Examination management system for face recognition and automatic examination position distribution

Abstract

The invention relates to the technical field of intelligent examination management, and discloses an examination management system for face recognition and automatic examination position distribution. The system comprises an identity certificate reading module, a biological characteristic acquisition module, an identity verification module, an examination position management module and an output interaction module which are arranged in sequence. The identity certificate reading module obtains examinee identity certificate information; the biological characteristic acquisition module performs living body detection through an infrared camera and generates a dynamic face biological characteristic sequence; the identity verification module performs matching verification on the voucher information and the biological characteristics; after the verification is passed, the examination position management module automatically retrieves idle examination positions and dynamically allocates the idle examination positions; and a final result is informed to examinees through the display screen and the loudspeaker. According to the system, identity pretending is effectively prevented through a living body detection technology, through real-time linkage of verification and distribution, whole-course automation of entrance of examinees is realized, and the safety and efficiency of examination room management are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent examination management technology, specifically to an examination management system that uses facial recognition and automatic allocation of examination seats. Background Technology

[0002] Existing examination room management procedures generally rely on manual operation and static information comparison. When candidates enter the examination room, they need to present a paper admission ticket or ID card, which invigilators will manually verify to ensure that the photo on the ID card matches the candidate's appearance. After verification, candidates must find their seat on a paper seating chart at the entrance of the examination room or be informed of their seat by staff. This model constitutes the first step in examination room management, and its core is visual comparison and manual query.

[0003] This traditional method has its flaws. Manual identity verification is inefficient and can easily cause congestion at the entrance during peak periods. The accuracy of human judgment of photos is easily affected by subjective factors, changes in the candidate's makeup, and ambient lighting, and it cannot effectively identify identity impersonation using high-resolution photos or screen captures, thus creating security vulnerabilities. The transmission of test slot information relies on static paper lists or manual memorization, and information updates are not timely. If there are temporary changes to test slots, it can easily cause confusion among candidates and disrupt the examination room order.

[0004] Existing technologies also include solutions that attempt to use electronic means for identity verification. While these methods improve verification efficiency to some extent, they still fall short in terms of anti-counterfeiting capabilities. Simple static facial recognition cannot withstand more complex deception methods such as digital screen capture and 3D masks. These electronic verification systems often operate independently and do not form a data loop with the examination seat allocation system. After successful verification, candidates still need to obtain examination seat information through other channels, and the entire entry process is fragmented into multiple independent steps, failing to achieve a fully automated process from identity confirmation to location guidance.

[0005] Current technology cannot meet the urgent needs of large-scale, high-security examinations for efficient entry, accurate identity verification, and seamless process integration. The initial stage of examination room management requires an integrated solution capable of automatically, accurately, and fraudulently verifying identity and instantly triggering subsequent seat allocation and guidance. Summary of the Invention

[0006] The purpose of this invention is to provide an examination management system with facial recognition and automatic exam seat allocation to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an examination management system for facial recognition and automatic allocation of examination seats, the system comprising: The identity credential reading module uses radio frequency identification technology or optical character recognition technology to obtain the encoded data in the candidate's identity credential and parses it to obtain the identity identification information. The biometric acquisition module captures facial images of examinees through an infrared camera and runs a liveness detection program to distinguish between real faces and fake images, generating a facial biometric sequence. An identity verification module receives the identity information and the facial biometric sequence, performs feature matching using a pattern recognition algorithm, and outputs the identity verification status. The exam slot management module accesses the exam slot status database based on the authentication status, retrieves currently available exam slot information, and executes a dynamic allocation algorithm to generate exam slot allocation instructions. The output interaction module receives the examination seat allocation instruction, synthesizes visual interface content and voice broadcast content, and presents them to the examinees through an LCD screen and a speaker.

[0008] Preferably, the identity credential reading module uses radio frequency identification (RFID) or optical character recognition (OCR) technology to obtain the encoded data in the candidate's identity credential and parses it to obtain identity identification information, including: The candidate registration data stored in the chip embedded in the ID card is read using radio frequency signals; Alternatively, you can scan the barcode on the paper voucher with a camera to decode the digital code; The original data is verified and parsed to extract the candidate's unique identification code and name information; The extracted information is packaged into a standard format of identity information.

[0009] Preferably, the biometric acquisition module captures the examinee's facial image through an infrared camera and runs a liveness detection program to distinguish between real faces and fake images, generating a facial biometric sequence, including: Control the infrared camera to capture a sequence of multiple facial images; Each frame of the image is subjected to illumination normalization and noise filtering. The convolutional neural network model is used to analyze texture features and blood flow signals in images to determine the probability of a living organism. When the probability of a live subject exceeds a preset threshold, facial key point feature vectors are extracted to form a facial biometric sequence.

[0010] Preferably, the identity verification module receives the identity information and the facial biometric sequence, performs feature matching using a pattern recognition algorithm, and outputs the identity verification status, including: Retrieve pre-stored standard facial feature templates from the identity information; Calculate the Euclidean distance between the facial biometric sequence and the standard facial feature template; The Euclidean distance is compared with a similarity threshold. If the distance is less than the threshold, the label verification passes; otherwise, the label verification fails. The identity verification status, along with the confidence score, is sent to the test slot management module.

[0011] Preferably, the authentication module triggers a retry mechanism when it outputs an authentication failure status, including: Control the biometric acquisition module to re-acquire the facial image; Repeat the feature extraction and matching process, accumulating the number of retries; If the cumulative number of retries reaches the maximum and the verification still fails, an alarm signal will be activated and the test slot allocation process will be suspended.

[0012] Preferably, the examination slot management module executes a dynamic allocation algorithm to generate examination slot allocation instructions, including: Retrieve the real-time exam slot occupancy status table from the exam slot status database; filter the available exam slots based on the matching rules of candidate category and exam slot type; Apply a load balancing strategy to select the optimal test slot from the available test slot set; generate a test slot allocation instruction and update the occupancy status in the test slot status database.

[0013] Preferably, the examination slot management module applies a load balancing strategy to select the optimal examination slot from the available examination slot set, including: Calculate the historical usage frequency and current distribution density of each available exam slot; select the exam slot with the lowest usage frequency and uniform distribution based on a greedy algorithm; record the selection results in the allocation log and mark the exam slot as occupied.

[0014] Preferably, the output interaction module receives the examination seat allocation instruction, synthesizes visual interface content and voice broadcast content, and presents them to the examinees through an LCD screen and a speaker, including: Map the test slot number and location information in the test slot allocation instruction to the graphical interface template; generate dynamic display content including the test slot route map; The system converts text prompts into speech signals, adjusts the speech rate and volume parameters, renders the content on an LCD screen, and simultaneously plays the speech signal through a speaker.

[0015] Preferably, when the authentication module performs feature matching using a pattern recognition algorithm, it performs the following steps: Retrieve pre-stored multi-factor authentication data from the identity information, including password hash values ​​and backup contact information; The facial biometric sequence is matched with a pre-stored standard facial feature template, and the matching degree is calculated. Simultaneously verify whether the password hash value matches the input password; If both face matching and password verification are successful, the authentication status will be output as successful. If the face matching is successful but the password verification fails, or if the face matching fails but the password verification is successful, a backup authentication process is triggered, which involves secondary verification via SMS verification code or fingerprint recognition.

[0016] Preferably, when the examination slot management module accesses the examination slot status database based on the authentication status, it also integrates environmental monitoring data and performs the following steps: Real-time data on exam room temperature, light intensity, and noise levels are obtained from environmental sensors. Analyze the temperature, light intensity, and noise level data of the examination room to calculate the comfort index of each examination seat; The priority of test seat allocation is adjusted according to the comfort index, with priority given to test seats with higher comfort indices; Based on the matching rules of candidate category and test slot type, generate optimized test slot allocation instructions.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This technological approach, which uses infrared cameras to capture images and runs a liveness detection program to generate facial biometric sequences, fundamentally changes the benchmark for identity verification. Infrared imaging can perceive the blood flow thermogram and three-dimensional structural features under the facial skin—biosignatures unique to living organisms and difficult to replicate in two-dimensional or static ways. The liveness detection program analyzes facial micro-movements, blinking, and other physiological activities to ensure the dynamic authenticity of the collected features. Encoding this series of dynamic, live biosignatures into feature sequences for matching allows the identity verification model to move beyond simple static appearance comparison and construct a dynamic authentication model incorporating liveness information. This enhances the system's resistance to various deception methods, such as photos, videos, and highly realistic masks, elevating the security level of identity verification from image recognition to liveness biometric identification.

[0018] Using identity verification status as the sole trigger signal, the examination slot management module automatically retrieves data from the database and executes a dynamic allocation algorithm, achieving seamless integration of business logic. This design eliminates the delays caused by manual intervention and information transmission in traditional processes. Successful verification directly activates the examination slot allocation process, with the system querying the latest examination slot status database in real time to ensure the accuracy and timeliness of allocation information. The dynamic allocation algorithm can instantly make decisions and generate allocation instructions based on preset rules. This automated chain allows candidates to receive clear examination slot guidance the moment they complete identity verification, compressing the previously separate steps of verification, query, allocation, and guidance into a single, instantaneous operation. This simplifies the candidate entry route, reduces crowding and waiting at the entrance, and improves the smoothness and orderliness of the entire examination room management. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the working principle of the examination management system for face recognition and automatic exam seat allocation as described in this invention. Figure 2 A flowchart illustrating the operation of the identity credential reading module; Figure 3 A flowchart illustrating the operation of the biometric acquisition module; Figure 4 A graph showing the relationship between pass rate and confidence level under the retry mechanism of the identity verification module; Figure 5 To output a diagram showing the response time distribution of each functional component of the interactive module. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 This invention provides an examination management system for facial recognition and automatic exam seat allocation. The system includes: an identity credential reading module that first acquires the coded data from the candidate's identity credential and parses out the identity identification information; a biometric feature acquisition module that simultaneously captures the candidate's facial image, performs liveness detection, and generates a facial biometric feature sequence; an identity verification module that performs feature matching between the identity identification information and the facial biometric feature sequence and outputs the identity verification status; an exam seat management module that accesses the exam seat status database based on the verification status, dynamically allocates available exam seats, and generates allocation instructions; and an output interaction module that finally presents the exam seat information to the candidate through a visual interface and voice broadcast.

[0022] Example 1: See Figure 2 In practical implementation, the identity credential reading module's hardware platform integrates a high-frequency radio frequency (RF) reader unit and an optical scanning camera unit. The RF reader unit operates at a frequency of 13.56MHz, conforming to the ISO / IEC 14443-A standard. The optical scanning camera unit is equipped with autofocus and supplementary LED lighting, with a resolution of at least 5 megapixels. When using RFID technology, the examinee places their second-generation ID card within the reader's effective sensing area. The reader then activates the chip embedded in the ID card by transmitting an RF signal via its antenna. After power-on, the chip sends the stored examinee registration data back to the reader via load modulation. This registration data includes structured data fields such as name, ID number, photo information, and exam subject codes. The data stream received by the reader is decoded by the baseband processing chip. The decoding process includes Manchester encoding parsing and frame structure reassembly. Subsequently, a cyclic redundancy check (CRC) is performed on the reassembled data packet to verify the integrity of the data transmission. If the CRC code match fails, the identity credential reading module controls the reader to retransmit the RF signal for a second reading until successful or the maximum number of retries is reached.

[0023] In practical implementation, the identity credential reading module parses the raw data read by RFID technology. The parsing process includes field separator recognition, encoding format conversion, and information extraction. Candidate registration data is typically stored in TLV format, where the tag byte identifies the data type, the length byte indicates the size of the numerical field, and the numerical field contains specific candidate information. The parsing algorithm locates the data segment containing the candidate's unique identifier and name based on a predefined tag mapping table. The candidate's unique identifier may be a citizen identification number or a globally unique identifier generated internally by the system. The name information may be encoded in GB2312 or UTF-8, requiring correct character set conversion based on the identifier. The extracted raw string undergoes regular expression matching verification to ensure it conforms to the format specifications of the ID card number or name, such as the correctness of the number of digits and check digits in the ID card number. In practical implementation, the identity credential reading module uses optical character recognition (OCR) technology as an alternative or supplementary solution. OCR technology is primarily used to recognize barcodes or QR codes on paper credentials. After the camera captures the image of the voucher, the image preprocessing stage performs grayscale conversion, binarization, and perspective correction to eliminate the effects of paper wrinkles and uneven lighting. Then, open-source libraries such as ZBar or ZXing are used to decode the barcode, or FinderPattern positioning and version information decoding are performed on the QR code, resulting in a string of numerical codes. This string of numerical codes needs to be queried and matched against the backend database to retrieve complete candidate identity information.

[0024] Understandably, the identity credential reading module encapsulates the extracted information into standard-format identity information, using the JavaScriptObjectNotation lightweight data exchange format. The encapsulation process creates a root object containing multiple key-value pairs. For example, the "student_id" key corresponds to the string value of the candidate's unique identifier, the "name" key corresponds to the string value of the candidate's name, and it may also contain a "photo_url" key pointing to the path of a pre-stored photo or binary data. In practice, the encapsulation operation ensures data type uniformity; all numeric data is converted to string representation, and all timestamps use the ISO8601 standard format. The encapsulated identity information object is serialized into a byte stream and transmitted to the authentication module via an internal message queue or remote procedure call interface. Optionally, the identity credential reading module may attach a digital signature or message authentication code before encapsulation, using a pre-shared key to perform a hash operation on the identity information to generate an authentication code, which is then encapsulated along with the data. This allows the receiver to verify the authenticity and integrity of the data source and prevents data tampering during transmission.

[0025] In some embodiments, the error handling mechanism of the identity credential reading module includes invalid credential alarm and duplicate credential detection. When the chip data read by RFID technology fails to pass parsing verification, or the digital code decoded by optical character recognition technology has no corresponding record in the background database, the module will determine that the credential is invalid, trigger an audible and visual alarm locally, and record the error code and original data snapshot in the operation log. The duplicate credential detection function is implemented by maintaining a cache queue of recently successfully read credentials. When a new credential code matches an entry in the cache queue, the module will send a command to the output interaction module, prompting "This credential has been used, please do not verify again," and reject the read operation. This mechanism effectively prevents the same ID card or paper credential from being used multiple times for exam seat allocation. It is understood that the performance parameters of the identity credential reading module need to meet the high throughput requirements of the actual scenario, and the average single credential reading and parsing time should be controlled within 500 milliseconds to ensure the smooth entry process for candidates.

[0026] Optionally, the identity credential reading module supports mixed reading modes for multiple credential media. The system configuration file can set the priority order; for example, it can prioritize attempting to read ID cards using RFID technology, and automatically switch to optical character recognition (OCR) technology to scan paper admission tickets if the attempt fails within a timeout period. The mode switching logic is controlled by a state machine, which monitors the reader's readiness status and the camera's availability status, making decisions based on preset priorities and timeout thresholds. In practical implementation, the module's hardware interface is designed to be modular. The RFID reader connects to the host computer via a USB interface or serial port, and the optical scanning camera transmits image data via USB video protocol. This design facilitates independent hardware replacement or upgrades without modifying the upper-layer parsing logic software. The software portion of the identity credential reading module runs as an independent process or service, communicating decoupled from other modules in the system through well-defined application programming interfaces.

[0027] Example 2: See Figure 3 In practical implementation, the core hardware of the biometric acquisition module is an active near-infrared camera with a wavelength range of 850 nm to 940 nm. It is equipped with an infrared LED array to provide uniform illumination in low-light environments. The camera resolution is set to 1920x1080 pixels, and the frame rate is fixed at 30 frames per second to capture a multi-frame sequence of facial images. The camera connects to the main processing unit via a USB 3.0 interface. The main processing unit runs an embedded Linux operating system, on which an image acquisition driver is deployed. The driver controls the camera's startup, focusing, and image capture process. When the examinee stands in the designated area, the camera automatically triggers continuous shooting, capturing a sequence of 5 to 10 images. Each image is temporarily stored in a memory buffer in RAW format. In practice, the image preprocessing stage performs illumination normalization on each frame of RAW image. Illumination normalization adopts an algorithm based on Retinex theory. The algorithm decomposes the image into illumination components and reflection components. The illumination component is estimated and corrected by Gaussian filtering to ensure that facial images under different lighting conditions have a consistent brightness distribution. Noise filtering is then applied using a nonlocal mean denoising algorithm. The algorithm calculates the similarity weight between each pixel in the image and its neighboring pixels, and outputs a denoised image after weighted averaging, effectively suppressing Gaussian noise and salt-and-pepper noise.

[0028] In its implementation, the liveness detection program is based on a convolutional neural network (CNN) model. The CNN uses the ResNet-18 architecture as its backbone, incorporating residual blocks to mitigate the vanishing gradient problem. During training, a large-scale liveness detection dataset is used, containing real face images and various forged images such as printed photos, electronic screen images, and 3D masks. The training process employs a cross-entropy loss function and a stochastic gradient descent optimizer. In the forward inference phase, the preprocessed image sequence is input frame-by-frame into the CNN model. The model first extracts low-level features such as edges and textures through convolutional layers, then downsamples through pooling layers. Deep convolutional layers capture high-level features, including micro-texture patterns and subtle brightness changes caused by subcutaneous blood flow signals. The model ultimately outputs a two-dimensional vector. The first element represents the probability of a forged image, the second element represents the probability of a real face, and the liveness probability is the softmax value of the second element. The liveness detection program sets a preset threshold of 0.95; when the liveness probability exceeds this threshold, the image is considered a real face. The calculation of the liveness probability involves feature transformation, mathematically expressed as follows:

[0029] in: This represents the probability of a living organism. It is the fake class score output by the fully connected layer of the convolutional neural network. It is the true class score output by the fully connected layer. It is an exponential function. This formula transforms the raw score into a probability distribution, ensuring that the probability values ​​are between 0 and 1.

[0030] In some embodiments, the biometric acquisition module supports multimodal liveness detection. In addition to infrared images, it can optionally integrate near-infrared video stream analysis. Facial micro-motion signals are detected by calculating changes in optical flow between consecutive frames. These micro-motion signals are correlated with heart rhythm, thereby enhancing the reliability of liveness detection. Data augmentation strategies, including random rotation, scaling, and color jitter, are employed during convolutional neural network model training to improve model generalization ability. In specific implementations, when the liveness probability does not reach a preset threshold, the system does not immediately reject the image but initiates a multi-frame fusion decision. This involves averaging the liveness probability across all frames in the image sequence. If the average probability is still below the threshold, the image is determined to be a forged image, triggering an alarm and logging suspicious events. Optionally, the biometric acquisition module simultaneously acquires depth information while capturing the image sequence. It obtains facial 3D point cloud data through an integrated time-of-flight sensor or structured light projector. The 3D data and 2D images are input into the convolutional neural network model for multimodal feature fusion. The fused features are integrated through early or late fusion strategies to further improve liveness detection accuracy.

[0031] In practice, the facial biometric sequence generation process begins after liveness detection is successful. Feature extraction employs a deep learning-based keypoint detection algorithm using an Hourglass network structure. The network outputs coordinate mappings for 68 facial keypoints. For each frame, the network locates keypoints such as the corners of the eyes, the tip of the nose, and the corners of the mouth, then calculates the Euclidean distance and angular relationships between these keypoints to form a geometric feature vector. Simultaneously, local binary pattern features are extracted around the keypoints. These local binary pattern features compare the grayscale values ​​of each pixel with its neighboring pixels to generate a binary pattern histogram. The geometric feature vector and the local binary pattern feature vector are concatenated into a 128-dimensional composite feature vector, representing the facial biometric features of a single frame. For the entire image sequence, the composite feature vectors of all frames are arranged chronologically to form the facial biometric sequence. The sequence data is stored as a floating-point array with timestamps and frame index metadata. It is understood that the feature extraction process must ensure rotation invariance and scale invariance. The Hourglass network uses an hourglass encoder-decoder structure to achieve multi-scale feature learning, thereby addressing changes in facial pose. In some embodiments, the biometric acquisition module further includes a quality assessment submodule. This submodule calculates a sharpness score and an occlusion score for each image frame. If the score is below a threshold, the frame is automatically discarded and re-acquisition is triggered, ensuring reliable image quality for input feature extraction. Optionally, after the facial biometric sequence is generated, it undergoes Undergo normalization. This normalization uses the Z-score standardization method to scale the feature values ​​to zero mean and unit variance, eliminating the influence of dimensions and facilitating matching by subsequent modules. The software implementation of the biometric acquisition module is based on the OpenCV and TensorFlow frameworks. The module runs as an independent service, communicating with the authentication module via a gRPC interface to transmit facial biometric sequences and liveness detection results.

[0032] Example 3: In specific implementation, the identity verification module runs as an independent software service on the central processing unit. The service communicates with the identity credential reading module and the biometric feature acquisition module via a remote procedure call interface. It receives a standard-format identity information packet from the identity credential reading module and a facial biometric feature sequence from the biometric feature acquisition module. The identity information packet contains the candidate's unique identifier and name field. The facial biometric feature sequence is a multi-dimensional floating-point array representing the time series of facial key point feature vectors. The identity verification module first parses the identity information packet, extracts the candidate's unique identifier as the query key, and accesses the encrypted candidate feature database. The database uses an SQLite relational database management system and stores a pre-stored standard facial feature template for each candidate. The pre-stored standard facial feature template is a feature vector collected by the biometric feature acquisition module during the candidate registration stage and normalized. The query operation uses parameterized SQL statements to prevent injection attacks. It retrieves the corresponding pre-stored standard facial feature template based on the candidate's unique identifier. If the query result is empty, the identity verification module returns an error status and logs it. In practical implementation, the feature matching process employs the distance metric method from pattern recognition algorithms. It calculates the Euclidean distance between the facial biometric sequence and a pre-stored standard facial feature template. The Euclidean distance calculation is based on a vector space model, treating feature vectors as points in a multi-dimensional space; a smaller distance indicates higher similarity. The calculation process iterates through each frame of the facial biometric sequence, performing element-wise difference, summation of squares, and finally taking the square root to obtain the distance value. The formula is as follows:

[0033] in: Represents Euclidean distance. It is the dimension of the feature vector. It is the k-th element of the feature vector of the current frame in the facial biometric sequence. This is the k-th element of the pre-stored standard facial feature template. The identity verification module presets a similarity threshold of 0.6. It compares the calculated Euclidean distance with the threshold; if the Euclidean distance is less than the threshold, the verification is marked as successful; otherwise, the verification is marked as failed. The verification status, along with the confidence score, is encapsulated into a message structure. The confidence score is converted from the Euclidean distance to a value between 0 and 1 through a linear mapping; a higher score indicates higher confidence. The message is sent to the examination slot management module via a message queue.

[0034] In some embodiments, the authentication module triggers a retry mechanism when the authentication status is "authentication failed." This retry mechanism is controlled by a state machine. The initial state of the state machine is the first authentication attempt. When the authentication failure signal arrives, the state machine transitions to the retry state and sends a re-acquisition command to the biometric acquisition module. Upon receiving the command, the biometric acquisition module initiates a new image capture process, controlling the infrared camera to acquire multiple frames of facial image sequences again, repeatedly performing liveness detection and feature extraction to generate a new facial biometric sequence. The authentication module accumulates the number of retries. The retry counter initially has a value of 0 and increments by 1 with each retry. If the accumulated number of retries reaches the maximum attempt value of 3 and authentication still fails, the state machine transitions to the alarm state, activating the audible and visual alarm connected to the GPIO interface, pausing the examination allocation process, and sending an alarm message to the system monitoring center. It can be understood that the retry mechanism is designed with user experience and system security in mind, reducing the false rejection rate by providing additional opportunities in case of failure, while preventing unlimited attempts through a limit on the number of attempts. Optionally, the authentication module may adjust the similarity threshold during the retry process. For example, the threshold may be temporarily relaxed to 0.65 during the second retry to increase the probability of success, but the original threshold may be restored during the third retry to ensure security.

[0035] In its implementation, the identity verification module supports multi-factor authentication to enhance security. The multi-factor authentication process integrates password verification and backup contact information in addition to basic face matching. When the identity verification module retrieves pre-stored multi-factor authentication data from the identity information, this data includes a password hash value and a backup mobile phone number. The password hash value is a fixed-length string stored after hashing the candidate's password using the SHA-256 algorithm. The backup mobile phone number is used for SMS verification. During the feature matching phase, the identity verification module executes two verification threads in parallel: the first thread calculates the Euclidean distance between the facial biometric sequence and the pre-stored standard facial feature template to obtain the face matching degree; the second thread receives the password entered by the candidate via keyboard, performs the same SHA-256 hash operation on the password, and compares the result with the pre-stored password hash value. If the face matching degree is successful (Euclidean distance less than the threshold) and the password verification is successful, the identity verification status is output as successful; if the face matching degree is successful but the password verification fails, or vice versa, the backup authentication process is triggered. The backup authentication process first checks the reason for failure. If password verification fails, the identity verification module sends a one-time verification code to a backup mobile phone number via an integrated SMS gateway. The verification code is a 6-digit number, valid for 5 minutes. After the candidate enters the verification code, the system compares its correctness. If face matching fails, the identity verification module optionally integrates a fingerprint recognition module, guiding the candidate to press a fingerprint sensor to collect fingerprint features and match them with a pre-stored template. It can be understood that multi-factor authentication reduces the risk of impersonation by superimposing independent factors, while the backup authentication process acts as a fault-tolerance mechanism to improve system robustness. In some embodiments, the software implementation of the identity verification module adopts a microservice architecture, with each verification factor acting as an independent sub-service, coordinating its work through event-driven communication. For example, the face verification service publishes a verification result event, and the password verification service listens for this event and decides whether to trigger the backup process. Optionally, the identity verification module appends a timestamp and session identifier when outputting the verification status for audit trail purposes, and all operation logs are recorded to a security log server for post-event analysis. Performance optimization of the identity verification module includes feature vector indexing technology, constructing a KD-tree index on pre-stored standard face feature templates to accelerate nearest neighbor search in Euclidean distance calculation, thereby reducing matching latency and meeting real-time requirements.

[0036] See Figure 4The diagram illustrates the retry mechanism of the identity verification module: After receiving the identity information and facial biometric sequence, the module performs feature matching and outputs the verification status. If verification fails, the biometric acquisition module is triggered to re-acquire the image, repeating the feature extraction and matching process, with a maximum of 3 retries. The number of retries in the diagram covers the range of 0 to 3, with the pass rate increasing with the number of retries: approximately 87% at 0 retries and exceeding 95% at 3 retries; the confidence level also increases, approximately 0.65 at 0 retries and approaching 0.9 at 3 retries. This correlation reflects the core logic of the retry mechanism: by increasing the number of facial image acquisitions, the effectiveness of the facial biometric sequence is optimized, thereby improving the accuracy of feature matching; the simultaneous increase in confidence level reflects the increased reliability of the matching results with the number of retries. This mechanism reduces the false rejection rate through multiple rounds of acquisition, ensuring a better verification experience for test takers, while also preventing malicious unlimited attempts by setting a 3-retry limit. It achieves a balance between experience and security during the identity verification process, directly reflecting the fault tolerance and security control strategy of the identity verification module, and aligning with the module's operational logic outside of multi-factor authentication.

[0037] Example 4: In specific implementation, the exam slot management module runs as a background service on the application server. This service establishes a persistent connection with the exam slot status database through a database connection pool. The exam slot status database uses a MySQL database management system to store exam room layout, exam slot attributes, and real-time status information. When the authentication status sent by the authentication module is "verification passed," the exam slot management module initiates the allocation process. First, it sends an SQL query statement to the exam slot status database, with the query condition being that the occupancy status field equals "available." The query result returns a real-time exam slot occupancy status table, which includes fields such as exam slot number, physical location coordinates, exam slot type, and candidate category compatibility flag. The dynamic allocation algorithm then filters the available exam slots based on the candidate category and exam slot type matching rules. The candidate category field is obtained from the identity information, such as "regular candidate," "disabled candidate," or "special arrangement candidate." Exam slot types include "standard exam slot," "accessible exam slot," and "reserve exam slot." The matching rule is defined as: disabled candidates can only be allocated accessible exam slots, while regular candidates are given priority for standard exam slots. In practical implementation, a load balancing strategy is used to select the optimal exam slot from the available slots. The strategy calculates the historical usage frequency and current distribution density of each available slot. The historical usage frequency is calculated from the allocation log table, noting the number of times each slot was allocated in the past hour. The current distribution density is obtained by averaging the Euclidean distances between each slot and its three most recent allocated slots; a larger distance indicates a sparser distribution. A greedy algorithm iterates through the available slots, selecting the slot with the lowest historical usage frequency and the highest current distribution density as the optimal slot. The selection process can be quantified using the following formula to determine the priority score for each slot:

[0038] in: Indicates priority score, This is the current distribution density value. This is the historical usage frequency value. It is a logarithmic function with base 10. The formula is designed so that test slots with low usage frequency and sparse distribution also score high, and the algorithm selects the test slot with the highest score. After generating the test slot allocation instruction, the test slot management module executes an UPDATESQL statement to update the occupancy status field in the test slot status database to "occupied" and inserts a record into the allocation log table. Refer to Table 1, the allocation log table records the test slot number, the candidate's unique identification code, and the allocation timestamp.

[0039] Table 1: Examination Seat Status Database Table

[0040] In some embodiments, the examination seat management module integrates environmental monitoring data to optimize allocation decisions. Environmental monitoring data is acquired in real-time from an environmental sensor network deployed within the examination room. These sensors include digital temperature sensors, illuminance sensors, and noise sensors. The sensors transmit data to a gateway via the ZigBee wireless protocol, and the gateway aggregates the data and provides it to the examination seat management module via an HTTP interface. The examination seat management module analyzes the examination room temperature data, illuminance data, and noise level data to calculate a comfort index for each examination seat. The comfort index is calculated based on a weighted average model, which assigns a weight coefficient to each environmental factor: temperature suitability (0.4), illuminance uniformity (0.3), and noise quietness (0.3). Temperature suitability is obtained by comparing the reciprocal of the absolute difference between the current temperature and the ideal temperature of 22 degrees Celsius. Illuminance uniformity is obtained by normalizing the ratio of the current illuminance to the standard illuminance of 500 lux. Noise quietness is obtained by normalizing the difference between the current noise level and the maximum permissible noise level of 60 decibels. The comfort index is used to adjust the priority of exam seat allocation. Based on the priority score calculated by the load balancing strategy, the comfort index is multiplied to obtain the final score, thus prioritizing the allocation of exam seats with higher comfort indices. This integration of environmental factors ensures that exam seat allocation considers not only resource utilization but also the actual experience of examinees, improving exam fairness. Optionally, the exam seat management module supports dynamic weight adjustments, automatically increasing the weight of temperature suitability during hot summer weather and increasing the weight of noise level in exam rooms located near streets. The weight coefficients are externalized through configuration files for easy operation and maintenance management.

[0041] In practical implementation, the exception handling mechanism of the exam slot management module includes exam slot conflict detection and automatic recovery. When two allocation requests almost simultaneously hit the same available exam slot, the database transaction isolation level is set to repeatable read, and with the optimistic locking mechanism, the version number or timestamp is checked when updating the occupancy status. If the data has been modified by other transactions, an exception is thrown. After the exam slot management module catches the exception, it rolls back the transaction, re-queries the exam slot status, and executes the allocation algorithm. It can be understood that the exam slot status database uses a master-slave replication architecture to achieve high availability. The master database handles write operations, and the slave database handles read operations. When the master database fails, the exam slot management module automatically switches to the slave database to ensure service continuity. In some embodiments, the exam slot management module also maintains an exam slot reservation list for pre-allocation of group registrations or candidates with special needs. Reserved exam slots are marked as "reserved" in the real-time exam slot occupancy status table. The dynamic allocation algorithm skips these exam slots, and the reservation information is pre-configured through the administrator interface. Optionally, after the exam slot allocation instruction is generated, the exam slot management module will publish an allocation success event to the message middleware. The event content includes the candidate's unique identification code and exam slot number. Downstream systems, such as the guidance system or monitoring system, can subscribe to this event to trigger subsequent operations. The performance monitoring of the exam slot management module is achieved by recording the processing time of each allocation request. The processing time includes database query time, algorithm execution time, and database update time. If the average processing time exceeds a preset threshold, such as 1 second, the system will issue an alarm indicating a performance bottleneck.

[0042] Example 5: In specific implementation, the output interaction module runs as a front-end service on an embedded terminal device. The device is equipped with a high-performance graphics processing unit and an audio codec chip. The service receives the examination seat allocation instruction from the examination seat management module through a message queue. The instruction includes the examination seat number, the physical location coordinates of the examination seat, the examination room layout identifier, and text prompts such as "Please go to examination seat A101". The visualization interface content synthesis thread first parses the examination seat allocation instruction, extracts the examination seat number and location information, and then accesses the locally stored graphical interface template library. The template library contains a variety of pre-designed interface styles, such as the default style and the high-contrast style. The templates are defined in HTML5 and CSS3 formats and include placeholder variables. The synthesis engine maps the examination seat number and location information to the corresponding placeholders in the template, performs string replacement operations to generate static text content, and dynamically generates an examination seat route map. The route map generation is based on the examination room layout vector data, and the vector data is stored as an SVG format file, containing elements such as the coordinates, channels, and exits of all examination seats. The route map rendering engine calculates the optimal path based on the physical coordinates of the test center. The path calculation uses the A* algorithm to find the shortest path from the entrance to the target test center, and then overlays this path onto the test center layout map, ultimately generating dynamic display content containing text and graphics. The graphical interface template supports responsive design, adapting to different sized LCD screens and automatically adjusting font size and layout proportions based on screen resolution detected through media queries.

[0043] In practice, the speech delivery is synthesized by a parallel speech synthesis thread. This thread receives text prompts from the test slot allocation instructions, such as "Your test slot is A101, please follow the green route." The speech synthesis engine is based on a deep learning acoustic model using the Tacotron2 architecture. It converts the input text sequence into a Mel spectrogram, which is then converted into waveform data using a WaveNet vocoder. Speech rate and volume parameters are adjusted in real-time during the synthesis process. The speech rate parameter controls the number of synthesized speech frames per unit time, while the volume parameter controls the amplitude gain of the output waveform. Parameter adjustments are based on a preset configuration file and the real-time system status. The speech rate adjustment formula is as follows:

[0044] in: This represents the final speech rate, measured in syllables per second. The default standard speech rate is set to 4 syllables per second. The speech rate adjustment factor is fixed at 0.1. The priority factor, derived from the test slot allocation instruction, ranges from 0 to 1 and is used to fine-tune the speech rate based on the urgency of test slot allocation. Volume adjustment is similar, dynamically compensated based on ambient noise sensor readings, but the core adjustment is still set via configuration parameters. The synthesized speech signal undergoes audio post-processing, including equalizer adjustment and compression limiting, to optimize sound quality and avoid distortion. In some embodiments, the output interaction module supports multilingual speech synthesis. The speech synthesis engine integrates multilingual phoneme sets and acoustic models. When the identity information includes a language preference field, the engine automatically switches the synthesized language, for example, from the default Chinese to English, ensuring that the speech content matches the candidate's native language. Optionally, the speech content supports interruption and replay mechanisms. When a candidate's questioning expression or prolonged pause is detected, the system captures the candidate's behavior through the camera, triggering a replay of the speech content or providing more detailed prompts.

[0045] The rendering of the visual interface content is achieved through a 10.1-inch IPS touchscreen with a resolution of 1920x1200. The rendering engine is based on the OpenGLES 3.0 graphics API, converting dynamically displayed content into frame buffer data. The rendering process includes vector graphics rasterization, texture mapping, and blending operations. Anti-aliasing is used to smooth the edges of paths in the examination location route map, and text rendering uses the FreeType library to generate glyph textures, ensuring clear readability at high resolution. The display driver controller receives frame buffer data and transmits it to the LCD panel via the LVDS interface, continuously updating the displayed content at a 60Hz refresh rate. Simultaneously, the voice broadcast content is output through a 2-watt stereo speaker. The audio signal is amplified by a Class D amplifier before driving the speaker, achieving a sound pressure level of over 70 decibels to ensure clear audibility in noisy environments. The audio-visual output employs a synchronization mechanism, using a hardware timer to ensure that the start time of the voice broadcast is aligned with the time of changes in the interface display, avoiding inconsistencies. In some embodiments, the output interaction module integrates touch interaction functionality, with the LCD screen serving as the touch input device. Candidates can click on elements on the interface to request help or repeat information. Touch events are handled by the Qt framework, triggering corresponding callback functions to update the interface or re-synthesize speech. Optionally, the output interaction module supports an offline working mode. When the network is interrupted, the module continues to run using locally cached routing data and the speech synthesis model, ensuring system reliability. The software architecture of the output interaction module adopts an event-driven model. The main loop listens for examination seat allocation instruction messages and user input events, dynamically creating synthesis and rendering threads. Threads exchange data through shared memory to ensure low-latency output.

[0046] See Figure 5The diagram corresponds to the core functional flow of the output interaction module: after receiving the test seat allocation instruction, the module executes three types of tasks in parallel: interface rendering, speech synthesis, and route generation, and integrates the output through a synchronization mechanism. Interface rendering relies on pre-stored HTML5 / CSS3 graphic templates, so its response time is the shortest; speech synthesis is based on the conversion operation between the Tacotron2 acoustic model and the WaveNet vocoder, which takes slightly longer; route generation needs to call the test room SVG vector data and calculate the optimal path through an algorithm, which further increases the time taken; synchronous output needs to coordinate the timing of interface display and voice broadcast, which adds the overhead of multi-task collaboration, so its response time is the longest. This distribution reflects the module's task scheduling logic: lightweight template rendering is processed first, sufficient resources are allocated to computationally intensive speech and route tasks, and finally, the consistency of information transmission is ensured through a synchronization mechanism. This not only matches the hardware performance of the embedded terminal but also meets the test takers' needs for receiving real-time information, which is a performance characteristic of its dynamic synthesis + synchronous presentation architecture.

[0047] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A test management system for face recognition and automatic allocation of test seat, characterized in that, The system comprises: An identity credential reading module that acquires coded data in an examinee's identity credential using radio frequency identification technology or optical character recognition technology, and analyzes to obtain identity identification information; A biological feature acquisition module that captures an examinee's facial image through an infrared camera, and runs a live body detection program to distinguish between a real face and a fake image, and generates a face biological feature sequence; An identity verification module that receives the identity identification information and the face biological feature sequence, and uses a pattern recognition algorithm to perform feature matching, and outputs an identity verification state; An examination seat management module that accesses an examination seat state database based on the identity verification state, retrieves current idle examination seat information, and executes a dynamic allocation algorithm to generate an examination seat allocation instruction; An output interaction module that receives the examination seat allocation instruction, synthesizes visual interface content and voice broadcast content, and presents to the examinee through a liquid crystal display screen and a loudspeaker.

2. The test management system for face recognition and automatic allocation of seat for examination as claimed in claim 1 wherein, The identity credential reading module acquires coded data in an examinee's identity credential using radio frequency identification technology or optical character recognition technology, and analyzes to obtain identity identification information, including: Reading examinee registration data stored in an embedded chip in an identity card through a radio frequency signal; Or decoding a digital code obtained by scanning a bar code on a paper certificate through a camera; Verifying and analyzing the read raw data to extract an examinee's unique identification code and name information; Packaging the extracted information into identity identification information in a standard format.

3. The test management system for face recognition and automatic allocation of seat for examination as claimed in claim 1 wherein, The biological feature acquisition module captures an examinee's facial image through an infrared camera, and runs a live body detection program to distinguish between a real face and a fake image, and generates a face biological feature sequence, including: Controlling the infrared camera to collect a plurality of facial image sequences; Performing light normalization and noise filtering processing on each image; Applying a convolutional neural network model to analyze texture features and blood flow signals in the image to determine a live body probability; When the live body probability exceeds a preset threshold, extracting a face key point feature vector to form a face biological feature sequence.

4. The test management system for face recognition and automatic allocation of seat for examination as claimed in claim 1 wherein, The identity verification module receives the identity identification information and the face biological feature sequence, uses a pattern recognition algorithm to perform feature matching, and outputs an identity verification state, including: Retrieving a pre-stored standard face feature template from the identity identification information; Calculating the Euclidean distance between the face biological feature sequence and the standard face feature template; Comparing the Euclidean distance with a similarity threshold, and if the distance is less than the threshold, marking the verification as passed, otherwise marking the verification as failed; Sending the identity verification state together with a confidence score to the examination seat management module.

5. The test management system for face recognition and automatic allocation of seat for examination as claimed in claim 4 wherein, The identity verification module triggers a retry mechanism when the output identity verification state is verification failed, including: Controlling the biological feature acquisition module to re-collect a face image; Repeating the feature extraction and matching process, and accumulating the number of retries; If the accumulated number of retries reaches a maximum attempt value and the verification still fails, activating an alarm signal and pausing the examination seat allocation process.

6. The test management system for face recognition and automatic allocation of seat for examination as claimed in claim 1 wherein, The examination seat management module executes a dynamic allocation algorithm to generate an examination seat allocation instruction, including: Obtain real-time seat occupation state table from seat state database; filter available seat set according to examinee category and seat type matching rule; Select optimal seat from available seat set by applying load balancing strategy; generate seat allocation instruction and update occupation state in seat state database.

7. The test management system for face recognition and automatic allocation of seat for examination as claimed in claim 6 wherein, The seat management module selects the optimal seat from the available seat set by applying the load balancing strategy, comprising: Calculate the historical usage frequency and current distribution density of each available seat; select the seat with the lowest usage frequency and uniform distribution based on the greedy algorithm; record the selection result in the allocation log and mark the seat as occupied.

8. The test management system for face recognition and automatic allocation of seat for examination as claimed in claim 1 wherein, The output interaction module receives the seat allocation instruction, synthesizes visual interface content and voice broadcast content, and presents them to the examinee through the liquid crystal display and the loudspeaker, comprising: Map the seat number and position information in the seat allocation instruction to the graphical interface template; generate dynamic display content containing seat route map; Convert the text prompt information into a voice signal, adjust the speech speed and volume parameters; render the display content through the liquid crystal display and play the voice signal through the loudspeaker.

9. The test management system for face recognition and automatic allocation of seat for examination as claimed in claim 1 wherein, When the identity verification module uses pattern recognition algorithm for feature matching, the following steps are performed: Retrieve pre-stored multi-factor authentication data, including password hash value and backup contact information, from the identity information; Match the face biometric feature sequence with the pre-stored standard face feature template and calculate the matching degree; Simultaneously verify whether the password hash value is consistent with the input password; If both face matching degree and password verification pass, output the identity verification status as pass; If the face matching degree passes but the password verification fails, or the face matching degree fails but the password verification passes, trigger the backup authentication process for secondary verification through SMS verification code or fingerprint recognition.

10. The test management system for face recognition and automatic allocation of seat for examination as claimed in claim 6 wherein, When the seat management module accesses the seat state database based on the identity verification status, it also integrates environmental monitoring data and performs the following steps: Obtain real-time examination room temperature data, light intensity data and noise level data from environmental sensors; Analyze the examination room temperature data, light intensity data and noise level data to calculate the comfort index of each seat; Adjust the seat allocation priority according to the comfort index, and preferentially allocate seats with high comfort index; Generate optimized seat allocation instruction in combination with examinee category and seat type matching rule.

Images