personnel to be checked does not belong to the department to be checked.

[0015] Further, the determination method of the matching degree threshold value is as follows:

[0016] The company-entered face data set, the company personnel-to-be-checked face data set, and the non-company face data set are collected.

[0017] Based on the above three data sets, corresponding dimensionality-reduced feature matrices are obtained.

[0018] The dimensionality-reduced feature matrix corresponding to the non-company face data set and the company-entered face data set is matched to obtain a first matching vector P1, and the median p1 in the first matching vector P1 is extracted.

[0019] The dimensionality-reduced feature matrix corresponding to the company personnel-to-be-checked face data set and the company-entered face data set is matched to obtain a second matching vector P2, and the median p2 in the second matching vector P2 is extracted.

[0020] The mean value of p1 and p2 is taken as the matching degree threshold value.

[0021] Further, the determination method of the dimension dim to which the dimensionality reduction is performed is as follows:

[0022] Calculate the corresponding evaluation index E under different dimensions dim:

[0023] E = 1 - (ω1O TN + ω2P TP + ω3P M )

[0024] Wherein, P TN represents the probability of non-employee identification, P TP represents the probability of employee identification as non-employee, P M represents the probability of internal employee identification error, ω1, ω2, ω3 all represent weight coefficients, and ω2> ω3> ω4;

[0025] The dimension dim corresponding to the maximum evaluation index E is taken as the dimension for dimension reduction.

[0026] Further, the value range of the copy number c is [100, 200].

[0027] Further, when used for binary classification, the value range of the dimension dim for dimension reduction is set to [50, 150].

[0028] Further, when used for multi-classification, the value range of the dimension dim for dimension reduction is set to [40, 100].

[0029] A face recognition terminal device comprises a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method of the above-mentioned embodiment of the application when executing the computer program.

[0030] A computer readable storage medium stores a computer program, wherein the computer program is executed by a processor to implement the steps of the method of the above-mentioned embodiment of the application.

[0031] The technical scheme is adopted, valuable features of the face are extracted, and inter-class variation as a standard for distinguishing individuals is amplified, so that the face recognition efficiency is improved and the large-scale face data storage cost is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The flowchart of the first embodiment of the application is shown.

[0033] Figure 2 The flowchart of the matching degree threshold solving process in the embodiment is shown.

[0034] Figure 3 The face recognition process diagram in the embodiment is shown.

[0035] Figure 4 Fig. 4 shows the influence of the number of copies on the recognition result in this embodiment.

[0036] Figure 5 Fig. 5 shows the influence of the dimension of dimension reduction on the binary classification recognition result in this embodiment.

[0037] Figure 6 Fig. 6 shows the influence of the dimension of dimension reduction on the multi-classification recognition result in this embodiment.

[0038] Figure 7 Fig. 7 shows the influence of the dimension of dimension reduction on the overall recognition result in this embodiment. DETAILED DESCRIPTION

[0039] To further illustrate the embodiments, the present application provides accompanying drawings. These accompanying drawings are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can be used to explain the operating principle of the embodiments in conjunction with the related description of the specification. Those of ordinary skill in the art should understand other possible implementations and advantages of the present application in conjunction with these contents.

[0040] The present application will be further described in conjunction with the accompanying drawings and specific embodiments.

[0041] Embodiment one:

[0042] The embodiment of the present application provides a face recognition method, as shown in the figure, the method comprises the following steps: Figure 1

[0043] S1: Collecting public face image dataset and face image dataset of employees in the department to be checked.

[0044] In this embodiment, the public face image dataset is denoted as: Wherein, x i represents the i-th face image data in the public face image dataset D1.

[0045] Suppose there are N employees in the department to be checked, and the identity information of all employees is denoted as: C={NaM, ID1, ID2, …, ID N}, wherein NaM represents an employee who is not in the department. The face image dataset of the department to be checked is denoted as: Wherein, y i represents the i-th face image data in the face image dataset D2 of the department to be checked, represents the identity information of the i-th face.

[0046] S2: Extracting the face features of each face image data in the public face image dataset and the face image dataset of the department to be checked by the face feature extractor. ​

[0047] The embodiment summarizes the face feature extractor (Feature Extractor) as FE, assuming that the extracted features are n-dimensional column vectors. The face features of x i in the public face image dataset D1 are recorded as The face features of y i in the face image dataset D2 of the department to be checked are recorded as That is:

[0048]

[0049]

[0050] wherein, and are only n-dimensional column vectors.

[0051] S3: Copy all face features corresponding to the face image dataset of the department to be checked c times and merge them with the face features corresponding to the public face image dataset to form a face feature matrix F.

[0052] In this embodiment, the matrix composed of the face features corresponding to the public face image dataset is recorded as F x That is, The matrix composed of all face features corresponding to the face image dataset of the department to be checked is recorded as F y That is, Copy F y c times and merge it with F x to form a new matrix, which is recorded as the face feature matrix F, that is,

[0053] S4: Perform dimension reduction processing on the face feature matrix F to obtain a dimension reduction matrix M and a dimension-reduced feature matrix F ld , and calculate the mean value of the dimension-reduced feature matrix F ld

[0054] In this embodiment, the calculation formula of the mean value is set as:

[0055]

[0056] S5: Based on the mean value and the dimension reduction matrix M, perform dimension reduction processing on each face feature corresponding to the face image dataset of the department to be checked and form a dimension-reduced feature matrix F y ld corresponding to the face image dataset of the department to be checked, and the corresponding calculation formula is:

[0057] ​

[0058]

[0059] Where the superscript T denotes the transpose of the matrix, Representing facial features Dimensionality reduction characteristics.

[0060] S6: Based on mean With the dimensionality reduction matrix M, the facial features of the personnel to be checked are reduced in dimensionality to obtain the dimensionality reduction features corresponding to each personnel.

[0061] Let z be the facial data of a person awaiting attendance, and f be the facial features. z Its dimensionality reduction feature is denoted as f. z ld .

[0062] Record the face datasets of multiple (N3) people awaiting attendance as The facial features of each face data point are denoted as follows: Dimensionality reduction features are denoted as

[0063] S7: Dimensional reduction features for personnel awaiting attendance. Combine it with the dimensionality-reduced feature matrix F y ld For each dimensionality reduction feature, calculate the matching degree and determine whether the maximum matching degree is greater than the matching degree threshold θ. If so, extract the dimensionality reduction feature matrix F that has the maximum matching degree. y ld The facial image data of the department to be checked and recorded, corresponding to the dimensionality reduction feature, is used as the identity information of the person to be checked and recorded; otherwise, it is determined that the person to be checked and recorded does not belong to the department to be checked and recorded.

[0064] In this embodiment, the cosine similarity algorithm is used to calculate the matching degree in the dataset. Specifically, for z... k , and the characteristic matrix F y ld Each feature is matched with cosine similarity for one-to-one matching, and the maximum matching degree s is calculated. * And the facial recognition data of the corresponding attendance department

[0065] For the binary classification problem of distinguishing between company employees and non-company employees: Set a matching degree threshold θ, and use the following function to determine the face recognition result: If the matching degree reaches its maximum value s... * If the match score is greater than the matching threshold θ, it indicates that face z k and Match successful; if the maximum matching degree is less than or equal to the matching degree threshold θ, it means that face z kThe dimensionality reduction feature matrix F corresponding to the face image dataset of people not in the attendance department y ld Among them:

[0066]

[0067] Based on the recognition results, if there are r correctly classified faces, then the recognition accuracy P of the binary classification problem is... C The calculation formula is:

[0068]

[0069] Where D represents the total number of classification results.

[0070] like Figure 2 The method for determining the matching degree threshold θ in the embodiment shown is as follows:

[0071] S101: Collect the company's face data set, the company's face data set of employees awaiting attendance, and non-company face data sets;

[0072] S102: Obtain the corresponding dimensionality-reduced feature matrix based on the above three datasets;

[0073] S103: Calculate the matching degree between the dimensionality reduction feature matrix corresponding to the non-company face dataset and the company's entered face dataset to obtain the first matching vector P1, and extract the median p1 in the first matching vector P1.

[0074] S104: Calculate the matching degree between the face dataset of employees waiting to be clocked in the company and the dimensionality-reduced feature matrix corresponding to the face dataset entered by the company, and obtain the second matching vector P2. Extract the median p2 from the second matching vector P2.

[0075] S105: Use the mean of p1 and p2 as the matching threshold.

[0076] When calculating the matching degree between the non-company face dataset and the reduced feature matrix corresponding to the company's entered face dataset, the first matching vector P1 is composed of the maximum matching degree corresponding to each non-company face dataset. The maximum matching degree is the maximum value among the matching degrees of a non-company face dataset and all company-entered face datasets. The same method is used to calculate the matching degree between the company's attendance-pending face dataset and the reduced feature matrix corresponding to the company's entered face dataset, which will not be elaborated here.

[0077] like Figure 3 As shown, this addresses the multi-class classification problem for accurate identification of company employees: the dimensionality-reduced faces of employees awaiting attendance are... and company face dataset Code y Code xAnd one by one matching, in turn, select the highest degree of face matching face to be attendance, and matching code:

[0078]

[0079] Because the result severity is different due to different recognition error judgments, this embodiment also includes weighting different recognition results to obtain a comprehensive evaluation index E:

[0080] E = 1-(ω1P TN + ω2P Tp + ω3P m )

[0081] Wherein, P TN represents the probability of non-employee identifying employees, P TP represents the probability of employee identification as non-employee, P M represents the probability of internal employee identification error, ω1, ω2, ω3 all represent weight coefficients. Through analysis, the non-employee release has the greatest severity, followed by the prevention of employees from entering, and finally the internal employee identification error, so the size relationship between the weights is: ω1> ω2> ω3.

[0082] By calculating the corresponding evaluation index E under different dimensions dim, the dimension dim corresponding to the maximum evaluation index E is taken as the dimension for dimension reduction.

[0083] Experiment 1

[0084] In this embodiment, the influence of the feature replication times c on the technical effect of the embodiment in the dimension reduction process is verified by experiment. From the experiment, it is obtained that for different feature extraction algorithms and face data sets, the face recognition effect is most ideal when c is between [100, 200]. The best replication times are different when different face feature extraction algorithms are used. The best replication times of Seetaface is 145, and the best replication times of Retinaface is 165.

[0085] The experimental steps are as follows:

[0086] (1) First, divide the two face recognition data sets CASIA and LFW data sets into Positive_Sample (company input face data set), Positive_Test (company face set to be attendance), and Negative_Test (non-company face set);

[0087] (2) Perform feature matrix dimension reduction on the face data set: copy the company face feature matrix c times, combine it with the public face feature matrix, perform PCA dimension reduction, and obtain the feature matrix F;

[0088] (3) According to the method flow of the embodiment, the feature matrix F of different face data sets and the attendance face feature matrix F y ld As input, face matching is performed, and the face feature extraction algorithm uses Retinaface and Seetaface. The Retinaface model identifies a picture to obtain a 512-dimensional feature vector, and the Seetaface model identifies a picture to obtain a 2048-dimensional feature matrix.

[0089] (4) In order to investigate the influence of different copy times c on the face recognition effect, c is sequentially assigned from the interval [100, 1000] in the implementation, and then the recognition result is evaluated using the weighted accuracy, and the greater the weighted accuracy, the better the face recognition result. Here, two face data sets are tested at the same time;

[0090] (5) The experimental results of the two face data sets are shown in Figure 4 (Wherein, the left side is the Retinaface algorithm, the right side is the Seetaface algorithm, and from top to bottom, CASIA, LFW) respectively. From the results, for different face data sets, whether using Retinaface algorithm or Seetaface algorithm for face feature extraction, the influence of copy times c on face recognition results has the same trend, and the weighted accuracy reaches the maximum when c is 165 and 145 respectively (there is a certain deviation between the results of different data sets and extraction algorithms, but the results when c is in [100, 200] are very small from the maximum value), that is, the face recognition effect is most ideal when the copy times c is in [100, 200].

[0091] Experiment 2

[0092] In this embodiment, the influence of the dimension dim of the dimension reduction on the recognition accuracy of the two-class problem of company employees and non-company employees is verified by experiment. It is concluded from the experiment that for different face feature extraction algorithms and face data sets, when the dimension dim of the dimension reduction is [50, 150], the accuracy of the two-class face recognition problem is most ideal.

[0093] The experimental steps are as follows:

[0094] (1) The selected face data source picture and the face feature processing process are consistent with experiment 1;

[0095] (2) In order to investigate the influence of the dimension of the face feature matrix F dimension reduction on the accuracy of the technology, in the implementation, for the Retinaface feature extraction algorithm, the copy number c = 165 is taken, and the value of the interval [32, 512] is taken in turn to assign the dimension dim; for the Seetaface feature extraction algorithm, the copy number c = 145 is taken, and the value of the interval [32, 2048] is taken in turn to assign the dimension dim, and then the binary classification effect of the technology is evaluated. The binary classification effect evaluation mainly refers to the accuracy of the technology in correctly judging the employees and non-employees, and the larger the value is, the better the face recognition effect is;

[0096] (3) The binary classification effect evaluation results are shown in Figure 5 (Wherein, the left side is the Retinaface algorithm, the right side is the Seetaface algorithm, and from top to bottom, it is CASIA, LFW). From the results, for the Retinaface algorithm, when the dimension dim is reduced to 54, the face recognition effect is the best; for the Seetaface algorithm, when the dimension dim is reduced to 122, the face recognition effect is the best (the results of different face data sets are slightly different, but the results when the dimension dim interval is [50, 150] are not much different from the maximum value), which reflects the usability of the technology. In addition, from Figure 5 each figure in

[0097] Experiment 3

[0098] In this embodiment, the influence of the dimension dim of the dimension reduction on the accuracy of the multi-classification problem of the accurate recognition of the employees in the company is verified by experiment. From the experimental results, for different face feature extraction algorithms, when the dimension dim is [40, 100], the multi-classification face recognition accuracy is the most ideal.

[0099] The experimental steps are as follows:

[0100] (1) The selected face data source image and the face feature processing process are consistent with those of the implementation example 1;

[0101] (2) In order to investigate the influence of the dimension of the face feature matrix F dimension reduction on the accuracy, in the implementation, for the Retinaface feature extraction algorithm, the value of the interval [32, 512] is taken in turn to assign the dimension dim; for the Seetaface feature extraction algorithm, the value of the interval [32, 2048] is taken in turn to assign the dimension dim, and then the multi-classification effect of the technology is evaluated. The multi-classification effect evaluation mainly refers to the accuracy of the technology in correctly identifying the identity of the employees, and the larger the value is, the better the face recognition effect is.

[0102] (3) Multi-classification effect evaluation results are as shown in the following table Figure 6 (Wherein, the left side is the Retinaface algorithm, and the right side is the Seetaface algorithm) as shown. From the results, for the Retinaface algorithm, the face recognition effect is best when the dimension reduction dimension dim is 46; for the Seetaface algorithm, the face recognition effect is best when the dimension reduction dimension dim is 68 (the results of different face data sets are slightly biased, but the results when the dimension reduction dimension dim interval is [40, 100] do not differ much from the maximum value), and the accuracy value results can reflect that the method has high recognition accuracy.

[0103] Experiment 4:

[0104] In this embodiment, the influence of the dimension dim of dimension reduction on the overall effect of face recognition is verified by experiment. From the experimental results, for different face feature extraction algorithms and face data sets, the face recognition overall effect is most ideal when the dimension dim is [50, 150].

[0105] The experimental steps are as follows:

[0106] (1) The selected face data source image, the number of copies c, the face feature processing process, and the dimension interval setting of feature dimension reduction are consistent with experiment 2;

[0107] (2) In order to investigate the simulation effect of the technology in actual application, weighted accuracy is used as an evaluation index in implementation. According to the actual situation, non-employee identification members are given 0.5, 0.3, and 0.2, respectively, for employee identification, and the greater the weighted accuracy, the better the face recognition effect;

[0108] (3) The weighted accuracy evaluation of face recognition is as shown in the following table Figure 7 (Wherein, the left side is the Retinaface algorithm, and the right side is the Seetaface algorithm) as shown. From the results, for the Retinaface algorithm, the face recognition effect is best when the dimension reduction dimension dim is 54; for the Seetaface algorithm, the face recognition effect is best when the dimension reduction dimension dim is 122 (the results of different face data sets are slightly biased, but the results when the dimension reduction dimension dim interval is [50, 150] do not differ much from the maximum value).

[0109] The following improvement measures are adopted in the embodiment of the application:

[0110] (1) The feature dimension is determined, and too little dimension reduction will not greatly improve the inter-class change, and too much dimension reduction will easily remove valuable features, affecting the face recognition result.

[0111] (2) determining the face feature matrix F of the staff in the department to be checked y The number of copying times has an impact on the dimension reduction matrix M and the feature average value Too many copying times will lead to reduced efficiency of the recognition algorithm and redundant data storage.

[0112] The embodiments of the present application have the following technical effects:

[0113] (1) By feature dimension reduction, high-value features are retained, redundant features are removed, and the efficiency and accuracy of the face recognition algorithm are improved.

[0114] (2) By feature dimension reduction, feature selection is achieved, only high-value features are stored, and large-scale face feature storage costs are reduced.

[0115] Embodiment two:

[0116] The present application also provides a face recognition terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to realize the steps in the above method embodiments of embodiment one of the present application.

[0117] Further, as an executable solution, the face recognition terminal device can be a desktop computer, a notebook, a palm computer, and a cloud server, etc. The face recognition terminal device can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the above-mentioned composition structure of the face recognition terminal device is only an example of the face recognition terminal device, and does not constitute a limitation on the face recognition terminal device, and can include more or fewer components than the above, or combine certain components, or different components, for example, the face recognition terminal device can also include an input / output device, a network access device, a bus, etc., and the embodiments of the present application do not limit this.

[0118] Further, as an executable solution, the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA), or other programmable logic devices, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The processor is the control center of the face recognition terminal device, and connects all parts of the face recognition terminal device through various interfaces and lines.

[0119] The memory can be used to store the computer program and / or modules, and the processor realizes various functions of the face recognition terminal device by running or executing the computer program and / or modules stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application program required by a function; and the data storage area can store data created according to the use of the mobile phone and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0120] The application further provides a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to realize the steps of the method of the embodiments of the application.

[0121] The modules / units integrated in the face recognition terminal device, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, the application realizes all or part of the processes in the above-mentioned embodiment methods, and can also be completed by instructing related hardware through a computer program. The computer program can be stored in a computer readable storage medium. The computer program is executed by a processor to realize the steps of the above-mentioned various method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, an executable file or some intermediate form, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a U disk, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM) and a software distribution medium, etc.

[0122] Although the application is specifically shown and described in connection with the preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the spirit and scope of the application as defined in the appended claims.

Claims

1. A face recognition method, characterized by, The method comprises the following steps: S1: collecting a public face image dataset and a face image dataset of employees to be checked in a department; S2: extracting face features of each face image data in the public face image dataset and the face image dataset of employees to be checked in a department by a face feature extractor respectively; S3: copying all face features corresponding to the face image dataset of employees to be checked in a department c times and merging the face features with face features corresponding to the public face image dataset to form a face feature matrix F; S4: Dimensionality reduction is performed on the face feature matrix F to obtain a dimensionality reduction matrix M and a dimensionality-reduced feature matrix F ld , and the mean value of the dimensionality-reduced feature matrix F ld is calculated S5: based on the mean value and the dimension reduction matrix M, the face feature corresponding to the face image data set of the department to be checked is processed by dimension reduction, and a dimension reduction feature matrix F corresponding to the face image data set of the department to be checked is formed y ld ; S6: based on the mean value and the dimension reduction matrix M, the face features of the to-be-attended personnel are processed to obtain the corresponding dimension reduction features of each to-be-attended personnel S7: match the dimensionality reduction features corresponding to the to-be-attended personnel one by one with the dimensionality reduction features in the dimensionality reduction feature matrix F y ld y ld If yes, the face image data of the to-be-attended personnel corresponding to the dimensionality reduction feature with the maximum matching degree in the dimensionality reduction feature matrix Fis extracted as the identity information of the to-be-attended personnel; otherwise, it is determined that the to-be-attended personnel does not belong to the to-be-attended department.​ 2. The face recognition method of claim 1, wherein: The determination method of the matching degree threshold is: Collecting a company-entered face dataset, a company-employee-to-be-checked face dataset, and a non-company face dataset; Obtaining corresponding dimension-reduced feature matrices based on the company-entered face dataset, the company-employee-to-be-checked face dataset, and the non-company face dataset; Calculating matching degrees of the dimension-reduced feature matrices corresponding to the non-company face dataset and the company-entered face dataset to obtain a first matching vector P1, and extracting a median p1 in the first matching vector P1; Calculating matching degrees of the dimension-reduced feature matrices corresponding to the company-employee-to-be-checked face dataset and the company-entered face dataset to obtain a second matching vector P2, and extracting a median p2 in the second matching vector P2; Taking an average of p1 and p2 as the matching degree threshold.

3. The face recognition method of claim 1, wherein: The determination method of the dimension dim to which the dimension reduction is performed is: Calculating corresponding evaluation indexes E under different dimensions dim: E = 1 - (ω1P TN + ω2P TP + ω3P M ) wherein P TN represents the probability of identifying a non-employee as an employee, P TP represents the probability of identifying an employee as a non-employee, P M represents the probability of an internal employee identification error, and ω1, ω2, ω3 each represent a weight coefficient, and have ω1>ω2>ω3. Taking the dimension dim corresponding to the maximum evaluation index E as the dimension to which the dimension reduction is performed.

4. The face recognition method of claim 1, wherein: The value range of the copy number c is [100, 200].

5. The face recognition method of claim 1, wherein: When used for binary classification, the value range of the dimension dim to which the dimension reduction is performed is set to [50, 150].

6. The face recognition method of claim 1, wherein: When used for multi-classification, the value range of the dimension dim to which the dimension reduction is performed is set to [40, 100].

7. A face recognition terminal device, characterized by: The computer program is executed by the processor to implement the steps of the method according to any one of claims 1-6.

8. A computer-readable storage medium, the computer-readable storage medium storing a computer program, characterized in that: The computer program is executed by the processor to implement the steps of the method according to any one of claims 1-6.

Images