AED operator information tracking and monitoring method based on transform face recognition technology

Abstract

According to the AED operator information tracking and monitoring method based on the transform face recognition technology, AED equipment and various sensors are deployed in public places, and real-time collection and monitoring of the surrounding environment and the state information of the AED equipment are achieved. Data processing and analysis are carried out by utilizing a Bochuan F407 pocket machine and Raspberry Pi 4B, an intelligent positioning system is constructed, and accurate positioning and path planning navigation are provided for a user in combination with Beidou positioning, a WeChat applet and a map plug-in. The system transmits data to the server for storage and processing through the 4G communication module, realizes real-time monitoring and management at the same time, and performs face recognition by using a transform algorithm so as to improve the safety and traceability of equipment. And finally, an administrator carries out equipment maintenance and guarantee according to the monitoring condition of the system, so that the integrity and timeliness of the AED equipment are ensured.

Description

Technical Field

[0001] The present invention relates to the field of AED, and in particular to an AED operator information tracking and monitoring method based on transformer face recognition technology. Background Art

[0002] In today's society, sudden cardiac conditions such as cardiac arrest have become a serious public health issue. As a key emergency device, the automated external defibrillator (AED) plays a vital role in saving lives. However, the management and use of AED devices currently face many challenges.

[0003] With the development of cities and the growth of population, the demand for AEDs in public places is increasing. Existing AED devices are not widely distributed and their locations are unclear, making them difficult to find quickly in an emergency. At the same time, traditional AED devices lack the ability to monitor the surrounding environment and their own status in real time, and cannot timely understand the power level of the device, fall damage and other information, and may not be able to be used normally at critical moments. In addition, the safety and traceability of AED devices is also an important issue. They are prone to theft or misoperation, and there is a lack of effective means to control them.

[0004] In terms of technology, the efficiency of data transmission and processing needs to be improved. Traditional AED management systems may not be able to transmit device status information to the management end in a timely manner, resulting in administrators being unable to grasp the device operation status in real time and unable to handle abnormalities in a timely manner. Moreover, the lack of accurate positioning and navigation systems makes users spend a lot of time looking for AED devices.

[0005] In order to solve these problems, an advanced AED operator information tracking and monitoring system is urgently needed. By integrating multiple technologies, comprehensive management and monitoring of AED equipment can be achieved, the accessibility, safety and efficiency of the equipment can be improved, and strong protection for public life safety can be provided. Summary of the invention

[0006] To solve the above problems, the present invention proposes an AED operator information tracking and monitoring method based on transformer face recognition technology. The specific steps are as follows, which are characterized by:

[0007] Step 1: Deploy AED equipment and sensors. Widely deploy AED equipment in public places and install light intensity sensors, attitude sensors, power detectors, Beidou positioning modules, etc. to ensure that the equipment can collect information about the surrounding environment and its own status in real time;

[0008] Step 2: Data collection and processing: The data from each sensor is received by combining the Broadcom F407 pocket computer with the Raspberry Pi 4B, and then processed and analyzed. For example, the inertial sensor data is used to calculate the AED device fall status, and the power detector data is used to monitor the AED power level.

[0009] Step 3: Build an intelligent positioning system, use the Beidou positioning module to determine the location of the AED device, and combine the WeChat applet and map plug-in to provide users with accurate positioning and path planning navigation of the AED device;

[0010] Step 4: Data transmission and storage. The system uses the 4G communication module to transmit the collected data to the server in real time. The server stores and processes the data after receiving it. At the same time, the face picture of the Raspberry Pi is also uploaded to the server.

[0011] Step 5: Real-time monitoring and management. The management platform webpage and WeChat applet provide real-time monitoring of the AED device status. Administrators can check the device operation status at any time, automatically identify abnormal situations and issue alarm notifications;

[0012] Step 6: Face recognition and control. The system uses Raspberry Pi for face recognition. Only after successful recognition can the AED device be operated. At the same time, personnel information is recorded to improve equipment safety and traceability.

[0013] Step 7, equipment maintenance and guarantee. The administrator will arrange maintenance personnel to handle faulty equipment in a timely manner based on system monitoring conditions, or remind users to return equipment in a timely manner to ensure the integrity and timeliness of the AED equipment.

[0014] The present invention provides an AED operator information tracking and monitoring method based on transformer face recognition technology, and the technical effects of the present invention are:

[0015] 1. The present invention uses face recognition technology to ensure that only authorized personnel can operate the AED device, effectively preventing misoperation and device theft, improving the efficiency and safety of the device, and recording operator information to facilitate traceability and accountability;

[0016] 2. The present invention provides users with convenient AED device search, positioning and navigation services through WeChat applet and map plug-in, helping users to quickly find the nearest AED device, shortening rescue time, providing rescue knowledge popularization, and improving public first aid awareness and skills;

[0017] 3. The present invention realizes remote monitoring and management of AED equipment through a management platform, timely discovers and handles equipment failures, and ensures that the equipment is always in good condition. The data analysis function can help administrators optimize equipment layout and usage strategies and improve management efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the present invention;

[0019] Figure 2 It is the home page of the AED monitoring system of the present invention;

[0020] Figure 3 This is a login page diagram of the AED monitoring system of the present invention;

[0021] Figure 4 This is a diagram showing the device information page of the AED monitoring system of the present invention;

[0022] Figure 5 This is a large-screen display effect diagram of the overall data of the AED monitoring system of the present invention;

[0023] Figure 6 It is an AED position information monitoring diagram of the AED monitoring system of the present invention;

[0024] Figure 7 This is the WeChat applet interface diagram of the present invention;

[0025] Figure 8 It is the unlocking personnel information interface of the AED monitoring system of the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0027] The present invention proposes an AED operator information tracking and monitoring method based on transformer face recognition technology. The invention flow chart is as follows: Figure 1 As shown, the steps of the present invention are described in detail below.

[0028] Step 1: Deploy AED equipment and sensors. Widely deploy AED equipment in public places and install light intensity sensors, inertial sensors, power detectors, Beidou positioning modules, etc. to ensure that the equipment can collect information about the surrounding environment and its own status in real time;

[0029] Step 2: Data collection and processing: The data from each sensor is received by combining the Broadcom F407 pocket computer with the Raspberry Pi 4B, and then processed and analyzed. For example, the inertial sensor data is used to calculate the AED device fall status, and the power detector data is used to monitor the AED power level.

[0030] Step 2.1, sensor data collection:

[0031] Step 2.1.1, inertial sensor data acquisition, the MPU6050 used in the AED monitoring system is an inertial measurement unit that integrates a three-axis gyroscope and a three-axis accelerometer. It communicates with the microcontroller through a digital interface to collect the acceleration and angular velocity data of the AED device;

[0032] Step 2.1.2, power detector data collection, use H52CUP battery meter to measure the remaining power of the AED defibrillator battery, equipped with a display screen, which can display the remaining power of the battery and other related information, such as voltage, current, etc., and transmit the power information to the development board through the serial port;

[0033] Step 2.1.3, light intensity sensor data collection, use ADC to collect sensor data, when the ambient light brightness is sufficient, the AO end outputs a high level, when the ambient light brightness is low, the AO end outputs a low level, the microcontroller processes the value obtained by the sensor, and compares the value with the set threshold. If it is lower than the threshold, the system will turn on the LED light for fill light;

[0034] Step 2.2, inertial sensor data processing, the posture information obtained from the inertial sensor through the IIC bus is P = (a x ,a y ,a z ,w x ,w y ,w z ), where a x 、a y 、a z are the three-axis acceleration, w x 、w y 、w z The three-axis angular velocity is respectively, and the pitch angle θ, roll angle φ and yaw angle are calculated by the following formula respectively

[0035]

[0036] Step 2.3, data processing of the power detector, making corresponding judgments and processing according to the power feedback, and issuing a low power alarm when the received power information is lower than a certain threshold;

[0037] Step 2.4, processing the data of the light intensity sensor, when the collected light intensity data is less than the threshold, controlling the LED light to light up for supplementary light;

[0038] Through the control unit composed of Broadcom F407, the system can accurately collect various information inside the AED box: falling, power, light intensity, etc. Real-time monitoring of this information can help the system detect any potential problems or abnormalities in time, ensuring that the AED equipment is always in good working condition. The system controls the buzzer, servo unlocking and LED light strip through hardware devices to achieve remote control of the AED equipment. When the system detects that the AED device is in a low-battery state, it can automatically sound an alarm and notify relevant personnel so that the battery can be replaced or charged in time to ensure the continuous availability of the AED equipment. The remote control function can be achieved through the 4G module, making the system more flexible and scalable.

[0039] Step 3: Build an intelligent positioning system, use the Beidou positioning module to determine the location of the AED device, and combine the WeChat applet and map plug-in to provide users with accurate positioning and path planning navigation of the AED device;

[0040] The WeChat applet provides a detailed query function for the locations of all available AED boxes around the user. Users can use this function to accurately understand the location information of surrounding AED boxes, helping users to quickly find the nearest AED device in an emergency. In addition to providing basic positioning functions, the system also uses the intelligent navigation function to automatically generate the best route based on the user's current location and the location of the target AED box, guiding the user to quickly reach the destination. This intelligent navigation can greatly shorten the time users spend looking for AED devices and improve the efficiency of rescue.

[0041] Step 4: Data transmission and storage. The system uses the 4G communication module to transmit the collected data to the server in real time. The server stores and processes the data after receiving it. At the same time, the face picture of the Raspberry Pi is also uploaded to the server.

[0042] The system uses a 4G communication module to achieve stable data communication with the server. Through this communication module, the system can transmit real-time monitoring data and control commands in a timely manner. The server can instantly receive monitoring data and alarm information from the AED device and process it according to the preset logic. The face picture of the Raspberry Pi can also be uploaded to the server through the built-in communication module. The administrator can check the device status and operation status at any time through the management platform webpage to ensure the smooth progress of the rescue work.

[0043] Step 5: Real-time monitoring and management. The management platform webpage and WeChat applet provide real-time monitoring of the AED device status. Administrators can check the device operation status at any time, automatically identify abnormal situations and issue alarm notifications;

[0044] The management platform webpage provides a real-time monitoring function for the status of all AED boxes. Administrators can access the platform through the network anytime and anywhere to view the operation status of AED equipment. The system automatically identifies abnormal situations, such as equipment failure, low battery, etc. Administrators can immediately receive relevant alarm notifications and take timely treatment measures. For example, timely arrange maintenance personnel to deal with faulty equipment, or remind users to return equipment in time to ensure the integrity and timeliness of AED equipment.

[0045] The homepage of the AED monitoring system is a website developed based on Spring Boot backend and Vue frontend. It provides administrators with detailed introduction and usage methods to help administrators understand the AED monitoring system and the correct usage process. By expanding the background introduction, policy reports and usage methods on the homepage of the website, users can fully understand the importance of AED smart defibrillators and the correct usage process of the AED monitoring system. The system homepage is shown in the figure below. Figure 2 shown.

[0046] The AED monitoring system webpage is configured with a login and registration page. The page provides two input boxes for the account and password, which conforms to the normal login and registration logic. A jump between the two pages is added below the login / register button to avoid the situation where the user cannot log in by directly entering the login page before registering. The login page is as follows Figure 3 Shown

[0047] The AED device sends the data collected by each module to the EMQ proxy server through the 4G module, and then the backend processes the data in the database. The front end obtains the corresponding data from the database according to the functional requirements of the back end, and displays the device's latitude and longitude information, usage times, power consumption, and usage time on the management platform. The manager can view the device's data in real time and understand the specific situation of the device in a timely manner. The device information page is shown as follows Figure 4 shown.

[0048] The AED monitoring system is equipped with a large visual data screen for administrators to view the current status of each device in real time. The large screen uses a scrolling data refresh method to monitor the status of each device and facilitate the handling of abnormal events. This large screen provides the management department with an effective visual data status display, enabling them to understand the current environmental information of the device and propose corresponding solutions and action suggestions, thereby achieving unified management capabilities for drivers.

[0049] The Visual Tools open source platform was used during the development process. The middle part of the large screen shows the distribution of AEDs, and the two sides show the charts summarizing the monitoring information of the AED system, including the number of AED uses, power status, and location. The overall data display effect is as follows: Figure 5 shown.

[0050] The AED location information web interface uses the Beidou module to achieve a more accurate positioning function. The Beidou module interacts with the map to accurately mark the location of each AED device and provide real-time longitude and latitude information. Users can find nearby AED devices through the map interface, and the Beidou module ensures the accuracy and reliability of location information. In this interface, administrators can more intuitively see the basic information of all AED devices in the area, such as longitude and latitude, number of uses, power status of AED devices, unlocking status, and light intensity of accessories, which can effectively enhance the management efficiency and coordination ability of urban administrators. The AED location information monitoring chart is as follows: Figure 6 shown.

[0051] The WeChat mini program uses Beidou positioning technology to allow users to easily view the details of all AED devices, and through the map route planning plug-in, realize navigation planning from their own location to the target AED device. Users can choose different travel methods to obtain the best route and estimated arrival time, providing a full range of AED equipment search and navigation functions. The mini program also provides rescue classes, which use text, pictures, videos and other forms to help users understand what AED is, how to operate it, and cardiopulmonary resuscitation, so that users can obtain more comprehensive first aid knowledge and skills. In addition, the mini program is also connected to the Baidu Wenxin Yiyan API, which can help users solve various sudden first aid problems. The WeChat mini program interface is shown below. Figure 7 shown.

[0052] Step 6: Face recognition and control. The system uses Raspberry Pi for face recognition. Only after successful recognition can the AED device be operated. At the same time, personnel information is recorded to improve equipment safety and traceability.

[0053] The system uses Raspberry Pi to write relevant codes for face recognition, realizing fast and accurate face recognition function. After the face information is recognized, the user operates the AED device. Through face recognition technology, the system can effectively prevent irrelevant personnel from misoperating or stealing the AED device, improving the safety and reliability of the AED box. At the same time, the system can also record the personnel information of each operation, which facilitates subsequent management and tracing.

[0054] The unlocking personnel information interface is as follows Figure 8As shown, the unlocking personnel information includes the unlocking records of administrators and ordinary users. The AED monitoring system will divide the unlocking personnel into two categories. If the facial information has been logged in to the system, it will be regarded as the administrator who regularly maintains the AED equipment. If it has not been logged in, it will be regarded as an ordinary user who uses the AED equipment. In the administrator unlocking record interface, the webpage will display the ID, unlocking time, name and face picture information of the device opened by the administrator; in the ordinary user interface, the webpage will also display the ID, unlocking time and face information of the device opened, and display the unlocking method. Through this function, the administrator can accurately trace the source of each unlocking personnel, ensure the return of the AED or the maintenance of the AED, and ensure that the public can use normal AED equipment.

[0055] The face recognition function is implemented based on the Transformer framework. The specific steps are as follows:

[0056] Step 6.1, image normalization processing, the face input image is I, the size of height, width and number of channels is H×W×C, for each pixel point (i,j,k) in the image, where i∈{1,2,…,H}, j∈{1,2,…,W}, k∈{1,2,…,C}, the image normalization formula is:

[0057]

[0058] Among them, I norm (i, j, k) is the normalized pixel point, I max is the maximum value of the pixel, I min is the minimum value of the pixel;

[0059] Step 6.2, image segmentation, divide the normalized image into multiple small blocks, each of which has a size of p×p. Assuming that the image is divided into N small blocks, we have:

[0060]

[0061] Step 6.3, linear embedding, for each small block, flatten it into a one-dimensional vector and map it to a high-dimensional feature space through linear transformation to obtain the input embedding vector x i is the input embedding vector obtained after the linear transformation of the i-th image block, which is a dimension of D. Representing a real number set with dimension D, the linear transformation can be expressed as:

[0062] x i =W embed v i

[0063] Among them, W embed is the embedding weight matrix, vi is the vector of the flattened image patch;

[0064] Step 6.4, multi-head attention mechanism;

[0065] Step 6.4.1, calculate the query Q, key K and value V matrices: For the embedding vector sequence X = [x1, x2, ..., x N ], and obtain the Q, K, and V matrices through linear transformation:

[0066] Q=XW Q

[0067] K=XW K

[0068] V=XW V

[0069] Among them, W Q , W K , W V is the weight matrix, dimension is D×d k , d k is the dimension of each attention head;

[0070] Step 6.4.2, calculate the attention score. For each attention head h, calculate the attention score:

[0071]

[0072] Among them, Q h is the query of attention head h, is the transpose of the key of the attention head h, A h is the attention score of attention head h;

[0073] Step 6.4.3, calculate the attention output. For each attention head h, calculate the attention output:

[0074] O h =A h V h

[0075] Among them, O h is the output of the attention head h, V h is the value of the attention head h;

[0076] Step 6.4.4, multi-head attention output, concatenate the outputs of multiple attention heads, and then perform linear transformation to obtain the output of the multi-head attention mechanism:

[0077] O=Concat(O1,O2,…,O H )W O

[0078] Where H is the number of attention heads, W O is the linear transformation weight matrix output by the multi-head attention mechanism;

[0079] Step 6.5, feedforward neural network, for the output O of the multi-head attention mechanism, passes through a feedforward neural network composed of two fully connected layers to output image features, and the activation function of the fully connected layer is ReLU;

[0080] Step 6.6: The feature vector extracted by the feedforward neural network is input into the classification layer. The classification layer calculates the feature into the classification output y through a fully connected layer plus a Softmax activation function. i ,y i Used to predict the category of a face, indicating the probability that the input face belongs to the i-th category;

[0081] In step 6.7, the built Transformer-based face recognition model is trained through the back propagation algorithm, and the parameters are updated using the stochastic gradient descent algorithm to obtain the trained Transformer-based face recognition model.

[0082] Step 7, equipment maintenance and guarantee. The administrator will arrange maintenance personnel to handle faulty equipment in a timely manner based on system monitoring conditions, or remind users to return equipment in a timely manner to ensure the integrity and timeliness of the AED equipment. In addition, the applicant promises that the facial recognition technology provided will strictly comply with the provisions of the "Cybersecurity Law of the People's Republic of China", "Data Security Law of the People's Republic of China", "Personal Information Protection Law of the People's Republic of China" and other relevant laws and administrative regulations. The applicant guarantees that the use of facial recognition technology will not infringe upon the user's personal privacy, will not harm the public interest, and will not be used for any activities that violate laws and regulations, social ethics or the public interest.

[0083] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. AED operator information tracking and monitoring method based on transformer face recognition technology, the specific steps are as follows, characterized in that: Step 1: Deploy AED equipment and sensors. Widely deploy AED equipment in public places and install light intensity sensors, inertial sensors, power detectors, Beidou positioning modules, etc. to ensure that the equipment can collect information about the surrounding environment and its own status in real time; Step 2: Data collection and processing: By combining the Broadcom F407 pocket computer with the Raspberry Pi 4B, the data from each sensor is received, processed and analyzed, such as solving the inertial sensor data to determine the fall of the AED device, and monitoring the AED power level through the power detector data. Step 3: Build an intelligent positioning system, use the Beidou positioning module to determine the location of the AED device, and combine the WeChat applet and map plug-in to provide users with accurate positioning and path planning navigation of the AED device; Step 4: Data transmission and storage. The system uses the 4G communication module to transmit the collected data to the server in real time. The server stores and processes the data after receiving it. At the same time, the face picture of the Raspberry Pi is also uploaded to the server. Step 5: Real-time monitoring and management. The management platform webpage and WeChat applet provide real-time monitoring of the AED device status. Administrators can check the device operation status at any time, automatically identify abnormal situations and issue alarm notifications; Step 6, face recognition and control, the system uses Raspberry Pi for face recognition. After successful recognition, the user operates the AED device and records personnel information to improve equipment safety and traceability; Step 7, equipment maintenance and guarantee. The administrator will arrange maintenance personnel to handle faulty equipment in a timely manner based on system monitoring conditions, or remind users to return equipment in a timely manner to ensure the integrity and timeliness of the AED equipment.

2. The AED operator information tracking and monitoring method based on transformer face recognition technology according to claim 1 is characterized in that: The process of data collection and processing in step 2 can be expressed as: Step 2.1, sensor data collection: Step 2.1.1, inertial sensor data acquisition, the MPU6050 used in the AED monitoring system is an inertial measurement unit that integrates a three-axis gyroscope and a three-axis accelerometer. It communicates with the microcontroller through a digital interface to collect the acceleration and angular velocity data of the AED device; Step 2.1.2, power detector data collection, use H52CUP battery meter to measure the remaining power of the AED defibrillator battery, equipped with a display screen, which can display the remaining power of the battery and other related information, such as voltage, current, etc., and transmit the power information to the development board through the serial port; Step 2.1.3, light intensity sensor data collection, use ADC to collect sensor data, when the ambient light brightness is sufficient, the AO end outputs a high level, when the ambient light brightness is low, the AO end outputs a low level, the microcontroller processes the value obtained by the sensor, and compares the value with the set threshold. If it is lower than the threshold, the system will turn on the LED light for fill light; Step 2.2, inertial sensor data processing, the posture information obtained from the inertial sensor through the IIC bus is P = (a x ,a y ,a z ,w x ,w y ,w z ), where a x 、a y 、a z are the three-axis acceleration, w x 、w y 、w z The three-axis angular velocity is respectively, and the pitch angle θ, roll angle φ and yaw angle are calculated by the following formula respectively Step 2.3, data processing of the power detector, making corresponding judgments and processing according to the power feedback, and issuing a low power alarm when the received power information is lower than a certain threshold; Step 2.4, processing the data of the light intensity sensor, when the collected light intensity data is less than the threshold, control the LED light to light up for fill light.

3. The AED operator information tracking and monitoring method based on transformer face recognition technology according to claim 1 is characterized in that: The construction of the intelligent positioning system in step 3 is as follows: The WeChat mini program provides a detailed query function for the locations of all usable AED boxes around the user. Users can use this function to accurately understand the location information of surrounding AED boxes, helping users to quickly find the nearest AED device in an emergency. In addition to providing basic positioning functions, the system also uses intelligent navigation functions to automatically generate the best route based on the user's current location and the location of the target AED box, guiding the user to quickly reach the destination. This intelligent navigation can greatly shorten the time users spend looking for AED devices and improve rescue efficiency.

4. The AED operator information tracking and monitoring method based on transformer face recognition technology according to claim 1 is characterized in that: The data transmission and storage in step 4 are represented as follows: The system uses a 4G communication module to achieve stable data communication with the server. Through this communication module, the system can transmit real-time monitoring data and control commands in a timely manner. The server can instantly receive monitoring data and alarm information from the AED device and process it according to the preset logic. The Raspberry Pi's facial images can also be uploaded to the server through its own communication module. The administrator can check the device status and operation status at any time through the management platform webpage to ensure the smooth progress of the rescue work.

5. The AED operator information tracking and monitoring method based on transformer face recognition technology according to claim 1 is characterized in that: The process of implementing the face recognition function based on the Transformer framework in step 6 is as follows: Step 6.1, image normalization processing, the face input image is I, the size of height, width and number of channels is H×W×C, for each pixel point (i,j,k) in the image, where i∈{1,2,…,H}, j∈{1,2,…,W}, k∈{1,2,…,C}, the image normalization formula is: Among them, I norm (i, j, k) is the normalized pixel point, I max is the maximum value of the pixel, I min is the minimum value of the pixel; Step 6.2, image segmentation, divide the normalized image into multiple small blocks, each of which has a size of p×p. Assuming that the image is divided into N small blocks, we have: Step 6.3, linear embedding, for each small block, flatten it into a one-dimensional vector and map it to a high-dimensional feature space through linear transformation to obtain the input embedding vector x i is the input embedding vector obtained after the linear transformation of the i-th image block, which is a dimension of D. Representing a real number set with dimension D, the linear transformation can be expressed as: x i =W embed v i Among them, W embed is the embedding weight matrix, v i is the vector of the flattened image patch; Step 6.4, multi-head attention mechanism; Step 6.4.1, calculate the query Q, key K and value V matrices: For the embedding vector sequence X = [x1, x2, ..., x N ], and obtain the Q, K, and V matrices through linear transformation: Q=XW Q K=XW K V=XW V Among them, W Q , W K , W V is the weight matrix, dimension is D×d k , d k is the dimension of each attention head; Step 6.4.2, calculate the attention score. For each attention head h, calculate the attention score: Among them, Q h is the query of attention head h, is the transpose of the key of the attention head h, A h is the attention score of attention head h; Step 6.4.3, calculate the attention output. For each attention head h, calculate the attention output: About h =A h In h Among them, O h is the output of the attention head h, V h is the value of the attention head h; Step 6.4.4, multi-head attention output, concatenate the outputs of multiple attention heads, and then perform linear transformation to obtain the output of the multi-head attention mechanism: O=Concat(O1,O2,…,O H )W O Where H is the number of attention heads, W O is the linear transformation weight matrix output by the multi-head attention mechanism; Step 6.5, feedforward neural network, for the output O of the multi-head attention mechanism, passes through a feedforward neural network composed of two fully connected layers to output image features, and the activation function of the fully connected layer is ReLU; Step 6.6: The feature vector extracted by the feedforward neural network is input into the classification layer. The classification layer calculates the feature into the classification output y through a fully connected layer plus a Softmax activation function. i ,y i Used to predict the category of a face, indicating the probability that the input face belongs to the i-th category; In step 6.7, the built Transformer-based face recognition model is trained through the back propagation algorithm, and the parameters are updated using the stochastic gradient descent algorithm to obtain the trained Transformer-based face recognition model.

Images