Intelligent physical examination equipment based on face recognition and physical examination method thereof

Abstract

According to the intelligent health physical examination equipment based on face recognition and the physical examination method thereof, a control host is used as a core and is connected with a first camera, a second camera, a pulse diagnosis device, a blood pressure detection wrist strap, a main display screen, an auxiliary display screen and a printer, and multi-dimensional health data collection and processing are achieved. Starting the control host and performing self-inspection on the equipment, acquiring a face image through a first camera, performing gray standardization, improved MTCNN detection, 3D posture correction and improved ResNet-50 feature extraction, completing identity verification and associating with an electronic health record in combination with dynamic threshold judgment; a second camera, a pulse diagnosis device and a touch screen respectively collect tongue diagnosis and face diagnosis, pulse diagnosis and inquiry diagnosis data to generate preliminary physical examination data; the health data and the identity information are bound, are uploaded to a cloud platform after being mixed and encrypted, and are locally backed up; the control host generates a health report based on feedback of the cloud platform or local data and pushes the health report to a user mobile phone and a hospital HIS system, and the cloud platform constructs a user portrait and provides personalized health services and intervention plans.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of physical examination equipment, in particular to an intelligent health physical examination equipment based on face recognition and a physical examination method thereof. BACKGROUND

[0002] With the continuous improvement of people's health awareness and the wide penetration of intelligent technology in the field of medical health, intelligent health physical examination terminals gradually become an important tool for the public to obtain convenient health services, which can break the dependence on fixed medical institutions in traditional physical examination and provide users with nearby and efficient health data collection and preliminary evaluation services.

[0003] However, the existing intelligent health physical examination terminals still have many technical defects in actual application, which are difficult to meet the needs of users for convenience, data integrity and security management. In the user identity verification link, the existing terminals mostly rely on user manual input of identity card number, mobile phone number and other information to complete identity verification, the operation process is cumbersome and prone to identity and data mismatch due to input errors, and there is a lack of efficient multi-modal identity verification means, which cannot adapt to the use habits and scene needs of different users; in the health data collection dimension, most terminals only focus on the detection of basic physiological indicators such as heart rate and blood pressure, and do not integrate the data collection capabilities of traditional medical diagnosis dimensions such as TCM four diagnoses, resulting in limited coverage of health data and difficulty in fully reflecting the physical characteristics and health status of users; in the data management and service link, the existing terminals often have problems such as loose binding of health data and user identity information, insufficient data transmission and storage security, and inability to efficiently synchronize physical examination data to the cloud platform and hospital HIS system, which is not conducive to user long-term tracking of health trends and is difficult to form linkage with medical institution diagnosis and treatment services, and lack of ability to generate targeted health advice and intervention programs based on user personalized health data, which cannot provide users with whole-cycle health management support.

[0004] These technical defects restrict the service quality and popularization of intelligent health physical examination terminals, and it is urgent to realize the integration and optimization of identity verification, data collection, security management and personalized services through technological innovation. SUMMARY

[0005] The main purpose of the present application is to provide an intelligent health physical examination equipment based on face recognition and a physical examination method thereof, which solves the problems of tedious and error-prone identity verification operation of existing intelligent health physical examination terminals, limited and incomplete health data collection dimension, loose binding of data and identity and insufficient security management, inability to efficiently link cloud platform and hospital system and lack of personalized health services.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an intelligent health check-up device based on face recognition, including a control host, the control host being connected to a main display screen and a secondary display screen, and a first camera being provided on one side of the secondary display screen for face recognition; A pulse diagnosis device is installed on one side of the control host, and a second camera is installed on the pulse diagnosis device. The second camera is used to capture images of the patient's face and tongue. The pulse diagnosis device has blood pressure monitoring wristbands on both sides; The control host is electrically connected to the second camera of the pulse diagnosis device.

[0007] In the preferred embodiment, the secondary display screen is mounted on the support table via a rotatable bracket, and the secondary display screen is larger than the main display screen.

[0008] In the preferred embodiment, the main display screen is a touch screen and is connected to the control host via a wireless transmission device.

[0009] In the preferred embodiment, the second camera is connected to the top of the pulse diagnosis device via a multi-section adjustable support arm.

[0010] In the preferred embodiment, the printer is electrically connected to the control host.

[0011] In the preferred embodiment, the method includes: S1. Start the control host terminal. The control host performs a self-test on the pulse diagnosis device, blood pressure monitoring wristband, first camera and second camera. If normal communication is displayed, the device initialization is completed. S2. The first camera verifies the user's face, and the control host verifies the legality of the user's identity. After the verification is successful, it automatically links to the user's electronic health record. S3, the pulse diagnosis device, and the second camera sequentially complete the collection of user tongue and face diagnosis, pulse diagnosis, and consultation data to generate preliminary physical examination data; S4. The control host binds the collected health data with the user's identity information, encrypts the bound data, uploads the encrypted data to the cloud platform, and stores a local backup. S5. The control host generates a health report based on the processed data fed back from the cloud platform or the data packets stored locally, and pushes the health report to the user's mobile phone and the hospital's HIS system simultaneously. The cloud platform provides personalized health services based on the user's health data.

[0012] In the preferred embodiment, in step S2: S21. Capture the user's facial image using the first camera and perform grayscale value standardization processing on the original image; S22. Adaptive illumination compensation algorithm is used to optimize image quality; S23. An improved multi-task cascaded convolutional neural network is used to realize face region detection, including three stages: P-Net, R-Net, and O-Net. An IoU weighted loss function is introduced, and the formula is: ; In the formula, Let α, β, and γ be the total loss function, and let α, β, and γ be the weighting coefficients of each loss term. For classifying losses, For bounding box regression loss, For keypoint regression loss, The area of ​​the candidate box. Let N be the average area of ​​all candidate boxes, and N be the total number of candidate boxes. P-Net constructs a multi-scale image pyramid on the preprocessed image to generate initial face candidate boxes. R-Net filters candidate bounding boxes by a face probability threshold of 0.6 and resizes them to 24×24×3, outputting the corrected offset. And correct the candidate box Then, overlapping candidate boxes are removed using non-maximum suppression with an IoU threshold of 0.5; O-Net resizes the candidate boxes to 48×48×3 and outputs precise offsets. The coordinates of 5 facial landmarks and the final face probability are used to correct and obtain the final face detection box. And processed by NMS with an IoU threshold of 0.3; S24. Pre-build a 3D face averaging model. ,in Let F be the set of coordinates of M 3D vertices, and F be the set of triangle patch indices; 2D key points output by O-Net Corresponding key points in the 3D average model Matching, establishing correspondence through perspective projection transformation. ,in This is the camera intrinsic parameter matrix. Focal length (where R is the coordinate of the image center point) R is a 3×3 rotation matrix, and t is a 3×1 translation vector; The optimal rotation matrix is ​​solved by minimizing the reprojection error using the Levenberg-Marquardt algorithm. With translation vector The objective function is optimized as follows: ; In the formula, 2D key points pixel coordinates, Let the element in the m-th row and n-th column of the rotation matrix R be... Let be the components of the translation vector t. 3D key points The coordinates; according to and Pose correction is performed on the 3D average model to obtain a frontal 3D face model. Then, a standardized frontal face image of 224×224 pixels is obtained through inverse perspective projection. ; S25, will An improved ResNet-50 neural network that integrates channel attention and spatial attention is input, and the feature map is optimized through the channel attention module, as shown in the formula: ,in σ is the Sigmoid activation function, and MLP is the input dimension. The initial feature map output by ResNet-50. This is channel-by-channel multiplication; The feature map is optimized using a spatial attention module, as shown in the formula. ,in , This is a 7×7 convolutional layer for output channel 1, with Concat being the channel concatenation operation; Will The original feature vector is obtained by inputting the GAP layer. The final face feature vector is then obtained through L2 normalization. The formula is: ; In the formula, , For the k-th component of the final eigenvector, The k-th component of the original feature vector; S26. Retrieve the pre-stored user facial feature template library from the database. N is the total number of database users. For the pre-stored feature vector of the i-th user, calculate the weight of each component based on the ratio of the inter-class variance to the intra-class variance of the feature components. , Let be the inter-class variance of the k-th feature component. Let be the intraclass variance of the k-th feature component; Calculate the feature vector of the user to be verified With template feature vector The weighted cosine similarity is calculated using the following formula: ; In the formula, For weighted cosine similarity; S27. A dynamic threshold determination mechanism is adopted, and the threshold calculation formula is as follows: ; In the formula, T is the dynamic judgment threshold. Based on the threshold, This is the threshold adjustment coefficient. ( The distance between the user feature to be verified and the most similar feature in the database is denoted as . This is the average of the minimum distances between all features in the database; like Determine the user's identity is legitimate and record the matching user ID; if The system determines that the user's identity is invalid and triggers an alarm. S28. Association of Legitimate User Electronic Health Records: Based on the matching user ID, the control host connects to the database through the database query interface to retrieve the user's pre-stored electronic health records. The retrieved record information is then compared with the verification results for consistency. The formula is as follows: ; In the formula, C is the verification coefficient. For the m-th basic information item associated with the user to be verified, This refers to the m-th basic information item of the i-th user profile in the database; If C=1, bind the face verification information to the electronic health record and store it in the temporary cache area; if C<1, trigger a second query and re-verify until the association is accurate.

[0013] In the preferred embodiment, in step S3: S31. The user's tongue and face images are captured by the second camera, and after noise reduction using an improved bilateral filtering algorithm, they are converted to the LAB color space. The U-Net++ model with an attention fusion mechanism is used to perform semantic segmentation of the tongue coating, tongue body, cheek and forehead regions, and extract feature parameters such as tongue coating color, thickness and face color. These parameters were organized into a data matrix for tongue diagnosis and facial diagnosis; S32, Pulse diagnosis device (4) The built-in piezoresistive pulse sensor collects the user's pulse waveform signal. After adaptive wavelet threshold denoising and Chebyshev type I low-pass filtering preprocessing, the "adaptive threshold + slope change detection" algorithm is used to identify key feature points in the pulse waveform and calculate the pulse position, pulse rate, pulse width and pulse pressure ratio parameters to form a pulse diagnosis data vector. S33. Collect user-filled information by touching the main display screen (5), quantify the score of each physical dimension using dimension weight and reverse question correction algorithm, and generate a radar chart reflecting the user's physical characteristics; output the total score of each dimension and the set of radar chart coordinate points as consultation data.

[0014] S34. The control host receives and associates the timestamps of the above three types of data, formats them into a JSON structure, encrypts and stores them in the local cache, and marks them as pending upload to prepare for uploading to the cloud platform.

[0015] In the preferred embodiment, in step S41: S41. Binding health data with user identity information: Extract core user identity information from the face recognition verification results, including user ID, ID card number, and verification timestamp, to form an identity information tuple; standardize the format of the identity information, convert the user ID to 32-bit UUID format, process the ID card number into a mask format that retains the first 6 and last 4 digits and replaces the middle digits with *, and convert the timestamp to ISO 8601 standard format; A dual binding mechanism of timestamp association and hash verification is adopted. The time difference between the health data collection timestamp and the identity verification timestamp is calculated. If the time difference is ≤300 seconds, it is determined that it is the same user and the same operation and the binding is allowed. Otherwise, the binding is rejected and manual review is triggered. At the same time, the health data and identity information are jointly hashed to generate a unique binding identifier. The calculation method is to use the SHA-256 algorithm to perform a hash operation on the result of string concatenation of user ID, health data collection timestamp and health data MD5 hash value. The binding identifier, processed identity information, and health data are integrated into a complete data unit. This data unit consists of four parts: Header, Identity, HealthData, and Signature. The Header contains the binding identifier, user ID, health data collection timestamp, and authentication timestamp. The Identity contains the masked ID number. The HealthData is the initially collected health data. The Signature is the result of the control host signing the binding identifier using its own RSA private key. S42. Encryption of bound data: A 256-bit AES symmetric encryption key is generated using the NIST-recommended CTR-DRBG method. During the generation process, the 128-bit initial seed provided by the hardware random number generator, the device's unique identifier, and the hash value of the current timestamp are used as input parameters. The generated AES symmetric encryption key is encrypted using the RSA public key of the cloud platform to obtain the encrypted symmetric key. The integrated complete data unit is encrypted using the AES-256-GCM mode. First, a 96-bit initialization vector and 32-bit additional authentication data are generated. Then, the complete data unit is encrypted using the AES-256-GCM algorithm in combination with the AES symmetric encryption key, the initialization vector, and the additional authentication data to obtain the encrypted health data and 128-bit authentication tag. Subsequently, the initialization vector, additional authentication data, encrypted health data, authentication tag, and encrypted symmetric key are concatenated into a byte stream to form an encrypted data packet. The checksum of the encrypted data packet is calculated by XORing the result obtained by the Cyclic Redundancy Check algorithm with the result obtained by the HMAC algorithm combined with an HMAC key independent of the AES symmetric encryption key. The checksum is then appended to the end of the encrypted data packet to form the final transmitted data. S43. Uploading encrypted data to the cloud platform: The control host establishes a TLS 1.3 encrypted connection with the cloud platform. The connection process uses the ECDHE key exchange algorithm, AES-128-GCM encryption suite, and SHA-256 hash algorithm. The control host sends the device certificate to the cloud platform. After the cloud platform verifies the validity of the certificate, it generates a random challenge value. The control host uses the device private key to sign the random challenge value with ECDSA and sends it to the cloud platform. After the cloud platform verifies the signature, it completes the device identity authentication. The final transmitted data is fragmented into 1MB chunks to obtain a fragment sequence. A hash value is calculated for each fragment using the SHA-256 algorithm, generating a fragment list. The fragment list contains the total number of fragments, the hash value of each fragment, and the SHA-256 hash value of the final transmitted data. The fragment list and each fragment are uploaded to the cloud platform sequentially. The cloud platform verifies the hash value of each fragment upon receipt. If transmission is interrupted, the control host requests the set of successfully uploaded fragment indices from the cloud platform and only re-uploads the fragments that were not successfully uploaded. After receiving all fragments, the cloud platform concatenates the fragments into complete final transmission data, verifies the checksum and hash value of the final transmission data, decrypts the integrated complete data unit, and verifies the RSA signature of the binding identifier by the control host. After all verifications are successful, a successful upload receipt is returned to the control host. The control host stores this receipt as an upload credential. The receipt includes the upload completion timestamp, the SHA-256 hash value of the final transmission data, and the content of the cloud platform signing the result of concatenating the upload completion timestamp and the hash value of the final transmission data using its own RSA public key. S44. Local Backup Storage: The final data to be transmitted is stored in the device's built-in eMMC flash memory, using the EXT4 file system, with the file naming format "Backup_User ID_Health Data Collection Timestamp.enc"; Every day at 3:00 AM, the backup data stored in the eMMC flash memory in the past 24 hours is synchronized to the external SD card. During the synchronization process, the SHA-256 algorithm is used to calculate the hash value of the backup file in the eMMC flash memory and the synchronized file in the SD card. If the two hash values ​​are the same, the synchronization is considered successful; otherwise, the file is resynchronized and the synchronization log is recorded. An automatic deletion of expired backup data is achieved using a dynamic cleanup algorithm based on user activity. The data retention period is calculated as follows: a base retention period of 30 days is added, along with the product of 180 and the natural exponential function value, where the exponent of the natural exponential function is 0.5, and the user's activity level over the past 30 days is calculated as the sum of the number of logins and health checkups performed by the user in the past 30 days, divided by 30 and multiplied by 2. The activity level ranges from 0 to 1. When the number of days a backup file is stored exceeds the calculated data retention period, the backup file is automatically deleted. Every Sunday at 2:00 AM, an integrity check is performed on all local backup data. The check is performed by comparing the previously calculated checksum with the result obtained by recalculating the encrypted data packet using the Cyclic Redundancy Check algorithm, and the result obtained by XORing the result obtained by using the HMAC-SHA256 algorithm with the corresponding HMAC key. If they match, the backup data is considered valid; otherwise, it is considered invalid. For backup files that fail the check, an attempt is made to download and replace them from the cloud platform. If the download fails, the file is marked as corrupt and the administrator is notified. S45. Full-process verification: Set up verification mechanisms at each key node of health data processing, including verifying whether the bound identifier is consistent with the result of hashing the result obtained by concatenating the user ID, health data collection timestamp, and MD5 hash value of health data using the SHA-256 algorithm; verifying the validity of the authentication tag generated in AES-256-GCM mode after data decryption; verifying the checksum of the encrypted data packet and the hash value of the final transmitted data on the cloud platform; and periodically verifying the validity of backup data locally. All verification results are recorded in an immutable audit log. The log is stored in a chained manner. Each log entry contains the current operation timestamp, operation type, operation result, and the result of hashing the previous log entry using the SHA-256 algorithm.

[0016] In the preferred embodiment, step S5: S51. The control host (6) prioritizes calling the processed data fed back by the cloud platform after multi-dimensional data comparison and verification. If it does not receive the data, it retrieves the backup data packet stored locally in encryption. S52. Load the preset template containing the user basic information area, physical examination indicator display area, health status assessment area and suggestion and guidance area, fill in the masked user identity information, and present the tongue diagnosis, face diagnosis, pulse diagnosis and consultation data in the form of structured tables, waveform graphs and parameter combination, and constitution radar chart respectively. S53. Using a decision tree model constructed from clinical data, comprehensively analyze multi-dimensional data to generate a three-level health assessment conclusion of "no obvious abnormalities", "mild abnormalities", and "requires medical intervention", and complete the generation of a health report; S54. Push reports containing 24-hour valid encrypted links and verification codes to users' mobile phones via encrypted push interfaces, support PDF download, and push concise letters containing core conclusions when the corresponding APP is not installed; push standardized report data to the hospital HIS system via medical data interfaces that conform to the HL7 FHIR standard, add data traceability identifiers, establish a secure connection through two-way authentication, and receive and store confirmation receipts. S55: The cloud platform uses user ID as an identifier to integrate current and historical physical examination data to form a user profile containing long-term health trends, and associates user age and gender with key indicators; it uses an algorithm combining collaborative filtering and rule engine to generate categorized health suggestions, including specific and actionable measures; S56. Generate a 1-3 month phased intervention plan for users with "mild abnormalities" and set regular reminders; match users who "require medical intervention" with cooperating hospital departments and generate an appointment recommendation list; push de-identified health data to the corresponding doctor's terminal in advance; regularly review and analyze health data, dynamically adjust service content, ensure data security, and allow users to manage the scope of data authorization independently.

[0017] This invention provides an intelligent health checkup device and method based on facial recognition. This device and method effectively address the shortcomings of existing health checkup terminals in areas such as identity verification, data collection, management, and services, offering significant advantages in multiple aspects. In the identity verification process, by integrating facial recognition dynamic verification, manual input of information by the user is eliminated, making operation convenient and enabling rapid identity verification. This significantly reduces the risk of identity and data mismatch due to input errors, while also adapting to different user habits, improving the flexibility and accuracy of identity verification, and ensuring the uniqueness of data ownership.

[0018] In terms of health data collection, the system integrates a traditional Chinese medicine diagnostic instrument module, enabling comprehensive collection of multi-dimensional health data, including tongue diagnosis, facial diagnosis, pulse diagnosis, and oral diagnosis. This breaks through the limitations of traditional terminals that only detect basic physiological indicators, fully reflecting the user's physical characteristics and health status, and providing more comprehensive data support for subsequent health assessments. Regarding data management and security, the system achieves precise binding of health data with user identity information. A hybrid encryption mechanism combining symmetric and asymmetric encryption ensures the security of data transmission and storage. It also supports data synchronization to cloud platforms and hospital HIS systems, facilitating long-term tracking of health trends for users and enabling integration with medical institutions' diagnostic and treatment services, thereby enhancing the value of data utilization. In terms of personalized services, the cloud platform can generate categorized and actionable health recommendations based on user health data profiles, combined with collaborative filtering and rule engines. It can develop phased health intervention plans and provide regular reminders for users with "mild abnormalities," match hospital departments and appointment services for users who "require medical intervention," and dynamically adjust service content to meet users' full-cycle health management needs. In addition, users can independently manage the scope of data authorization, balancing service personalization with data privacy and security, thereby improving the overall service quality and user experience of intelligent health check-up equipment and promoting its wider application. Attached Figure Description

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the intelligent health checkup device of the present invention; Figure 2 This is a schematic diagram of the principle of this intelligent health checkup terminal; Figure 3 This is a diagram of the face recognition module of the present invention.

[0020] In the diagram: 1. Secondary display screen; 2. First camera; 3. Second camera; 4. Pulse diagnosis device; 5. Main display screen; 6. Control host; 7. Blood pressure monitoring wristband; 8. Printer. Detailed Implementation

[0021] Example 1 like Figure 1 As shown, an intelligent health checkup device based on facial recognition includes a control host 6, which is connected to a main display screen 5 and a secondary display screen 1. A first camera 2 is provided on one side of the secondary display screen 1, and the first camera 2 is used for facial recognition. A pulse diagnosis device 4 is provided on one side of the control host 6. A second camera 3 is provided on the pulse diagnosis device 4. The second camera 3 is used to capture images of the patient's face and tongue. The pulse diagnosis device 4 has blood pressure monitoring wristbands 7 on both sides; The control host 6 is electrically connected to the pulse diagnosis device 4 and the second camera 3.

[0022] In the preferred embodiment, the secondary display screen 1 is mounted on the support table via a rotatable bracket, and the size of the secondary display screen 1 is larger than that of the main display screen 5.

[0023] In the preferred embodiment, the main display screen 5 is a touch screen and is connected to the control host 6 via a wireless transmission device.

[0024] In the preferred embodiment, the second camera 3 is connected to the top of the pulse diagnosis device 4 via a multi-section adjustable support arm.

[0025] In the preferred embodiment, the printer 8 is electrically connected to the control host 6.

[0026] The intelligent health checkup device based on facial recognition uses a control host 6 as the core control unit. The control host 6 is connected to the main display screen 5 and the secondary display screen 1. The secondary display screen 1 is equipped with a first camera 2 on one side. The core function of the first camera 2 is to complete the facial recognition operation by collecting the user's facial image, so as to provide support for subsequent identity verification.

[0027] A pulse diagnosis device 4 is installed on one side of the control host 6. The second camera 3 on the pulse diagnosis device 4 is mainly used to collect images of the patient's face and tongue, providing image data for tongue and face diagnosis in traditional Chinese medicine. Blood pressure monitoring wristbands 7 are also installed on both sides of the pulse diagnosis device 4 to assist in collecting user blood pressure data, enriching the dimensions of physical examination indicators. Simultaneously, the control host 6 maintains an electrical connection with the second camera 3 on the pulse diagnosis device 4, ensuring that the image data collected by the second camera 3 can be stably transmitted to the control host 6 for processing. In a preferred embodiment, the secondary display screen 1 is mounted on a support table using a rotatable bracket. The secondary display screen 1 is larger than the main display screen 5. The rotatable bracket allows for easy adjustment of the angle of the secondary display screen 1 to accommodate users of different heights, and the larger size also makes the displayed content clearer.

[0028] The main display screen 5 is designed as a touch screen and is connected to the control host 6 via a wireless transmission device. The touch screen design allows users to directly perform interactive operations such as clicking and filling in information, while the wireless connection reduces tangled wiring and makes the device layout more flexible. The second camera 3 is connected to the top of the pulse diagnosis device 4 via a multi-section adjustable support arm. The multi-section adjustable support arm can flexibly adjust the height, angle, and position of the second camera 3 to ensure accurate alignment with the user's face or tongue, improving the accuracy of image acquisition. In addition, the printer 8 is electrically connected to the control host 6, which can quickly print out data such as physical examination reports, making it convenient for users to keep paper reports.

[0029] Before use, adjust the angle of the secondary display screen 1 using the rotatable bracket, and adjust the position of the second camera 3 using the multi-section adjustable support arm to ensure that the device is in a state that is convenient for user operation and data collection. When using the device, the user first faces the first camera 2 on one side of the secondary display screen 1. The first camera 2 captures the user's facial image, and the control host 6 processes the facial image and completes facial recognition authentication.

[0030] After identity verification, the user places their wrists on the blood pressure monitoring wristbands 7 on both sides of the pulse diagnosis device 4, and the wristbands 7 begin collecting the user's blood pressure data. Simultaneously, following the guidance of the touchscreen on the main display screen 5, the user adjusts their head or tongue position so that the second camera 3 can clearly capture facial and tongue images. The collected image data is transmitted to the control host 6 via electrical connection. During this period, the user can operate through the touchscreen on the main display screen 5 to complete interactive processes such as filling out a consultation questionnaire. After all the physical examination data, including blood pressure, facial images, tongue images, and consultation data, are integrated and processed by the control host 6, a health report is generated. The user can view the report on the main display screen 5 or the secondary display screen 1. If a paper report is required, it can be printed out through the printer 8, which is electrically connected to the control host 6.

[0031] In terms of ease of operation, the facial recognition achieved by the first camera 2 eliminates the need for users to manually input identity information, reducing operational steps and input errors. The touchscreen design of the main display screen 5, combined with wireless connectivity, makes user interaction more intuitive and the device layout more flexible. The rotatable stand of the secondary display screen 1 and the multi-section adjustable support arm of the second camera 3 can adapt to the usage needs of different users, reducing the difficulty of device operation. In terms of data collection, the device integrates facial recognition, blood pressure detection, and tongue image acquisition. Compared with single-function medical examination devices, the collected data is more comprehensive and can more completely reflect the user's health status. From a practical point of view, the printer 8 can quickly output paper medical examination reports, which are convenient for users to keep immediately and use for subsequent medical treatment. The design of the main and secondary dual displays can not only meet the user's interactive operation needs, but also provide clearer and larger-scale content display, improving the user experience. At the same time, the stable connection between each component and the control host 6 ensures the smooth transmission and processing of data, ensuring the efficient progress of the medical examination process and providing users with convenient, comprehensive, and efficient intelligent health examination services.

[0032] Example 2 Further explanation in conjunction with Example 1, such as Figures 1-3 The method, as shown in the diagram, includes: S1. Start the control host 6 terminal. The control host 6 performs a self-test on the pulse diagnosis device 4, blood pressure detection wristband 7, first camera 2 and second camera 3. If normal communication is displayed, the device initialization is completed. S2, the first camera 2 verifies the user's face identity, and the control host 6 verifies the legality of the user's identity. After the verification is successful, it automatically links to the user's electronic health record. S3, pulse diagnosis device 4, and second camera 3 sequentially complete the collection of user tongue diagnosis and facial diagnosis, pulse diagnosis and consultation data, and generate preliminary physical examination data; S4, the control host 6 binds the collected health data with the user's identity information, encrypts the bound data, uploads the encrypted data to the cloud platform, and stores a local backup. S5 and the control host 6 generate health reports based on the processed data fed back from the cloud platform or the data packets stored locally, and push the health reports to the user's mobile phone and the hospital's HIS system simultaneously. The cloud platform provides personalized health services based on the user's health data.

[0033] The first step is the S1 device initialization process. After starting the control host 6 terminal, the control host 6 will automatically perform a comprehensive self-test on the pulse diagnosis device 4, blood pressure monitoring wristband 7, first camera 2 and second camera 3. It will detect whether each component is responding normally through the built-in communication protocol. Once all components show a normal communication status, the device initialization is completed, laying the hardware foundation for the subsequent physical examination process.

[0034] Next is the S2 identity verification step, where the first camera 2 captures the user's facial image and transmits the image data to the control host 6. The control host 6 calls the facial recognition algorithm to compare the captured facial features with the user features pre-stored in the database to verify the user's identity. After successful verification, the system automatically associates the user's corresponding electronic health record to ensure that subsequent physical examination data can accurately match the user. Then comes the S3 data acquisition step, where the pulse diagnosis device 4 and the second camera 3 work together. The second camera 3 first captures images of the user's tongue and face to complete tongue and facial diagnosis data acquisition. The pulse diagnosis device 4 then collects the user's pulse data. At the same time, it collects data on the user's diet, rest, and past medical history in conjunction with the consultation interaction process. During this period, the blood pressure monitoring wristband 7 collects the user's blood pressure data simultaneously. After all the data is integrated, preliminary physical examination data is generated and transmitted to the control host 6.

[0035] The next step is the S4 data processing and storage step. The control host 6 binds the collected health data with the user identity information verified in the S2 step, and encrypts the bound data using a hybrid encryption algorithm to ensure data security. After encryption, the encrypted data is uploaded to the cloud platform and backed up locally to prevent data loss.

[0036] Finally, the S5 report generation and service steps involve the control host 6 prioritizing the access to the deeply processed physical examination data fed back from the cloud platform. If the cloud platform fails to retrieve the data, the locally stored data package is retrieved. Based on this data, a standardized health report is generated and simultaneously pushed to the user's mobile phone and the hospital's HIS system for easy viewing by the user and subsequent diagnosis and treatment by the hospital. The cloud platform then generates personalized health suggestions, intervention plans, and other services based on the user's historical and current health data, completing the health management loop.

[0037] Before use, staff or the user activates the control host 6 terminal. The control host 6 then completes self-tests of the pulse diagnosis device 4, blood pressure monitoring wristband 7, first camera 2, and second camera 3. Once all components are confirmed to be communicating normally, the device enters standby mode. The user first faces the first camera 2, keeping their face within the shooting range, and cooperates with the first camera 2 to complete facial image acquisition. The user then waits for the control host 6 to verify their identity. Upon successful verification, the system automatically links to the user's personal electronic health record. Following the instructions of the control host 6, the user first aligns the second camera 3 with their face and tongue to complete tongue and facial diagnosis data acquisition. Then, the user places their wrist on the pulse diagnosis device 4 to complete pulse diagnosis data acquisition. Simultaneously, the user fills out a questionnaire through the terminal interface. During this process, the user slips their arm onto the blood pressure monitoring wristband 7 to complete blood pressure data acquisition. After all data acquisition is complete, the control host 6 automatically generates preliminary physical examination data. No further user intervention is required. The control host 6 will automatically bind and encrypt the health data and identity information, upload it to the cloud platform, and store a local backup. After the control host 6 generates a health report, the user can receive the report via mobile phone. The hospital's HIS system will also receive the report synchronously, and the cloud platform will also push personalized health service content for the user's reference.

[0038] In terms of process efficiency, the automatic self-check function in step S1 eliminates the need for manual inspection of each device, quickly completing initialization. Step S2's facial recognition authentication eliminates the need for manual user input, reducing operational steps and errors, and significantly improving the efficiency of starting the physical examination process and verifying identity. Regarding data security and integrity, step S4's data binding and encryption ensures accurate correspondence between health data and user identity, preventing data leakage or tampering during transmission and storage. Dual protection through local backup and cloud platform uploads avoids the risk of data loss. Step S3 integrates tongue diagnosis, facial diagnosis, pulse diagnosis, consultation, and blood pressure monitoring data, providing a more comprehensive reflection of the user's health status compared to single data collection methods. In terms of service convenience and connectivity, step S5's multi-device report push allows users to obtain reports without waiting on-site. The hospital's HIS system simultaneously receives data, facilitating subsequent diagnosis and treatment. The personalized health services provided by the cloud platform offer targeted health management suggestions, realizing a full-cycle service from physical examination to health intervention, comprehensively improving the user's physical examination experience and health management efficiency.

[0039] In the preferred embodiment, in step S2: S21. Capture the user's facial image through the first camera 2, and perform grayscale value standardization processing on the original image; S22. Adaptive illumination compensation algorithm is used to optimize image quality; S23. An improved multi-task cascaded convolutional neural network is used to realize face region detection, including three stages: P-Net, R-Net, and O-Net. An IoU weighted loss function is introduced, and the formula is: ; In the formula, Let α, β, and γ be the total loss function, and let α, β, and γ be the weighting coefficients of each loss term. For classifying losses, For bounding box regression loss, For keypoint regression loss, The area of ​​the candidate box. Let N be the average area of ​​all candidate boxes, and N be the total number of candidate boxes. P-Net constructs a multi-scale image pyramid on the preprocessed image to generate initial face candidate boxes. R-Net filters candidate bounding boxes by a face probability threshold of 0.6 and resizes them to 24×24×3, outputting the corrected offset. And correct the candidate box Then, overlapping candidate boxes are removed using non-maximum suppression with an IoU threshold of 0.5; O-Net resizes the candidate boxes to 48×48×3 and outputs precise offsets. The coordinates of 5 facial landmarks and the final face probability are used to correct and obtain the final face detection box. And processed by NMS with an IoU threshold of 0.3; S24. Pre-build a 3D face averaging model. ,in Let F be the set of coordinates of M 3D vertices, and F be the set of triangle patch indices; 2D key points output by O-Net Corresponding key points in the 3D average model Matching, establishing correspondence through perspective projection transformation. ,in This is the camera intrinsic parameter matrix. Focal length (where R is the coordinate of the image center point) R is a 3×3 rotation matrix, and t is a 3×1 translation vector; The optimal rotation matrix is ​​solved by minimizing the reprojection error using the Levenberg-Marquardt algorithm. With translation vector The objective function is optimized as follows: ; In the formula, 2D key points pixel coordinates, Let the element in the m-th row and n-th column of the rotation matrix R be... Let be the components of the translation vector t. 3D key points The coordinates; according to and Pose correction is performed on the 3D average model to obtain a frontal 3D face model. Then, a standardized frontal face image of 224×224 pixels is obtained through inverse perspective projection. ; S25, will An improved ResNet-50 neural network that integrates channel attention and spatial attention is input, and the feature map is optimized through the channel attention module, as shown in the formula: ,in σ is the Sigmoid activation function, and MLP is the input dimension. The initial feature map output by ResNet-50. This is channel-by-channel multiplication; The feature map is optimized using a spatial attention module, as shown in the formula. ,in , This is a 7×7 convolutional layer for output channel 1, with Concat being the channel concatenation operation; Will The original feature vector is obtained by inputting the GAP layer. The final face feature vector is then obtained through L2 normalization. The formula is: ; In the formula, , For the k-th component of the final eigenvector, The k-th component of the original feature vector; S26. Retrieve the pre-stored user facial feature template library from the database. N is the total number of database users. For the pre-stored feature vector of the i-th user, calculate the weight of each component based on the ratio of the inter-class variance to the intra-class variance of the feature components. , Let be the inter-class variance of the k-th feature component. Let be the intraclass variance of the k-th feature component; Calculate the feature vector of the user to be verified With template feature vector The weighted cosine similarity is calculated using the following formula: ; In the formula, For weighted cosine similarity; S27. A dynamic threshold determination mechanism is adopted, and the threshold calculation formula is as follows: ; In the formula, T is the dynamic judgment threshold. Based on the threshold, This is the threshold adjustment coefficient. ( The distance between the user feature to be verified and the most similar feature in the database is denoted as . This is the average of the minimum distances between all features in the database; like Determine the user's identity is legitimate and record the matching user ID; if The system determines that the user's identity is invalid and triggers an alarm. S28. Association of Legitimate User Electronic Health Records: Based on the matching user ID, the control host connects to the database through the database query interface to retrieve the user's pre-stored electronic health records. The retrieved record information is then compared with the verification results for consistency. The formula is as follows: ; In the formula, C is the verification coefficient. For the m-th basic information item associated with the user to be verified, This refers to the m-th basic information item of the i-th user profile in the database; If C=1, bind the face verification information to the electronic health record and store it in the temporary cache area; if C<1, trigger a second query and re-verify until the association is accurate.

[0040] In actual deployment, first ensure the normal operation of the hardware, including the first camera and the control host. The software is developed using Python, utilizing the OpenCV library for image grayscale standardization and adaptive lighting compensation; an improved multi-task cascaded convolutional neural network is built based on the PyTorch framework to achieve face region detection, with custom functions handling loss function calculation and candidate box operations; the Open3D library is used for 3D face model construction and pose correction, and the NumPy library is used for perspective projection matrix operations; an improved ResNet-50 neural network with channel and spatial attention modules is built within the PyTorch framework; a MySQL database is used to store and retrieve face feature templates and user information, with queries implemented using SQL statements; and Python code is used to write logic for dynamic threshold determination and consistency verification. All software modules are integrated into the control host operating system to ensure normal communication with the hardware. Parameters such as loss term weight coefficients and probability thresholds are tested and adjusted to ensure the system accurately and efficiently completes user face authentication and electronic health record association.

[0041] Step S2: User facial recognition authentication and file association based on the first camera. I. Overall Algorithm Framework This step employs an eight-stage progressive algorithm framework: "face image acquisition → preprocessing → detection → key point localization and alignment → feature extraction → feature comparison → legality judgment → file association." Through multi-dimensional image optimization, high-precision feature calculation, and dynamic threshold determination, it achieves accurate user identity verification. The core algorithms include an adaptive illumination compensation algorithm, an improved MTCNN face detection algorithm, an alignment algorithm based on a 3D deformation model, a ResNet-50 feature extraction algorithm with an attention mechanism, and a weighted cosine similarity comparison algorithm. Each stage uses mathematical formulas to construct a quantitative calculation model to ensure the accuracy and robustness of the verification.

[0042] II. Detailed Algorithm Steps and Application of Mathematical Formulas (I) Step 1: Face image acquisition and digital signal conversion The first camera captures the user's facial image at a resolution of 1920×1080 pixels. The optical signal is then converted into an 8-bit grayscale digital signal by a CMOS image sensor to obtain the original facial image matrix. ,in , , Represents coordinates in the image The grayscale value of the pixel at that location ranges from [0, 255].

[0043] To eliminate the impact of differences in pixel value ranges under different lighting intensities on subsequent processing, the original image is standardized in grayscale using the following formula: $ ; The original image pixel grayscale values ​​are mapped to a standardized range with a mean of 0 and a standard deviation of 1. This avoids uneven pixel value distribution caused by excessively strong or weak ambient light, providing a unified data foundation for subsequent face detection. Symbol explanation: These are the standardized pixel grayscale values. The mean of the gray values ​​of all pixels in the original image ( ), The standard deviation of all pixel grayscale values ​​in the original image ( ).

[0044] (II) Step 2: Adaptive Illumination Compensation and Image Preprocessing To address potential issues such as localized shadows and backlighting in the acquired images, an adaptive illumination compensation algorithm is employed to optimize image quality, while Gaussian filtering is used for noise reduction.

[0045] By calculating the local illumination intensity of each pixel in the image, dark areas are enhanced and bright areas are suppressed, as shown in the following formula: ; Dynamically adjusting the brightness of different areas of the image eliminates the occlusion or interference of local lighting differences on facial features, making facial contours and details clearer, and improving the accuracy of subsequent detection and feature extraction.

[0046] These are the pixel values ​​after illumination compensation. To standardize the global average gray value of the image ( ), In pixels The average gray value within a local window centered at $15×15$. (Boundary pixels are filled with mirror images). To avoid the minimum value where the denominator is 0.

[0047] Using standard deviation A two-dimensional Gaussian filter is used to denoise the image after illumination compensation, filtering out random noise. The formula is as follows: ; Among them, Gaussian kernel function The window size is $2k+1 = 5$ (i.e., $k=2$).

[0048] Smooth image noise, preserve the edge and detail features of facial features, avoid noise interference with candidate box generation in subsequent face detection algorithms, and improve detection accuracy.

[0049] These are the pixel values ​​after noise reduction. For the Gaussian kernel function at position The weight value at that point satisfies This ensures that the overall brightness of the image remains unchanged after filtering.

[0050] (III) Step 3: Improved MTCNN face detection Face region detection is achieved using an improved multi-task cascaded convolutional neural network (MTCNN), which consists of three stages: P-Net (candidate box generation), R-Net (candidate box selection), and O-Net (face box localization). The detection effect of small-sized faces is improved by introducing an IoU (Intersection over Union) weighted loss function.

[0051] Step 3.1: P-Net Stage – Generating Face Candidate Boxes P-Net is a fully convolutional neural network. Its input is a 12×12×3 image patch, and its output is the position offset of a candidate bounding box and the probability of whether it is a face. The preprocessed image... Multi-scale scaling (scaling factor of 0.709) is performed to construct an image pyramid. Each scale image is processed by P-Net to generate initial face candidate boxes. ,in The coordinates of the top left corner of the candidate box. , These are the candidate box width and height, respectively.

[0052] Step 3.2: R-Net Stage – Filtering and Refining Candidate Boxes The candidate boxes generated by P-Net are filtered according to a face probability threshold. The filtered candidate boxes are then resized to 24×24×3 and input into R-Net (convolution + pooling + fully connected network). The output is the corrected offset of the candidate boxes. Based on the updated face probabilities, the candidate box positions are adjusted according to the offset: ; Then, the non-maximum suppression (NMS) algorithm is used, with an IoU threshold of 0.5, to remove candidate boxes with an overlap higher than the threshold and retain the best candidate box.

[0053] Step 3.3: O-Net Stage – Precisely Locating Face Boundaries The candidate boxes output by R-Net are resized to 48×48×3 and input into O-Net (a deeper convolutional + pooling + fully connected network) to output more accurate candidate box offsets. Coordinates of 5 facial landmarks (eyes, nose tip, left and right corners of mouth) The final face probability is used to refine the candidate bounding boxes, and NMS is performed again (with an IoU threshold of 0.3) to obtain the final face detection bounding boxes. .

[0054] To improve the accuracy of small-sized face detection, a candidate box area weighting factor is introduced into the loss function of MTCNN, as shown in the following formula: ; By analyzing small candidate boxes (area) Smaller bounding box regression loss By assigning higher weights, the problem of low localization accuracy of small-sized faces in traditional MTCNN is solved, ensuring that faces of different sizes can be accurately detected.

[0055] For the total loss function, , , These are the weighting coefficients for each loss term; The classification loss is used (using cross-entropy loss to distinguish between human faces and non-human faces). The bounding box regression loss is calculated using smooth L1 loss to determine the offset error between the predicted and ground truth boxes. The keypoint regression loss is used (using Euclidean distance loss to optimize keypoint localization). The area of ​​the candidate box; The average area of ​​all candidate boxes ( , (Total number of candidate boxes).

[0056] (iv) Step 4: Face alignment based on 3D deformation model Based on the five facial key points output by O-Net, a 3D deformation model (3DMM) is used to geometrically align the detected faces, eliminating the influence of facial pose (such as side profile, head up, head down) on feature extraction, and obtaining a standardized frontal face image.

[0057] Step 4.1: Construct a 3D face averaging model Pre-build 3D face average model ,in for A set of coordinates of 3D vertices ( ), This is a set of indices for triangle facets.

[0058] Step 4.2: Key point mapping and model fitting Detected 2D key points (Pixel coordinates) and the corresponding 5 key points in the 3D average model Matching is performed using (3D coordinates), and the correspondence between 2D and 3D key points is established through perspective projection transformation, as shown in the following formula: ; in, This is the camera intrinsic parameter matrix. Focal length The coordinates of the image center point are obtained from the camera calibration. A 3×3 rotation matrix describing the face pose. It is a 3×1 translation vector describing the position of the face.

[0059] Step 4.3: Pose Correction and Image Alignment The optimal rotation matrix is ​​obtained by minimizing the reprojection error between 2D keypoints and projected 3D keypoints using the Levenberg-Marquardt algorithm. With translation vector The formula is as follows: ; according to and Pose correction is performed on the 3D average model to obtain a frontal 3D face model. Then, the 3D frontal model is projected onto a 2D plane using inverse perspective projection to obtain a standardized 224×224 pixel frontal face image. .

[0060] The expansion of the objective function for solving the optimal attitude parameters is as follows: ; The deviation between 2D detection key points and 3D model projection key points is quantified. By minimizing this error, the facial pose can be accurately corrected, ensuring that the aligned facial image is a standard frontal view, eliminating pose interference, and providing facial data with a unified geometric shape for subsequent feature extraction.

[0061] For reprojection error, For the first 2D key points pixel coordinates, Rotation matrix The Line number Column elements, Translation vector The amount, 3D key points The coordinates.

[0062] (v) Step 5: ResNet-50 face feature extraction with attention mechanism An improved ResNet-50 neural network that integrates channel attention and spatial attention is used to process aligned frontal face images. Deep feature extraction is performed to obtain high-dimensional, highly discriminative facial feature vectors.

[0063] Step 5.1: Feature Extraction of ResNet-50 Basic Network Will Input an improved ResNet-50 network, which contains one convolutional layer (7×7 convolution, stride 2), one max pooling layer (3×3 pooling, stride 2), four residual block groups (containing 3, 4, 6, and 3 residual blocks respectively), and one global average pooling layer (GAP). Residual connections (shortcuts) are used to address the vanishing gradient problem in deep networks, and the output is an initial feature map. ( , , ).

[0064] Step 5.2: Channel Attention Module (CAM) Optimization A channel attention module is introduced to assign weights to the channel dimensions of the initial feature map, highlighting the contribution of key channels (such as channels corresponding to facial features), as shown in the following formula: ; Among them, channel attention weight , The activation function is Sigmoid, and $MLP$ is a two-layer fully connected network (input dimension 2048, hidden dimension 128, output dimension 2048). This is a channel-by-channel multiplication.

[0065] Step 5.3: Spatial Attention Module (SAM) Optimization Building upon the channel attention output, a spatial attention module is introduced to assign weights to the spatial dimensions of the feature map, highlighting the contributions of key facial regions (such as eyes, nose, and mouth). The formula is as follows: ; Among them, spatial attention weights , This is a 7×7 convolutional layer (output channel 1), and $Concat$ is the channel concatenation operation.

[0066] Step 5.4: Feature Vector Generation Feature maps output by the spatial attention module Inputting the global average pooling layer yields a feature vector with dimension 2048. Then, through L2 normalization, the final face feature vector is obtained. .

[0067] Mathematical Formula 6: L2 Standardized Eigenvector Formula ; By mapping high-dimensional feature vectors to a unit hypersphere, the impact of overall scale differences in feature vectors on subsequent similarity calculations is eliminated, ensuring that feature vectors from different users are comparable at the same scale, thus improving the fairness and accuracy of feature comparison.

[0068] For the final feature vector, the th One portion, The original feature vector after global average pooling is the first... One portion, To avoid the minimum value where the denominator is 0.

[0069] (vi) Step 6: Feature comparison based on weighted cosine similarity Retrieve the pre-stored user facial feature template library from the database. ( The total number of users in the database. For the first (Pre-stored feature vectors of each user), calculate the feature vector of the user to be verified. With each template feature vector We use weighted cosine similarity to achieve identity matching.

[0070] Step 6.1: Calculation of Feature Component Weights Based on the importance of each component of the feature vector in distinguishing users (determined by the ratio of inter-class variance to intra-class variance of each component calculated from the training set), each feature component is assigned a weight. The formula is as follows: ; in, For the first Inter-class variance of each feature component For the first Within-class variance of each feature component.

[0071] Step 6.2: Calculate weighted cosine similarity $ $ By analyzing the feature components with high discriminative power ( (Larger) assigns higher weights to improve the distinguishing ability of similarity calculation, and solves the matching error caused by the traditional cosine similarity treating all feature components equally. It is especially suitable for distinguishing similar faces.

[0072] For the user to be verified and the first The weighted cosine similarity of database users, with a value range of... The closer the value is to 1, the higher the similarity.

[0073] (vii) Step 7: Dynamic threshold validity judgment A dynamic threshold determination mechanism is adopted instead of a fixed threshold. The determination threshold is adjusted according to the distribution density of user features in the database to avoid misjudgment or missed judgment due to uneven distribution of user features.

[0074] ; When the feature density of the database surrounding the user's feature to be verified is high ( When the threshold is small, lower the threshold. To avoid misclassifying similar users as illegitimate; when feature density is low ( When the threshold is large, increase the threshold. This avoids misjudging illegal users as legitimate ones and improves the adaptive capability of legitimacy judgment.

[0075] To dynamically determine the threshold, Based on the threshold, This is the threshold adjustment coefficient; The distance between the user feature to be verified and the most similar feature in the database ( , ); The average of the minimum distances between all features in the database ( ).

[0076] Step 7.2: Legality Judgment Logic 1. Calculation ; 2. If If the user's identity is deemed legitimate, the ID of the successfully matched user (corresponding to the database) is recorded. (The user to whom the account belongs). 3. If If the user's identity is deemed invalid, the alarm unit will be triggered (a buzzer will sound and an indicator light will flash), alerting the administrator to the abnormal situation.

[0077] (viii) Step 8: Linking the electronic health records of legitimate users The control host (6) retrieves the user's pre-stored electronic health record through the database query interface (using JDBC protocol to connect to MySQL database) based on the user ID obtained in the legality judgment step. The record contains the user's basic information (name, age, gender, ID number), historical physical examination data (past tongue diagnosis, pulse diagnosis, consultation results), health warning records, etc. The face verification information of this time is bound to the electronic health record and stored in the temporary cache area to provide an identity association basis for subsequent health examination data collection and integration.

[0078] To ensure the accuracy of the file association, a consistency check is performed between the retrieved file information and the verification results. The formula is as follows: ; in, (like )or (like ), For the user to be verified, the first one associated with facial features Basic information (such as age, gender). For the first in the database The first user profile Basic information, (Age, gender, and the last 6 digits of ID number are selected as verification items).

[0079] When the verification coefficient At that time, the accuracy of the file association was determined; when If an anomaly is detected in the file association, a secondary query (re-retrieval and verification of the file) is triggered to avoid file mismatch caused by database data errors or query interface abnormalities, thus ensuring the accurate binding of subsequent physical examination data with identity.

[0080] In the preferred embodiment, in step S3: S31. The user's tongue and face images are captured by the second camera (2), and after noise reduction by the improved bilateral filtering algorithm, they are converted to the LAB color space. The U-Net++ model with an attention fusion mechanism is used to perform semantic segmentation of the tongue coating, tongue body, cheek and forehead regions, and extract feature parameters such as tongue coating color, thickness and face color. These parameters were organized into a data matrix for tongue diagnosis and facial diagnosis; S32, Pulse diagnosis device (4) The built-in piezoresistive pulse sensor collects the user's pulse waveform signal. After adaptive wavelet threshold denoising and Chebyshev type I low-pass filtering preprocessing, the "adaptive threshold + slope change detection" algorithm is used to identify key feature points in the pulse waveform and calculate the pulse position, pulse rate, pulse width and pulse pressure ratio parameters to form a pulse diagnosis data vector. S33. Collect user-filled information by touching the main display screen (5), quantify the score of each physical dimension using dimension weight and reverse question correction algorithm, and generate a radar chart reflecting the user's physical characteristics; output the total score of each dimension and the set of radar chart coordinate points as consultation data.

[0081] S34. The control host (6) receives and associates the timestamps of the above three types of data, formats them into a JSON structure, encrypts and stores them in the local cache, and marks them as pending upload status to prepare for uploading to the cloud platform.

[0082] Step S31 involves acquiring images of the user's tongue and face using a second camera. An improved bilateral filtering algorithm is used to denoise the images, removing noise while preserving edge information. The denoised images are then converted to the LAB color space, which better separates brightness and color information. A U-Net++ model with an attention fusion mechanism is used to perform semantic segmentation on the tongue coating, tongue body, cheeks, and forehead, accurately defining each target region. Feature parameters such as tongue coating color, thickness, and facial color are extracted and organized into tongue and facial diagnosis data matrices for structured data storage. In Step S32, the piezoresistive pulse sensor built into the pulse diagnosis device acquires the user's pulse waveform signal. This signal is first processed by adaptive wavelet threshold denoising, dynamically adjusting the threshold based on signal characteristics to remove noise. Then, a Chebyshev Type I low-pass filter further filters high-frequency interference. The preprocessed signal is then analyzed using an "adaptive threshold + slope change detection" algorithm to identify key feature points in the pulse waveform, such as peaks and troughs. Based on these feature points, pulse position, pulse rate, pulse width, and pulse pressure ratio parameters are calculated to form a pulse diagnosis data vector. Step S33 collects user-entered information via the touchscreen display, assigns corresponding weights to different constitution dimensions using a dimensional weighting algorithm, and adjusts the scores of reverse-scoring questions using a reverse-question correction algorithm. This quantifies the score for each constitution dimension, generating a radar chart reflecting the user's constitution characteristics. The total score for each dimension and the radar chart coordinates are then output as the consultation data. In Step S34, the control host receives the tongue and face diagnosis data matrix, pulse diagnosis data vector, and consultation data. The timestamps of the three types of data are correlated to ensure time consistency. The integrated data is formatted into a JSON structure for easy transmission and parsing, encrypted and stored in the local cache, and marked as pending upload, ready for uploading to the cloud platform.

[0083] In actual deployment, first ensure that the hardware connections of the second camera, pulse diagnosis device, main display screen, and control host are normal and stable. On the software side, Python is used as the development language. The OpenCV library is used for image acquisition, improved bilateral filtering for denoising, and LAB color space conversion. A U-Net++ model with an attention-integrated mechanism is built based on the PyTorch or TensorFlow framework to complete region semantic segmentation and feature parameter extraction. The NumPy library is used for the construction and processing of data matrices and vectors. For pulse signal processing, the PyWavelets library is used for adaptive wavelet thresholding denoising, the signal processing module in the SciPy library is used for Chebyshev type I low-pass filtering, and a custom Python algorithm is used for key feature point identification and parameter calculation. For consultation data processing, the PyQt or Tkinter library is used to develop a touch interface for the main display screen to collect user information. Dimension weighting and inverse error correction algorithms are implemented using Python code, and the Matplotlib library is used to generate a body constitution radar map. The control host uses Python's `json` module for JSON data formatting, combines it with the `cryptography` library for data encryption, and uses SQLite or Redis for local caching. A network module is used to prepare for communication with the cloud platform. All software modules are integrated into the control host's operating system, and joint testing is conducted to ensure that data acquisition, processing, and storage operate normally and meet actual usage requirements.

[0084] In the preferred embodiment, in step S41: S41. Binding health data with user identity information: Extract core user identity information from the face recognition verification results, including user ID, ID card number, and verification timestamp, to form an identity information tuple; standardize the format of the identity information, convert the user ID to 32-bit UUID format, process the ID card number into a mask format that retains the first 6 and last 4 digits and replaces the middle digits with *, and convert the timestamp to ISO 8601 standard format; A dual binding mechanism of timestamp association and hash verification is adopted. The time difference between the health data collection timestamp and the identity verification timestamp is calculated. If the time difference is ≤300 seconds, it is determined that it is the same user and the same operation and the binding is allowed. Otherwise, the binding is rejected and manual review is triggered. At the same time, the health data and identity information are jointly hashed to generate a unique binding identifier. The calculation method is to use the SHA-256 algorithm to perform a hash operation on the result of string concatenation of user ID, health data collection timestamp and health data MD5 hash value. The binding identifier, processed identity information, and health data are integrated into a complete data unit. This data unit consists of four parts: Header, Identity, HealthData, and Signature. The Header contains the binding identifier, user ID, health data collection timestamp, and authentication timestamp. The Identity contains the masked ID number. The HealthData is the initially collected health data. The Signature is the result of the control host signing the binding identifier using its own RSA private key. S42. Encryption of bound data: A 256-bit AES symmetric encryption key is generated using the NIST-recommended CTR-DRBG method. During the generation process, the 128-bit initial seed provided by the hardware random number generator, the device's unique identifier, and the hash value of the current timestamp are used as input parameters. The generated AES symmetric encryption key is encrypted using the RSA public key of the cloud platform to obtain the encrypted symmetric key. The integrated complete data unit is encrypted using the AES-256-GCM mode. First, a 96-bit initialization vector and 32-bit additional authentication data are generated. Then, the complete data unit is encrypted using the AES-256-GCM algorithm in combination with the AES symmetric encryption key, the initialization vector, and the additional authentication data to obtain the encrypted health data and 128-bit authentication tag. Subsequently, the initialization vector, additional authentication data, encrypted health data, authentication tag, and encrypted symmetric key are concatenated into a byte stream to form an encrypted data packet. The checksum of the encrypted data packet is calculated by XORing the result obtained by the Cyclic Redundancy Check algorithm with the result obtained by the HMAC algorithm combined with an HMAC key independent of the AES symmetric encryption key. The checksum is then appended to the end of the encrypted data packet to form the final transmitted data. S43. Uploading encrypted data to the cloud platform: The control host establishes a TLS 1.3 encrypted connection with the cloud platform. The connection process uses the ECDHE key exchange algorithm, AES-128-GCM encryption suite, and SHA-256 hash algorithm. The control host sends the device certificate to the cloud platform. After the cloud platform verifies the validity of the certificate, it generates a random challenge value. The control host uses the device private key to sign the random challenge value with ECDSA and sends it to the cloud platform. After the cloud platform verifies the signature, it completes the device identity authentication. The final transmitted data is fragmented into 1MB chunks to obtain a fragment sequence. A hash value is calculated for each fragment using the SHA-256 algorithm, generating a fragment list. The fragment list contains the total number of fragments, the hash value of each fragment, and the SHA-256 hash value of the final transmitted data. The fragment list and each fragment are uploaded to the cloud platform sequentially. The cloud platform verifies the hash value of each fragment upon receipt. If transmission is interrupted, the control host requests the set of successfully uploaded fragment indices from the cloud platform and only re-uploads the fragments that were not successfully uploaded. After receiving all fragments, the cloud platform concatenates the fragments into complete final transmission data, verifies the checksum and hash value of the final transmission data, decrypts the integrated complete data unit, and verifies the RSA signature of the binding identifier by the control host. After all verifications are successful, a successful upload receipt is returned to the control host. The control host stores this receipt as an upload credential. The receipt includes the upload completion timestamp, the SHA-256 hash value of the final transmission data, and the content of the cloud platform signing the result of concatenating the upload completion timestamp and the hash value of the final transmission data using its own RSA public key. S44. Local Backup Storage: The final data to be transmitted is stored in the device's built-in eMMC flash memory, using the EXT4 file system, with the file naming format "Backup_User ID_Health Data Collection Timestamp.enc"; Every day at 3:00 AM, the backup data stored in the eMMC flash memory in the past 24 hours is synchronized to the external SD card. During the synchronization process, the SHA-256 algorithm is used to calculate the hash value of the backup file in the eMMC flash memory and the synchronized file in the SD card. If the two hash values ​​are the same, the synchronization is considered successful; otherwise, the file is resynchronized and the synchronization log is recorded. An automatic deletion of expired backup data is achieved using a dynamic cleanup algorithm based on user activity. The data retention period is calculated as follows: a base retention period of 30 days is added, along with the product of 180 and the natural exponential function value, where the exponent of the natural exponential function is 0.5, and the user's activity level over the past 30 days is calculated as the sum of the number of logins and health checkups performed by the user in the past 30 days, divided by 30 and multiplied by 2. The activity level ranges from 0 to 1. When the number of days a backup file is stored exceeds the calculated data retention period, the backup file is automatically deleted. Every Sunday at 2:00 AM, an integrity check is performed on all local backup data. The check is performed by comparing the previously calculated checksum with the result obtained by recalculating the encrypted data packet using the Cyclic Redundancy Check algorithm, and the result obtained by XORing the result obtained by using the HMAC-SHA256 algorithm with the corresponding HMAC key. If they match, the backup data is considered valid; otherwise, it is considered invalid. For backup files that fail the check, an attempt is made to download and replace them from the cloud platform. If the download fails, the file is marked as corrupt and the administrator is notified. S45. Full-process verification: Set up verification mechanisms at each key node of health data processing, including verifying whether the bound identifier is consistent with the result of hashing the result obtained by concatenating the user ID, health data collection timestamp, and MD5 hash value of health data using the SHA-256 algorithm; verifying the validity of the authentication tag generated in AES-256-GCM mode after data decryption; verifying the checksum of the encrypted data packet and the hash value of the final transmitted data on the cloud platform; and periodically verifying the validity of backup data locally. All verification results are recorded in an immutable audit log. The log is stored in a chained manner. Each log entry contains the current operation timestamp, operation type, operation result, and the result of hashing the previous log entry using the SHA-256 algorithm.

[0085] Step S41 binds health data with user identity information. First, the user ID, ID card number, and verification timestamp are extracted from the face recognition verification result to form an identity information tuple. The identity information is then standardized: the user ID is converted to a 32-bit UUID format, the ID card number is processed into a mask format where the first 6 and last 4 digits are retained and the middle digits are replaced with *, and the timestamp is converted to the ISO 8601 standard format. A dual binding mechanism of timestamp association and hash verification is adopted. The time difference between the health data collection timestamp and the identity verification timestamp is calculated. If the time difference is ≤300 seconds, it is determined that it is the same user and the same operation, and binding is allowed; otherwise, binding is rejected and manual review is triggered. At the same time, the SHA-256 algorithm is used to perform a hash operation on the result of string concatenation of user ID, health data collection timestamp, and MD5 hash value of health data to generate a unique binding identifier. The binding identifier, processed identity information, and health data are integrated into a complete data unit. This unit consists of four parts: Header, Identity, HealthData, and Signature. The Header contains the binding identifier, user ID, health data collection timestamp, and authentication timestamp. The Identity contains the masked ID number. The HealthData is the initially collected health data. The Signature is the result of the control host signing the binding identifier with its own RSA private key.

[0086] Step S42 processes the encryption of the bound data. A 256-bit AES symmetric encryption key is generated using the NIST-recommended CTR-DRBG algorithm. During generation, a 128-bit initial seed provided by a hardware random number generator, the device's unique identifier, and the hash value of the current timestamp are used as input parameters. The generated AES symmetric encryption key is encrypted using the cloud platform's RSA public key to obtain the encrypted symmetric key. The complete data unit is then encrypted using the AES-256-GCM mode. First, a 96-bit initialization vector and 32-bit additional authentication data are generated. Then, the AES symmetric encryption key, initialization vector, and additional authentication data are combined and encrypted using the AES-256-GCM algorithm to obtain the encrypted health data and a 128-bit authentication tag. Finally, the initialization vector, additional authentication data, encrypted health data, authentication tag, and encrypted symmetric key are concatenated as a byte stream to form an encrypted data packet. The checksum of the encrypted data packet is calculated. The result of the calculation is performed using the Cyclic Redundancy Check (CR) algorithm. The result is then XORed with the result of the calculation using the HMAC-SHA256 algorithm combined with an HMAC key that is independent of the AES symmetric encryption key. The checksum is then appended to the end of the encrypted data packet to form the final transmitted data.

[0087] Step S43 implements the upload of encrypted data to the cloud platform. The control host establishes a TLS 1.3 encrypted connection with the cloud platform, using the ECDHE key exchange algorithm, AES-128-GCM encryption suite, and SHA-256 hash algorithm. The control host sends the device certificate to the cloud platform. After verifying the certificate's validity, the cloud platform generates a random challenge value. The control host uses the device's private key to perform an ECDSA signature on the random challenge value and sends it to the cloud platform. After the cloud platform verifies the signature, it completes device authentication. The final transmitted data is fragmented into 1MB segments to obtain a fragment sequence. The hash value of each segment is calculated using the SHA-256 algorithm, generating a fragment list containing the total number of fragments, the hash value of each fragment, and the SHA-256 hash value of the final transmitted data. The fragment list and each fragment are uploaded sequentially. The cloud platform receives each fragment and verifies its hash value. If the transmission is interrupted, the control host requests the set of successfully uploaded fragment indices from the cloud platform and only re-uploads the fragments that were not successfully uploaded. After receiving all fragments, the cloud platform concatenates them into complete final transmission data, verifies the checksum and hash value of the final transmission data, decrypts the complete data unit, and verifies the RSA signature of the binding identifier by the control host. Once all verifications are successful, a successful upload receipt is returned to the control host. The control host stores the receipt as an upload credential. The receipt contains the upload completion timestamp, the SHA-256 hash value of the final transmission data, and the content of the signature obtained by the cloud platform using its own RSA public key to concatenate the upload completion timestamp and the hash value of the final transmission data.

[0088] Step S44 performs local backup storage, storing the final transfer data to the device's built-in eMMC flash memory using the EXT4 file system, with file names in the format "Backup_UserID_HealthDataCollectionTimestamp.enc". Every day at 3 AM, the backup data from the eMMC flash memory over the past 24 hours is synchronized to an external SD card. During synchronization, the hash values ​​of the backup files in the eMMC flash memory and the synchronized files on the SD card are calculated using the SHA-256 algorithm. If the hash values ​​match, synchronization is considered successful; otherwise, synchronization is re-established and a log is recorded. A dynamic cleanup algorithm based on user activity is used to automatically delete expired backup data. The data retention period is calculated as a 30-day base retention period plus the product of 180 and the natural exponential function value. The exponent of the natural exponential function is 0.5 multiplied by the user's activity level over the past 30 days. The user's activity level over the past 30 days is the sum of the number of logins and health checkups in the past 30 days, divided by 30 and multiplied by 2, with the activity level ranging from 0 to 1. Backup files are automatically deleted when the calculated retention period is exceeded. Every Sunday at 2 AM, an integrity check is performed on all local backup data. The checksum is compared with the result of recalculating the encrypted data packet using the Cyclic Redundancy Check algorithm and then performing an XOR operation on the encrypted data packet using the HMAC-SHA256 algorithm combined with the corresponding HMAC key. If they match, the backup data is considered valid; otherwise, it is invalid. For invalid backup files, an attempt is made to download and replace them from the cloud platform. If the download fails, the backup file is marked as corrupted and the administrator is notified.

[0089] Step S45 performs full-process verification, setting up verification mechanisms at each key node of health data processing. This includes verifying whether the bound identifier matches the result of a hash operation performed using the SHA-256 algorithm on the user ID, health data collection timestamp, and the MD5 hash value of the health data; verifying the validity of the authentication tag generated in AES-256-GCM mode after data decryption; verifying the checksum of the encrypted data packet against the hash value of the final transmitted data on the cloud platform; and periodically verifying the validity of backup data locally. All verification results are recorded in an immutable audit log, stored in a chain. Each log entry contains the current operation timestamp, operation type, operation result, and the result of a hash operation using the SHA-256 algorithm on the previous log entry.

[0090] In actual deployment, the hardware deployment is completed first, ensuring that the control host, eMMC flash memory, external SD card, and other hardware connections are normal and functional. On the software side, Python is used as the primary development language. The uuid library is used to convert user IDs to 32-bit UUID format, the datetime library handles timestamp processing according to the ISO 8601 standard format, the hashlib library performs SHA-256 hash operations and MD5 hash value calculations, the cryptography library handles RSA key generation, signature verification, and AES-256-GCM encryption and decryption operations, and the pycryptodome library assists in CTR-DRBG key generation. For network communication, the ssl library establishes TLS 1.3 encrypted connections, the requests library implements data fragmentation upload and breakpoint resume functionality, and the json library handles JSON format processing. Local storage management uses the os and shutil libraries to synchronize files between the eMMC flash memory and SD card, and the sqlite3 library is used for chained storage and management of audit logs. The scheduled tasks are configured via the schedule library to perform data synchronization at 3 AM daily and backup data integrity verification at 2 AM every Sunday. During deployment, the required software libraries are first installed on the control host. Code integrating various functional modules is written and debugged. A full-process test of data binding, encryption, uploading, and backup is conducted to verify the normal functioning of each step and data security. Once the test is passed, it is officially put into use. During operation, audit logs are checked regularly to ensure that all operations are traceable and anomalies-free.

[0091] In the preferred embodiment, step S5: S51, the control host 6 prioritizes calling the processed data fed back by the cloud platform after multi-dimensional data comparison and verification. If it does not receive the data, it retrieves the backup data packet stored locally in encryption. S52. Load the preset template containing the user basic information area, physical examination indicator display area, health status assessment area and suggestion and guidance area, fill in the masked user identity information, and present the tongue diagnosis, face diagnosis, pulse diagnosis and consultation data in the form of structured tables, waveform graphs and parameter combination, and constitution radar chart respectively. S53. Using a decision tree model constructed from clinical data, comprehensively analyze multi-dimensional data to generate a three-level health assessment conclusion of "no obvious abnormalities", "mild abnormalities", and "requires medical intervention", and complete the generation of a health report; S54. Push reports containing 24-hour valid encrypted links and verification codes to users' mobile phones via encrypted push interfaces, support PDF download, and push concise letters containing core conclusions when the corresponding APP is not installed; push standardized report data to the hospital HIS system via medical data interfaces that conform to the HL7 FHIR standard, add data traceability identifiers, establish a secure connection through two-way authentication, and receive and store confirmation receipts. S55: The cloud platform uses user ID as an identifier to integrate current and historical physical examination data to form a user profile containing long-term health trends, and associates user age and gender with key indicators; it uses an algorithm combining collaborative filtering and rule engine to generate categorized health suggestions, including specific and actionable measures; S56. Generate a 1-3 month phased intervention plan for users with "mild abnormalities" and set regular reminders; match users who "require medical intervention" with cooperating hospital departments and generate an appointment recommendation list; push de-identified health data to the corresponding doctor's terminal in advance; regularly review and analyze health data, dynamically adjust service content, ensure data security, and allow users to manage the scope of data authorization independently.

[0092] In step S51, the control host prioritizes calling the processed data fed back from the cloud platform after multi-dimensional data comparison and verification. This processed data is generated by the cloud platform after cross-verification and integration of the previously collected tongue diagnosis, facial diagnosis, pulse diagnosis, and consultation data. If the cloud platform data is not received due to network failure or other reasons, a backup data packet stored locally in encrypted form is retrieved. This backup data packet contains complete and verified physical examination data stored previously, ensuring that the health report generation has a reliable data source. Step S52 loads a preset health report template, which includes four core modules: user basic information area, physical examination indicator display area, health status assessment area, and suggestion and guidance area. The user basic information area is filled with masked user identity information to avoid privacy leaks. In the physical examination indicator display area, tongue diagnosis and facial diagnosis data are presented in a structured table, pulse diagnosis data is presented in a waveform graph with parameters, and consultation data is presented in a constitution radar chart, making the data display more intuitive.

[0093] Step S53 invokes a decision tree model built based on clinical data. This model establishes decision logic by learning from a large amount of clinical health data, comprehensively analyzes multi-dimensional data from tongue diagnosis, facial diagnosis, pulse diagnosis, and medical history taking, and generates three levels of health assessment conclusions—"no obvious abnormalities," "mild abnormalities," and "requires medical intervention"—based on the correspondence between data and health status. The health report is then generated using a template. Step S54 pushes the health report to the user's mobile phone via an encrypted push interface. The push content includes an encrypted link and verification code valid for 24 hours. After verification, the user can view the report and download it in PDF format. If the user does not have the corresponding app installed, a concise email containing the core health assessment conclusions is pushed to ensure the user receives the key information.

[0094] Meanwhile, through a medical data interface that conforms to the HL7 FHIR international medical data exchange standard, the standardized report data is pushed to the hospital's HIS system. Before pushing, a data traceability identifier containing the device number and collection timestamp is added to the data. Through two-way authentication, the control host verifies the HIS system interface certificate, and the HIS system verifies the control host device certificate to establish a secure connection. After the transmission is completed, the confirmation receipt returned by the HIS system is received and stored to ensure that the push process is traceable.

[0095] In step S55, the cloud platform uses the user ID as a unique identifier to integrate the current physical examination data with the user's historical physical examination records (if any), analyzes data trends to form a user profile containing long-term health trends, and focuses on health indicators related to that age group and gender by combining the age and gender information in the user profile. A collaborative filtering and rule engine algorithm is used to generate health suggestions. The collaborative filtering references effective health plans for groups with similar health characteristics to the current user, while the rule engine presets health intervention rules based on traditional Chinese medicine theory and modern medical knowledge. Finally, health suggestions are generated categorized into "dietary regulation," "exercise guidance," and "lifestyle adjustment," with each suggestion containing specific actionable measures.

[0096] Step S56: For users whose health assessment conclusion is "mildly abnormal," the cloud platform generates a phased health intervention plan with a cycle of 1-3 months. The plan includes weekly diet, exercise, and rest arrangements, and sets regular reminders via SMS or the app. For users "requiring medical intervention," the cloud platform matches the corresponding departments of partner hospitals based on the type of abnormal indicators, generating an appointment recommendation list containing the hospital name, department, doctor's specialty, and available appointment time slots. Simultaneously, anonymized user health data is pushed to the corresponding doctor's diagnostic terminal in advance to provide reference for diagnosis. The cloud platform regularly reviews and analyzes user health data, dynamically adjusting health recommendations and intervention plans if abnormal changes in indicators are detected. User data security is ensured throughout the process, with no disclosure to third parties, and users are allowed to independently set data authorization scope through the app.

[0097] In actual deployment, first ensure normal network communication between the control host, cloud platform, and hospital HIS system, and that users' mobile phones can receive push notifications normally. On the software side, Python is used as the development language. The report generation and push module on the control host is developed using Flask or Django frameworks, and the requests library is used to implement interface calls with the cloud platform and HIS system. Health report templates are designed using HTML / CSS, and the Jinja2 template engine is used for dynamic data population. Matplotlib or Plotly libraries are used to generate pulse waveforms and body constitution radar charts, and the PyPDF2 library is used for PDF conversion and downloading of reports. On the cloud platform side, Java or Python is used. Spark or Flink frameworks are used to process user health data and build user profiles. Collaborative filtering algorithms are implemented using the Surprise library, and the rule engine is developed using Drools or Python's rule-engine library. Redis is used to cache frequently accessed user data, and MySQL or MongoDB is used to store user health records and historical data. The hospital's HIS system interface is adapted using the official HL7 FHIR SDK (such as HAPI FHIR) to ensure data format conformity. Two-way authentication utilizes the OpenSSL library for certificate generation and verification, while data encryption uses the cryptography library. Scheduled reminders are implemented using the APScheduler or Quartz framework, and the user app is developed using React Native or Flutter, supporting data authorization management and report viewing. During deployment, the corresponding software environment is first deployed on the control host and cloud platform server, requiring the installation of necessary dependencies and configuration of the database and caching services. The code for each functional module is then written and debugged, completing interface integration between the control host, cloud platform, and HIS system, and testing the entire process of report generation, push notifications, and data storage. After successful testing, the system is officially launched. During operation, the system's running status and data transmission security are regularly monitored to ensure stability and reliability at every stage.

[0098] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A smart health checkup device based on facial recognition, characterized in that: It includes a control host (6), which is connected to the main display screen (5) and the secondary display screen (1). The secondary display screen (1) is equipped with a first camera (2) on one side, which is used for face recognition. A pulse diagnosis device (4) is provided on one side of the control host (6), and a second camera (3) is provided on the pulse diagnosis device (4). The second camera (3) is used to collect images of the patient's face and tongue. The pulse diagnosis device (4) has blood pressure detection wristbands (7) on both sides; The control host (6) is electrically connected to the pulse diagnosis device (4) and the second camera (3).

2. The intelligent health checkup device based on facial recognition according to claim 1, characterized in that: The secondary display screen (1) is mounted on the support table via a rotatable bracket, and the size of the secondary display screen (1) is larger than that of the main display screen (5).

3. The intelligent health checkup device based on facial recognition according to claim 2, characterized in that: The main display screen (5) is a touch screen and is connected to the control host (6) via a wireless transmission device.

4. The intelligent health checkup device based on facial recognition according to claim 1, characterized in that: The second camera (3) is connected to the top of the pulse diagnosis device (4) via a multi-section adjustable support arm.

5. The intelligent health checkup device based on facial recognition according to claim 1, characterized in that: It also includes a printer (8), which is electrically connected to the control host (6).

6. A health checkup method using a facial recognition-based intelligent health checkup device according to any one of claims 1-5, characterized in that: The method includes: S1. Start the control host (6) terminal. The control host (6) performs a self-test on the pulse diagnosis device (4), blood pressure detection wristband (7), first camera (2) and second camera (3), and displays normal communication to complete the device initialization. S2. The first camera (2) verifies the user's face identity, and the control host (6) verifies the legality of the user's identity. After the verification is passed, the user's electronic health record is automatically associated. S3, pulse diagnosis device (4), and second camera (3) sequentially complete the collection of user tongue diagnosis, facial diagnosis, pulse diagnosis and consultation data, and generate preliminary physical examination data; S4. The control host (6) binds the collected health data with the user identity information, encrypts the bound data, uploads the encrypted data to the cloud platform, and stores a local backup. S5. The control host (6) generates a health report based on the processed data fed back by the cloud platform or the data packets stored locally, and pushes the health report to the user's mobile phone and the hospital's HIS system in a synchronized manner. The cloud platform provides personalized health services based on the user's health data.

7. The physical examination method of the intelligent health examination device based on facial recognition according to claim 1, Its characteristic is: In step S2: S21. Capture the user's facial image through the first camera (2) and perform grayscale standardization on the original image; S22. Adaptive illumination compensation algorithm is used to optimize image quality; S23. An improved multi-task cascaded convolutional neural network is used to realize face region detection, including three stages: P-Net, R-Net, and O-Net. An IoU weighted loss function is introduced, and the formula is: ； In the formula, Let α, β, and γ be the total loss function, and let α, β, and γ be the weighting coefficients of each loss term. For classifying losses, For bounding box regression loss, For keypoint regression loss, The area of ​​the candidate box. Let N be the average area of ​​all candidate boxes, and N be the total number of candidate boxes. P-Net constructs a multi-scale image pyramid on the preprocessed image to generate initial face candidate boxes. R-Net filters candidate bounding boxes by a face probability threshold of 0.6 and resizes them to 24×24×3, outputting the corrected offset. And correct the candidate box Then, overlapping candidate boxes are removed using non-maximum suppression with an IoU threshold of 0.5; O-Net resizes the candidate boxes to 48×48×3 and outputs precise offsets. The coordinates of 5 facial landmarks and the final face probability are used to correct and obtain the final face detection box. And processed by NMS with an IoU threshold of 0.3; S24. Pre-build a 3D face averaging model. ,in Let M be the set of coordinates of 3D vertices, and F be the set of triangle patch indices; 2D key points output by O-Net Corresponding key points in the 3D average model Matching, establishing correspondence through perspective projection transformation. ,in This is the camera intrinsic parameter matrix. Focal length (where R is the coordinate of the image center point) R is a 3×3 rotation matrix, and t is a 3×1 translation vector; The optimal rotation matrix is ​​solved by minimizing the reprojection error using the Levenberg-Marquardt algorithm. With translation vector The objective function is optimized as follows: ； In the formula, 2D key points pixel coordinates, Let the element in the m-th row and n-th column of the rotation matrix R be... Let be the components of the translation vector t. 3D key points The coordinates; according to and Pose correction is performed on the 3D average model to obtain a frontal 3D face model. Then, a standardized frontal face image of 224×224 pixels is obtained through inverse perspective projection. ; S25, will An improved ResNet-50 neural network that integrates channel attention and spatial attention is input, and the feature map is optimized through the channel attention module, as shown in the formula: ,in σ is the Sigmoid activation function, and MLP is the input dimension. The initial feature map output by ResNet-50. This is channel-by-channel multiplication; The feature map is optimized using a spatial attention module, as shown in the formula. ,in , This is a 7×7 convolutional layer for output channel 1, with Concat being the channel concatenation operation; Will The original feature vector is obtained by inputting the GAP layer. The final face feature vector is then obtained through L2 normalization. The formula is: ； In the formula, , For the k-th component of the final eigenvector, This is the k-th component of the original feature vector; S26. Retrieve the pre-stored user facial feature template library from the database. N is the total number of database users. For the pre-stored feature vector of the i-th user, calculate the weight of each component based on the ratio of the inter-class variance to the intra-class variance of the feature components. , Let be the inter-class variance of the k-th feature component. Let be the intraclass variance of the k-th feature component; Calculate the feature vector of the user to be verified With template feature vector The weighted cosine similarity is calculated using the following formula: ； In the formula, For weighted cosine similarity; S27. A dynamic threshold determination mechanism is adopted, and the threshold calculation formula is as follows: ； In the formula, T is the dynamic judgment threshold. Based on the threshold, This is the threshold adjustment coefficient. ( The distance between the user feature to be verified and the most similar feature in the database is denoted as . This is the average of the minimum distances between all features in the database; like Determine the user's identity is legitimate and record the matching user ID; if The system determines that the user's identity is invalid and triggers an alarm. S28. Association of Legitimate User Electronic Health Records: Based on the matching user ID, the control host connects to the database through the database query interface to retrieve the user's pre-stored electronic health records. The retrieved record information is then compared with the verification results for consistency. The formula is as follows: ； In the formula, C is the verification coefficient. For the m-th basic information item associated with the user to be verified, This refers to the m-th basic information item of the i-th user profile in the database; If C=1, the facial verification information will be bound to the electronic health record and stored in the temporary cache area; If C < 1, trigger a second query and re-validate until the association is accurate.

8. The physical examination method of the intelligent health examination device based on facial recognition according to claim 1, characterized in that: In step S3: S31. The user's tongue and face images are captured by the second camera (2), and after noise reduction by the improved bilateral filtering algorithm, they are converted to the LAB color space. The U-Net++ model with an attention fusion mechanism is used to perform semantic segmentation of the tongue coating, tongue body, cheek and forehead regions, and extract feature parameters such as tongue coating color, thickness and face color. These parameters were organized into a data matrix for tongue diagnosis and facial diagnosis; S32, Pulse diagnosis device (4) The built-in piezoresistive pulse sensor collects the user's pulse waveform signal. After adaptive wavelet threshold denoising and Chebyshev type I low-pass filtering preprocessing, the "adaptive threshold + slope change detection" algorithm is used to identify key feature points in the pulse waveform and calculate the pulse position, pulse rate, pulse width and pulse pressure ratio parameters to form a pulse diagnosis data vector. S33. Collect user-filled information by touching the main display screen (5), quantify the score of each physical dimension using dimension weight and reverse question correction algorithm, and generate a radar chart reflecting the user's physical characteristics; output the total score of each dimension and the set of radar chart coordinate points as consultation data. S34. The control host (6) receives and associates the timestamps of the above three types of data, formats them into a JSON structure, encrypts and stores them in the local cache, and marks them as pending upload status to prepare for uploading to the cloud platform.

9. The physical examination method of the intelligent health examination device based on facial recognition according to claim 1, characterized in that: In step S41: S41. Binding health data with user identity information: Extract core user identity information from the face recognition verification results, including user ID, ID card number, and verification timestamp, to form an identity information tuple; standardize the format of the identity information, convert the user ID to 32-bit UUID format, process the ID card number into a mask format that retains the first 6 and last 4 digits and replaces the middle digits with *, and convert the timestamp to ISO 8601 standard format; A dual binding mechanism of timestamp association and hash verification is adopted. The time difference between the health data collection timestamp and the identity verification timestamp is calculated. If the time difference is ≤300 seconds, it is determined that it is the same user and the same operation and the binding is allowed. Otherwise, the binding is rejected and manual review is triggered. At the same time, the health data and identity information are jointly hashed to generate a unique binding identifier. The calculation method is to use the SHA-256 algorithm to perform a hash operation on the result of string concatenation of user ID, health data collection timestamp and health data MD5 hash value. The binding identifier, processed identity information, and health data are integrated into a complete data unit. This data unit consists of four parts: Header, Identity, HealthData, and Signature. The Header contains the binding identifier, user ID, health data collection timestamp, and authentication timestamp. The Identity contains the masked ID number. The HealthData is the initially collected health data. The Signature is the result of the control host signing the binding identifier using its own RSA private key. S42. Encryption of bound data: A 256-bit AES symmetric encryption key is generated using the NIST-recommended CTR-DRBG method. During the generation process, the 128-bit initial seed provided by the hardware random number generator, the device's unique identifier, and the hash value of the current timestamp are used as input parameters. The generated AES symmetric encryption key is encrypted using the RSA public key of the cloud platform to obtain the encrypted symmetric key. The integrated complete data unit is encrypted using the AES-256-GCM mode. First, a 96-bit initialization vector and 32-bit additional authentication data are generated. Then, the complete data unit is encrypted using the AES-256-GCM algorithm in combination with the AES symmetric encryption key, the initialization vector, and the additional authentication data to obtain the encrypted health data and 128-bit authentication tag. Subsequently, the initialization vector, additional authentication data, encrypted health data, authentication tag, and encrypted symmetric key are concatenated into a byte stream to form an encrypted data packet. The checksum of the encrypted data packet is calculated by XORing the result obtained by the Cyclic Redundancy Check algorithm with the result obtained by the HMAC algorithm combined with an HMAC key independent of the AES symmetric encryption key. The checksum is then appended to the end of the encrypted data packet to form the final transmitted data. S43. Uploading encrypted data to the cloud platform: The control host establishes a TLS 1.3 encrypted connection with the cloud platform. The connection process uses the ECDHE key exchange algorithm, AES-128-GCM encryption suite, and SHA-256 hash algorithm. The control host sends the device certificate to the cloud platform. After the cloud platform verifies the validity of the certificate, it generates a random challenge value. The control host uses the device private key to sign the random challenge value with ECDSA and sends it to the cloud platform. After the cloud platform verifies the signature, it completes the device identity authentication. The final transmitted data is divided into 1MB segments to obtain a segment sequence. The hash value of each segment is calculated using the SHA-256 algorithm to generate a segment list. The segment list contains the total number of segments, the hash value of each segment, and the SHA-256 hash value of the final transmitted data. Upload the shard list and each shard to the cloud platform in sequence. The cloud platform receives each shard and verifies its hash value. If the transmission is interrupted, the control host requests the set of fragment indexes that have been successfully uploaded from the cloud platform and only re-uploads the fragments that were not successfully uploaded. After receiving all fragments, the cloud platform concatenates the fragments into complete final transmission data, verifies the checksum and hash value of the final transmission data, decrypts the integrated complete data unit, and verifies the RSA signature of the binding identifier by the control host. After all verifications are successful, a successful upload receipt is returned to the control host. The control host stores this receipt as an upload credential. The receipt includes the upload completion timestamp, the SHA-256 hash value of the final transmission data, and the content of the cloud platform signing the result of concatenating the upload completion timestamp and the hash value of the final transmission data using its own RSA public key. S44. Local Backup Storage: The final data to be transmitted is stored in the device's built-in eMMC flash memory, using the EXT4 file system, with the file naming format "Backup_User ID_Health Data Collection Timestamp.enc"; Every day at 3:00 AM, the backup data stored in the eMMC flash memory in the past 24 hours is synchronized to the external SD card. During the synchronization process, the SHA-256 algorithm is used to calculate the hash value of the backup file in the eMMC flash memory and the synchronized file in the SD card. If the two hash values ​​are the same, the synchronization is considered successful; otherwise, the file is resynchronized and the synchronization log is recorded. An automatic deletion of expired backup data is achieved using a dynamic cleanup algorithm based on user activity. The data retention period is calculated as follows: a base retention period of 30 days is added, along with the product of 180 and the natural exponential function value, where the exponent of the natural exponential function is 0.5, and the user's activity level over the past 30 days is calculated as the sum of the number of logins and health checkups performed by the user in the past 30 days, divided by 30 and multiplied by 2. The activity level ranges from 0 to 1. When the number of days a backup file is stored exceeds the calculated data retention period, the backup file is automatically deleted. Every Sunday at 2:00 AM, an integrity check is performed on all local backup data. The check is performed by comparing the previously calculated checksum with the result obtained by recalculating the encrypted data packet using the Cyclic Redundancy Check algorithm and the result obtained by XORing the encrypted data packet using the HMAC-SHA256 algorithm combined with the corresponding HMAC key. If they match, the backup data is considered valid; otherwise, it is considered invalid. For backup files that fail verification, attempt to download and replace them again from the cloud platform. If the download fails, mark the file as corrupt and notify the administrator. S45. Full-process verification: Set up verification mechanisms at each key node of health data processing, including verifying whether the bound identifier is consistent with the result of hashing the result obtained by concatenating the user ID, health data collection timestamp, and MD5 hash value of health data using the SHA-256 algorithm; verifying the validity of the authentication tag generated in AES-256-GCM mode after data decryption; verifying the checksum of the encrypted data packet and the hash value of the final transmitted data on the cloud platform; and periodically verifying the validity of backup data locally. All verification results are recorded in an immutable audit log. The log is stored in a chained manner. Each log entry contains the current operation timestamp, operation type, operation result, and the result of hashing the previous log entry using the SHA-256 algorithm.

10. The physical examination method of the intelligent health examination device based on facial recognition according to claim 1, Its characteristic is: In step S5: S51. The control host (6) prioritizes calling the processed data fed back by the cloud platform after multi-dimensional data comparison and verification. If it does not receive the data, it retrieves the backup data packet stored locally in encryption. S52. Load the preset template containing the user basic information area, physical examination indicator display area, health status assessment area and suggestion and guidance area, fill in the masked user identity information, and present the tongue diagnosis, face diagnosis, pulse diagnosis and consultation data in the form of structured tables, waveform graphs and parameter combination, and constitution radar chart respectively. S53. Using a decision tree model constructed from clinical data, comprehensively analyze multi-dimensional data to generate a three-level health assessment conclusion of "no obvious abnormalities", "mild abnormalities" and "requires medical intervention", and complete the generation of a health report; S54. Push reports containing 24-hour valid encrypted links and verification codes to users' mobile phones via encrypted push interfaces, support PDF download, and push concise letters containing core conclusions when the corresponding APP is not installed; push standardized report data to the hospital HIS system via medical data interfaces that conform to the HL7 FHIR standard, add data traceability identifiers, establish a secure connection through two-way authentication, and receive and store confirmation receipts. S55, the cloud platform uses user ID as an identifier to integrate current and historical physical examination data to form a user profile containing long-term health trends, and associates user age and gender to focus on corresponding indicators; An algorithm combining collaborative filtering and a rule engine is used to generate categorized health recommendations, including specific and actionable measures. S56. Generate a 1-3 month phased intervention plan for users with "mild abnormalities" and set regular reminders; match cooperative hospital departments for users who "require medical intervention" and generate an appointment recommendation list; and push desensitized health data to the corresponding doctor's terminal in advance. Regularly review and analyze health data, dynamically adjust service content, ensure data security, and allow users to independently manage the scope of data authorization.

Images