Face recognition security authentication system and method based on edge calculation

Abstract

The invention relates to the technical field of safety certification, in particular to a face recognition safety certification system and method based on edge computing, and the system comprises a binocular camera module which is used for synchronously capturing a face image pair; the ambient light sensor module is used for monitoring ambient light intensity and dynamically adjusting camera exposure parameters; the AI acceleration processing module is integrated with an NPU chip and operates a lightweight face recognition model to perform feature extraction and living body detection; the security encryption module is integrated with an encryption chip and is used for carrying out encryption storage and decryption comparison on the biological characteristic template; the data storage module is used for locally storing the encrypted biological characteristic template; and the system login module is registered to an operation system bottom layer authentication framework and submits an authentication token to realize password-free login. High-safety and high-precision offline face recognition and system-level seamless login are realized.

Description

Technical Field

[0001] This application relates to the field of security authentication technology, specifically to a face recognition security authentication system and method based on edge computing. Background Technology

[0002] With the development of information technology, facial recognition is increasingly widely used in office terminal identity authentication. However, existing technologies have significant limitations: First, traditional monocular cameras cannot acquire depth information, have weak liveness detection capabilities, are susceptible to two-dimensional forgery attacks such as photos and videos, and their recognition accuracy drops significantly under complex lighting conditions. Second, mainstream cloud-based facial recognition solutions rely on network transmission, with biometric data stored and processed in the cloud, posing a risk of leakage and failing to meet the high security requirements of offline scenarios such as enterprise intranets. Third, traditional edge computing devices have limited computing power, making it difficult to process the high-dimensional data streams generated by binocular cameras in real time, resulting in slow recognition speed, low accuracy, and a recognition accuracy of less than 85% in complex environments.

[0003] In addition, existing solutions face technical obstacles in integrating with the underlying operating system, making it impossible to seamlessly take over the system login process and resulting in a poor user experience.

[0004] Therefore, there is an urgent need in this field for a facial recognition authentication solution that can operate in weak network or offline environments, has high security and high recognition rate, and can be deeply integrated into the system login process. Summary of the Invention

[0005] To address the above problems, this invention provides a face recognition security authentication system and method based on edge computing.

[0006] In a first aspect, the present invention provides a face recognition security authentication system based on edge computing, comprising: A binocular camera module is used to simultaneously acquire two facial images that form an image pair; An ambient light sensor module is communicatively connected to the binocular camera module and is used to monitor the ambient light intensity in real time; the binocular camera module is configured to dynamically adjust exposure parameters based on the ambient light intensity. The AI ​​acceleration processing module is communicatively connected to the binocular camera module and integrates an NPU chip. The AI ​​acceleration processing module is configured to: run a lightweight face recognition model, extract features from the face images from the binocular camera module, generate biometric templates, and perform liveness detection based on the face images from the binocular camera module. A security encryption module, connected to the AI ​​acceleration processing module, integrates an encryption chip. The security encryption module is configured to: encrypt the biometric template from the AI ​​acceleration processing module and send it to the data storage module, and to retrieve the encrypted biometric template from the data storage module for decryption and comparison during authentication. The data storage module is connected to the security encryption module and the AI ​​acceleration processing module, and is used to locally store the encrypted biometric template. The system login module is connected to the security encryption module; the system login module is configured to: register with the underlying authentication framework of the operating system, and submit an authentication token to the authentication framework after the face authentication is successful, so as to trigger the operating system to perform a passwordless login action; As a preferred embodiment of the technical solution of this invention, the binocular camera module uses two cameras arranged side-by-side with a fixed baseline distance to synchronously expose and acquire left and right views constituting an image pair under the control of the same trigger signal.

[0007] As a preferred embodiment of the technical solution of the present invention, during the system initialization phase, the two cameras of the binocular camera module are binocularly calibrated to obtain camera parameters including focal length, distortion coefficient and relative position relationship; and a stereo correction mapping table for real-time image processing is generated based on the camera parameters so that the image rows captured by the two cameras are aligned.

[0008] By pre-compiling the mapping table between bi-target localization and stereo correction, complex geometric correction operations are transformed into efficient table lookup operations. This reduces the computational overhead and latency of real-time image processing, ensuring that the system meets real-time requirements.

[0009] As a preferred embodiment of the technical solution of the present invention, the binocular camera module integrates a controller and a fill light. The controller is configured to dynamically adjust the exposure parameters based on the ambient light intensity, specifically configured as follows: Receive real-time ambient light intensity data from the ambient light sensor module; Based on ambient light intensity data, the target exposure time and target fill light intensity are determined by querying a predefined exposure parameter table. The exposure parameter table stores the mapping relationship between different ambient light intensity ranges and the camera target exposure time and the fill light target fill light intensity. The target exposure time parameter is written into the camera sensor register via the I²C control bus, and the fill light is controlled to provide fill light at the target fill light intensity.

[0010] By dynamically adjusting exposure and fill light through a closed-loop control strategy, the system gains environmental adaptability. This ensures that the system can capture correctly exposed and detailed facial images in complex lighting scenarios such as strong light, low light, and backlight, improving the system's robustness and recognition rate in real-world environments.

[0011] As a preferred embodiment of the technical solution of the present invention, the AI ​​acceleration processing module runs a lightweight face recognition model to extract features from the face image from the binocular camera module and generate a biometric template, specifically including: Face detection and key point localization are performed on a single face image, and spatial normalization is performed on the detected face image based on the key points; A lightweight face recognition model is used to extract features from a normalized face image and generate a fixed-dimensional feature vector as the biometric template.

[0012] By preprocessing face alignment and spatial normalization, the influence of pose and angle changes on feature extraction is eliminated, making the subsequently extracted feature templates more discriminative and stable.

[0013] As a preferred embodiment of the technical solution of the present invention, the AI ​​acceleration processing module performs liveness detection based on the face image pair, specifically including: A stereo matching algorithm is performed on the synchronously captured image pairs to generate depth maps of the face regions; Analyze the spatial contour features of the depth map to determine whether the captured object is a real living body with a three-dimensional structure.

[0014] Using depth maps generated by binocular vision for liveness detection can effectively defend against two-dimensional attacks (such as photos and screen captures). By analyzing three-dimensional spatial contour features, it is possible to accurately distinguish between real and fake faces, improving the accuracy of liveness detection and enhancing the system's security defense capabilities.

[0015] As a preferred embodiment of the technical solution of this invention, the security encryption module encrypts and decrypts the biometric template, specifically including: The security encryption module receives the plaintext registration biometric template from the AI ​​acceleration processing module, calls the national cryptographic algorithm engine integrated in the security encryption module, uses the internally protected data encryption key to encrypt the plaintext registration biometric template, and outputs the generated ciphertext registration biometric template to the data storage module. During authentication, the security encryption module performs the following steps: Obtain the encrypted registration biometric template from the data storage module; In the secure environment inside the encryption chip, the encrypted registration biometric template is decrypted into a plaintext registration biometric template; Within the secure environment, the plaintext registered biometric template is compared with the biometric template to be authenticated from the AI-accelerated processing module, and the comparison result is output.

[0016] All encryption and decryption operations, as well as sensitive feature comparisons, are performed within the encryption chip, ensuring that the plaintext biometric template never leaves the physical boundaries of the secure chip. This "data-within-the-chip" architecture achieves the highest level of data security, ensuring that biometric data will not be leaked even if the terminal system is compromised.

[0017] As a preferred embodiment of the technical solution of the present invention, the system login module is configured to seamlessly take over the login process in the following manner: During system deployment, it registers itself as the corresponding underlying authentication component based on the operating system type; After successful authentication, an authentication token conforming to the operating system's security specifications is generated. The authentication token is submitted to the operating system's login management process through the registered authentication component interface; Once the operating system kernel verifies the validity of the token, it authorizes the user to access the system desktop, thus bypassing the password input step. The user does not need to actively trigger any applications; the authentication process is completed seamlessly in the background, achieving a seamless closed loop from identification to login.

[0018] As a preferred embodiment of the technical solution of the present invention, registration as the corresponding underlying authentication component according to the operating system type specifically includes: In Windows systems, register as a credential provider; In Linux systems, it is registered as a pluggable authentication module; In the domestically developed operating system, registration is performed through the provided biometric authentication security interface.

[0019] As a preferred embodiment of the technical solution of the present invention, the system is configured to execute a registration process, specifically including: In response to a registration command, the binocular camera module synchronously acquires facial images for registration; wherein, the binocular camera module dynamically adjusts exposure parameters based on the ambient light intensity obtained from the ambient light sensor module; Liveness detection is performed based on face image pairs using an AI-accelerated processing module. If the liveness detection result is passed, the registered biometric template is extracted from the face image through the AI ​​acceleration processing module; The registered biometric template is encrypted using a national cryptographic algorithm through a security encryption module to generate a ciphertext registered biometric template. The encrypted biometric template is stored in the data storage module.

[0020] As a preferred embodiment of the technical solution of the present invention, the system is configured to execute an authentication login process, specifically including: In response to an authentication request, the binocular camera module synchronously acquires the face image of the person to be authenticated; wherein, the binocular camera module dynamically adjusts the exposure parameters based on the ambient light intensity obtained from the ambient light sensor module; Liveness detection is performed based on face image pairs using an AI-accelerated processing module. If the liveness detection result is passed, the AI-accelerated processing module extracts the biometric template to be authenticated from the face image. The secure encryption module retrieves the encrypted registration biometric template from the data storage module and decrypts it into a plaintext registration biometric template. The security encryption module compares the biometric template to be authenticated with the plaintext registered biometric template. If the comparison is successful, an authentication pass signal is submitted to the operating system to take over the system login process and enable passwordless login.

[0021] Secondly, the present invention also provides a face recognition security authentication method based on edge computing applied to the system described in the first aspect, the method comprising: Registration process: The facial images of the person registering are captured using a binocular camera module; The AI-accelerated processing module extracts registered biometric templates and performs liveness detection. If the liveness detection passes, the registered biometric template is encrypted by the security encryption module and then stored in the data storage module. Authentication login process: The system uses a binocular camera module to capture facial images of the person to be authenticated. The AI-accelerated processing module extracts the biometric template to be authenticated and performs liveness detection. If the liveness detection passes, the biometric template to be authenticated is compared with the registered biometric template decrypted from the data storage module. If the comparison is successful, the authentication token is submitted to the authentication framework through the system login module to trigger the operating system to perform a passwordless login action, thereby seamlessly taking over the login process.

[0022] As a preferred embodiment of the technical solution of the present invention, the step of acquiring the face image for registration or the face image to be authenticated through the binocular camera module specifically includes: Two cameras, positioned side-by-side with a fixed baseline distance, are used to synchronously expose and capture facial images under the control of the same trigger signal. Hardware-level noise suppression and signal enhancement are performed on the acquired face images. The preprocessed face images are output as a set of spatiotemporally aligned image pairs to the AI ​​acceleration processing module.

[0023] As a preferred embodiment of the technical solution of the present invention, the following steps are performed during the system initialization phase: Perform binocular positioning on two cameras to obtain camera parameters including focal length, distortion coefficient, and relative positional relationship; A stereo correction mapping table for real-time image processing is generated based on the camera parameters to align the rows of images captured by the two cameras.

[0024] As a preferred embodiment of the technical solution of the present invention, the specific steps for dynamically adjusting the exposure parameters during image acquisition include: Receive real-time ambient light intensity data from the ambient light sensor module; Based on ambient light intensity data, the target exposure time and target fill light intensity are determined by querying a predefined exposure parameter table. The exposure parameter table stores the mapping relationship between different ambient light intensity ranges and the camera target exposure time and the fill light target fill light intensity. The target exposure time parameter is written into the camera sensor register via the I²C control bus, and the fill light is controlled to provide fill light at the target fill light intensity.

[0025] As a preferred embodiment of the technical solution of the present invention, the step of extracting the registered biometric template or the biometric template to be authenticated through the AI ​​acceleration processing module specifically includes: Face detection and key point localization are performed on a single face image, and spatial normalization is performed on the detected face image based on the key points; A lightweight face recognition model is used to extract features from a normalized face image and generate a fixed-dimensional feature vector as the biometric template.

[0026] As a preferred embodiment of the technical solution of the present invention, the liveness detection via the AI ​​acceleration processing module specifically includes: A stereo matching algorithm is performed on the synchronously captured image pairs to generate depth maps of the face regions; Analyze the spatial contour features of the depth map to determine whether the captured object is a real living body with a three-dimensional structure.

[0027] As a preferred embodiment of the technical solution of the present invention, the encryption and decryption of the biometric template through the security encryption module specifically includes: Encryption process: Receive plaintext registered biometric templates from the AI-accelerated processing module; The integrated national cryptographic algorithm engine is invoked, and the internally protected data encryption key is used to encrypt the plaintext registered biometric template; The generated encrypted registration biometric template is output to the data storage module; Decryption and comparison process: Retrieve the encrypted registration biometric template from the data storage module; In the secure environment inside the encryption chip, the encrypted registration biometric template is decrypted into a plaintext registration biometric template; Within the secure environment, the plaintext registered biometric template is compared with the biometric template to be authenticated from the AI-accelerated processing module, and the comparison result is output.

[0028] As a preferred embodiment of the technical solution of the present invention, the seamless takeover of the login process through the system login module specifically includes: During system deployment, it registers itself as the corresponding underlying authentication component based on the operating system type; After successful authentication, an authentication token conforming to the operating system's security specifications is generated. The authentication token is submitted to the operating system's login management process through the registered authentication component interface; Once the operating system kernel verifies the validity of the token, it authorizes the user to access the system desktop, thereby bypassing the password input step.

[0029] As a preferred embodiment of the technical solution of the present invention, the registration process specifically includes the following steps: In response to the registration command, the binocular camera module synchronously acquires facial images for registration; the exposure parameters of the binocular camera module are dynamically adjusted based on the ambient light intensity obtained from the ambient light sensor module. Liveness detection is performed based on face image pairs using an AI-accelerated processing module. If the liveness detection result is passed, the registered biometric template is extracted from the face image through the AI ​​acceleration processing module; The registered biometric template is encrypted using a national cryptographic algorithm through a security encryption module to generate a ciphertext registered biometric template. The encrypted biometric template is stored in the data storage module.

[0030] As a preferred embodiment of the technical solution of the present invention, the authentication login process specifically includes the following steps: In response to an authentication request, the binocular camera module synchronously captures the face image of the person to be authenticated; the exposure parameters of the binocular camera module are dynamically adjusted based on the ambient light intensity obtained from the ambient light sensor module. Liveness detection is performed based on face image pairs using an AI-accelerated processing module. If the liveness detection result is passed, the AI-accelerated processing module extracts the biometric template to be authenticated from the face image. The encrypted biometric template is obtained from the data storage module through the security encryption module and then decrypted into a plaintext biometric template. The biometric template to be authenticated is compared with the plaintext registered biometric template using a secure encryption module. If the comparison is successful, the system login module submits an authentication pass signal to the operating system to take over the system login process and enable passwordless login.

[0031] As can be seen from the above technical solutions, this application has the following advantages: Through the collaboration of the binocular camera module and the ambient light sensor module, high-quality image pairs suitable for stereoscopic vision analysis can be acquired under any lighting conditions, ensuring a high recognition success rate from the source. The AI ​​acceleration processing module integrates an NPU chip, achieving localized high-speed feature extraction and liveness detection, ensuring offline availability while significantly improving processing efficiency and energy efficiency. The cooperation between the security encryption module and the data storage module ensures that the biometric template exists in encrypted form throughout the process, fundamentally eliminating the risk of data leakage. Finally, the system login module, through deep integration with the underlying operating system, achieves a seamless connection from biometric recognition to system login, providing an extremely convenient user experience. Attached Figure Description

[0032] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 A block diagram of a system provided in an embodiment of the present invention.

[0034] Figure 2 This is a schematic diagram of the registration process provided in an embodiment of the present invention.

[0035] Figure 3 This is a schematic diagram of the authentication process provided in an embodiment of the present invention. Detailed Implementation

[0036] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0038] like Figure 1 As shown, this embodiment of the invention provides a face recognition security authentication system based on edge computing, comprising: A binocular camera module is used to simultaneously acquire two facial images that form an image pair; An ambient light sensor module is communicatively connected to the binocular camera module and is used to monitor the ambient light intensity in real time; the binocular camera module is configured to dynamically adjust exposure parameters based on the ambient light intensity. The AI ​​acceleration processing module is communicatively connected to the binocular camera module and integrates an NPU chip. The AI ​​acceleration processing module is configured to: run a lightweight face recognition model, extract features from the face images from the binocular camera module, generate biometric templates, and perform liveness detection based on the face images. A security encryption module, connected to the AI ​​acceleration processing module, integrates an encryption chip. The security encryption module is configured to: encrypt the biometric template from the AI ​​acceleration processing module and send it to the data storage module, and to retrieve the encrypted biometric template from the data storage module for decryption and comparison during authentication. The data storage module is connected to the security encryption module and the AI ​​acceleration processing module, and is used to locally store the encrypted biometric template. The system login module is connected to the security encryption module. The system login module is configured to register with the underlying authentication framework of the operating system and submit an authentication token to the authentication framework after the face authentication is successful, so as to trigger the operating system to perform a passwordless login action, thereby seamlessly taking over the login process.

[0039] The binocular camera module uses two cameras arranged side by side with a fixed baseline distance to synchronously expose and capture facial images under the control of the same trigger signal. The captured facial images are subjected to hardware-level noise suppression and signal enhancement. The pre-processed facial images are output as a set of spatiotemporally aligned image pairs to the AI ​​acceleration processing module.

[0040] In some embodiments, during the system initialization phase, the two cameras of the binocular camera module are binocularly calibrated to obtain parameters including focal length, distortion coefficients, and relative positional relationship; and a stereo correction mapping table for real-time image processing is generated based on the parameters to align the rows of images captured by the two cameras.

[0041] The specific steps for generating a stereo correction mapping table for real-time image processing based on the parameters include: Based on the rotation matrix R and translation vector T in the relative positional relationship obtained through dual-camera calibration, a rotation matrix R1 and a translation vector T are calculated for each of the two cameras for stereo correction, so that the image planes of the two cameras are coplanar and aligned. For each target pixel coordinate in the target correction image and the known world coordinates, the intrinsic parameter matrix and distortion coefficient of each camera are calculated respectively. Based on the corresponding rotation matrix, intrinsic parameter matrix and distortion coefficient, the source pixel coordinates in the original distorted image are calculated by reverse mapping. The mapping relationship between the target pixel coordinates and the calculated source pixel coordinates is stored as the stereo correction mapping table; In real-time image processing, the original distorted image is directly remapped into a row-aligned corrected image by looking up the stereo correction mapping table.

[0042] In this embodiment of the invention, during the system initialization phase, the classic Zhang Zhengyou calibration method is used to calibrate the two cameras in the binocular camera module, and the relative positional relationship between them is calculated. The specific implementation steps are as follows: 1. Fabrication of a high-precision calibration plate Prepare a checkerboard calibration board with known and precise dimensions (e.g., each black and white square has a side length of 30 mm ± 0.01 mm). The checkerboard pattern must have a sufficient number of corner points (e.g., an 8x10 grid) and ensure that the pattern is flat.

[0043] 2. Acquire multiple sets of calibration images Place the calibration board within the common field of view of the binocular cameras.

[0044] The control unit drives the binocular cameras to simultaneously acquire at least 10 sets (ideally 15-20 sets) of calibration board image pairs in different poses. These poses should cover the entire field of view and include various angles such as flat, tilted, and rotated to ensure the adequacy of parameter estimation.

[0045] The acquired left and right image pairs are stored separately.

[0046] 3. Extract corner pixel coordinates For each acquired calibration board image, an image processing algorithm is used to automatically identify and extract the sub-pixel precision image pixel coordinates of all interior corner points on the chessboard grid. u , v ).

[0047] Record the coordinates of each corner point in the calibration plate's world coordinate system. X , Y , Z Since the calibration plate is flat, it is usually set... Z =0.

[0048] 4. Monocular camera parameter initialization and optimization (performed separately) Step 4.1: Calculate the homography matrix For each image from each camera, based on the extracted corner pixel coordinates ( u , v ) and its corresponding world coordinates ( X , Y Calculate a homography matrix (0, 0). H This matrix describes the projection relationship from the calibration plate plane to the camera image plane in the world coordinate system.

[0049] Step 4.2: Solve for the initial intrinsic parameters Using the calculated multiple homography matrices H By solving a system of linear equations, the camera's intrinsic parameter matrix can be preliminarily estimated. K Initial intrinsic parameter matrix K The format is as follows:

[0050] in, The focal length (in pixels) in the x and y directions. () are the coordinates of the main point.

[0051] Step 4.3: Solve for the initial extrinsic parameters Based on the homography matrix of each image H and preliminary intrinsic parameter matrix K Calculate the extrinsic parameter matrix corresponding to the image, i.e., the rotation matrix. R Translation vectort This set of extrinsic parameters describes the orientation of the calibration board in the image.

[0052] Step 4.4: Nonlinear Optimization The initial parameters obtained in the above steps contain errors. Next, a camera model incorporating radial and tangential distortion is constructed. Distortion coefficients. D Typically includes [ k 1, k 2, p 1, p 2, k 3).

[0053] Starting with the initial intrinsic parameters, extrinsic parameters, and distortion coefficients, and aiming to minimize the sum of reprojection errors at all corner points, a nonlinear optimization algorithm (such as the Levenberg-Marquardt algorithm) is executed.

[0054] Reprojection error: Reprojecting corner points in the world coordinate system using the currently estimated camera parameters ( K , D , R , t The image is then projected back into the pixel coordinate system to obtain the calculated pixel coordinates. The Euclidean distance between these coordinates and the extracted actual corner pixel coordinates is the reprojection error.

[0055] After optimization, the accurate intrinsic parameter matrix of the camera is obtained. K and distortion coefficient D .

[0056] 5. Binocular joint calibration and stereo parameter calculation After obtaining the internal parameters of the left and right cameras respectively. K 1, K 2 and distortion coefficient D 1, D After step 2, perform binocular joint calibration.

[0057] By utilizing the corner pairs matched in the left and right image pairs, and through a combination of linear methods and nonlinear optimization, the relative positional relationship between the two cameras is jointly solved, i.e., the rotation matrix. R Translation vector T ; At the same time, the intrinsic parameters of the left and right cameras can be further fine-tuned to obtain the optimal solution with the minimum overall reprojection error.

[0058] Through the above steps, the system obtains all the camera parameters necessary for subsequent stereo vision calculations: Left camera internal reference K 1 and distortion coefficient D 1; Right camera internal reference K2 and distortion coefficient D 2; Rotation matrix between two cameras R Translation vector T ; Generate a mapping table for each camera.

[0059] Taking the left camera as an example, generate its calibration mapping table map1: Define the size of the target image to be corrected, for example, 1280×720.

[0060] For each target pixel coordinate in the corrected image ( For example, (300, 200): a. Convert it to a point in the normalized camera coordinate system. .

[0061] b. Apply the inverse of the correction rotation matrix calculated in step a. Then, transform it back to the normalized coordinate system of the original camera.

[0062] c. Using distortion coefficients D 1. Calculate the distorted normalized coordinates of the point in the original image.

[0063] d. Through the original intrinsic parameter matrix K 1. Transform the distorted normalized coordinates back to the original image pixel coordinate system to obtain the source coordinates. ).

[0064] e. Due to the calculated ( ) are usually floating-point numbers, while pixel coordinates are integers, so their positions need to be recorded for subsequent interpolation calculations (such as bilinear interpolation).

[0065] Will( ) and the calculated ( The correspondence between the target image and the target image is stored in the mapping table map1. map1 is an array with the same size as the target image, and each position stores one or a pair of floating-point coordinates.

[0066] Similarly, generate a calibration mapping table (map2) for the right camera, using... R 2, K 2, D 2. Repeat the above process.

[0067] During real-time system operation, for each newly acquired pair of original distorted images: The system no longer needs to perform complex matrix multiplication, rotation, or distortion calculations. Instead, it directly looks up the pre-generated mapping tables map1 and map2. Based on the source coordinates stored in the mapping tables, it obtains pixel values ​​from the original left and right images through interpolation and fills them into a new, row-aligned, corrected image.

[0068] In some embodiments, the binocular camera module integrates a controller and a fill light. The controller is configured to dynamically adjust exposure parameters based on the ambient light intensity, specifically configured as follows: Receive real-time ambient light intensity data from the ambient light sensor module; Based on ambient light intensity data, the target exposure time and target fill light intensity are determined by querying a predefined exposure parameter table. The exposure parameter table stores the mapping relationship between different ambient light intensity ranges and the camera target exposure time and the fill light target fill light intensity; the exposure parameter table is shown in Table 1. Table 1: Examples of Exposure Parameters

[0069] It should be noted that PWM % refers to the percentage of this duty cycle. 0% means the entire cycle is at a low level (the fill light is off); 100% means the entire cycle is at a high level (the fill light illuminates at maximum power); 50% means the fill light is at a high level for half the time and at a low level for the other half (the fill light illuminates at approximately 50% brightness); and 10% means that within a complete PWM signal cycle, the control signal is at a high level for only one-tenth of the entire cycle.

[0070] The target exposure time parameter is written into the camera sensor register via the I²C control bus, and the fill light is controlled to provide fill light at the target fill light intensity.

[0071] In some embodiments, the AI ​​acceleration processing module runs a lightweight face recognition model to extract and compare features from the face images from the binocular camera module, generating a biometric template, specifically including: Face detection and key point localization are performed on a single face image, and spatial normalization is performed on the detected face image based on the key points; A lightweight face recognition model is used to extract features from a normalized face image and generate a fixed-dimensional feature vector as the biometric template.

[0072] During authentication, the similarity between the feature vector to be authenticated and the registered feature vector is calculated and compared with a preset threshold to generate face comparison results.

[0073] In some embodiments, the AI ​​acceleration processing module performs liveness detection based on face image pairs, specifically including: A stereo matching algorithm is performed on the synchronously captured image pairs to generate depth maps of the face regions; Analyze the spatial contour features of the face region in the depth map; Analyze the reflectivity of human face images; Input spatial contour features and reflectivity into a lightweight classifier and output liveness detection confidence; The liveness detection result is output based on whether the confidence level exceeds a preset threshold.

[0074] 1. The specific process of face recognition: Step F1: Face Detection and Alignment The AI ​​module receives images from the binocular cameras.

[0075] Run lightweight face detection models (such as MTCNN, UltraLight, etc.) to locate face regions in images and output face bounding boxes.

[0076] Further detection of key feature points inside the face (such as the corners of the eyes, the tip of the nose, the corners of the mouth, etc., usually 5 points or 106 points).

[0077] Based on these key points, algorithms such as affine transformation are used to normalize (align) the face image space, eliminate the influence of pose tilt, and generate a standard frontal face image.

[0078] Step F2: Feature Extraction The normalized face image is then input into a lightweight backbone feature extraction model (such as the quantized MobileFaceNet you mentioned).

[0079] The NPU efficiently runs the model, converting the input image into a low-dimensional, highly discriminative feature vector. This feature vector is the biometric template.

[0080] This process is completed on the NPU, which is more than 5 times faster and consumes significantly less power compared to the CPU / GPU.

[0081] Step F3: Feature Alignment and Output During registration: The feature vector (registration template) is encrypted using a security encryption module and then stored in the database.

[0082] During authentication: The feature vector extracted in real time (the template to be authenticated) is compared with the decrypted registration template in the database to calculate similarity (such as calculating cosine distance and Euclidean distance).

[0083] Based on whether the similarity score exceeds a preset threshold, the authentication result (pass / fail) is output.

[0084] 2. The specific process of the liveness detection pathway: Step L1: Depth Map Generation The AI ​​module receives image pairs that have undergone stereo correction.

[0085] Run an efficient stereo matching algorithm (typically optimized for execution on an NPU) to calculate the disparity for each pixel in the image and generate a high-precision depth map based on the principles of triangulation.

[0086] Step L2: Activity Characterization Analyze the spatial contour features of the face region in the depth map. A real human face should have continuous and smooth depth variations and convexity and concavity that conform to anatomical structure (such as a prominent nasal bridge and sunken eye sockets), while a two-dimensional prosthesis will show invalid depth or abnormal planes / curves.

[0087] Step L3: Living Decision Input the above features into a lightweight classifier (such as an SVM or a small neural network).

[0088] The classifier integrates all the information and outputs a final liveness detection confidence score.

[0089] Based on whether the confidence level exceeds a preset threshold, output the liveness detection result (pass / fail).

[0090] In some embodiments, the security encryption module encrypts and decrypts the biometric template, specifically including: The security encryption module receives the plaintext registration biometric template from the AI ​​acceleration processing module, calls the national cryptographic algorithm engine integrated in the security encryption module, uses the internally protected data encryption key to encrypt the plaintext registration biometric template, and outputs the generated ciphertext registration biometric template to the data storage module. During authentication, and after the AI ​​acceleration processing module passes the liveness detection, the security encryption module performs the following steps: Obtain the encrypted registration biometric template from the data storage module; In the secure environment inside the encryption chip, the encrypted registration biometric template is decrypted into a plaintext registration biometric template; Within the secure environment, the plaintext registered biometric template is compared with the biometric template to be authenticated from the AI-accelerated processing module, and the comparison result is output.

[0091] The national cryptographic algorithm engine is the SM4 algorithm engine, used to encrypt and decrypt biometric templates; and / or, the security encryption module also integrates the SM2 algorithm engine, used to generate digital signatures for the comparison results.

[0092] In some embodiments, the system login module is configured to seamlessly take over the login process in the following ways: During system deployment, it registers itself as the corresponding underlying authentication component based on the operating system type; After successful authentication, an authentication token conforming to the operating system's security specifications is generated. The authentication token is submitted to the operating system's login management process through the registered authentication component interface; The operating system kernel, after verifying the validity of the token, authorizes the user to access the system desktop, thereby bypassing the password input step. This is done by registering the corresponding underlying authentication component based on the operating system type, specifically including: In Windows systems, register as a credential provider; In Linux systems, it is registered as a pluggable authentication module; In the domestically developed operating system, registration is performed through the provided biometric authentication security interface.

[0093] In summary, the system is configured to execute a registration process that specifically includes: In response to a registration command, the binocular camera module synchronously acquires facial images for registration; wherein, the binocular camera module dynamically adjusts exposure parameters based on the ambient light intensity obtained from the ambient light sensor module; Liveness detection is performed based on face image pairs using an AI-accelerated processing module. If the liveness detection result is passed, the registered biometric template is extracted from the face image through the AI ​​acceleration processing module; The registered biometric template is encrypted using a national cryptographic algorithm through a security encryption module to generate a ciphertext registered biometric template. The encrypted biometric template is stored in the data storage module.

[0094] Phase 1: High-quality liveness image acquisition Trigger command: The administrator or user initiates a registration command in the registration management interface.

[0095] Environmental perception: The system activates the ambient light sensor module to obtain real-time light intensity data of the current environment.

[0096] Parameter optimization: Based on the acquired lighting data, the binocular camera module dynamically adjusts its exposure time, gain, and fill light intensity to ensure that it can capture correctly exposed and detailed images in complex environments such as strong light and low light.

[0097] Synchronous capture: After the parameters are adjusted, the binocular camera module simultaneously captures facial images, forming an image pair.

[0098] Data transmission: The high-quality image pair is transmitted to the AI ​​acceleration processing module.

[0099] Phase Two: Liveness Detection and Feature Extraction Perform liveness detection: The AI-accelerated processing module initiates the liveness detection process based on the received image pairs. Depth information calculation: Perform stereo matching on image pairs to generate depth maps.

[0100] Comprehensive analysis of the spatial contour of the depth map (to determine whether it is a 3D entity).

[0101] Liveness Decision: Determines whether a captured object is a real live object using an internally integrated lightweight classification model.

[0102] If the liveness detection fails: the process terminates immediately, returning a "liveness detection failed" message to the user, and no further action is taken. This fundamentally prevents registration using illegal means such as photos, videos, or masks.

[0103] If successful: the process continues.

[0104] Feature template extraction: After the liveness detection is passed, the AI ​​acceleration processing module uses its lightweight face recognition model to extract features from the high-quality face image and generate a plaintext registration biometric template (a feature vector) that can uniquely represent the user's identity.

[0105] Phase Three: Secure Encryption and Storage Initiate encryption request: The AI ​​acceleration processing module sends the generated plaintext registered biometric template to the secure encryption module.

[0106] Hardware encryption is performed: The security encryption module uses a national cryptographic algorithm (such as SM4) and an internally protected key to encrypt the plaintext template in its internal encryption chip, converting it into a ciphertext registered biometric template.

[0107] Secure storage: The generated encrypted biometric template is sent to the data storage module for permanent local storage. At this point, the user's original biometric information is protected.

[0108] Completion feedback: The system returns a "registration successful" signal to the interface, and the entire process ends.

[0109] In some embodiments, the system is configured to execute an authentication login process, specifically including: In response to an authentication request, the binocular camera module synchronously acquires the face image of the person to be authenticated; wherein, the binocular camera module dynamically adjusts the exposure parameters based on the ambient light intensity obtained from the ambient light sensor module; Liveness detection is performed based on face image pairs using an AI-accelerated processing module. If the liveness detection result is passed, the AI-accelerated processing module extracts the biometric template to be authenticated from the face image. The secure encryption module retrieves the encrypted registration biometric template from the data storage module and decrypts it into a plaintext registration biometric template. The security encryption module compares the biometric template to be authenticated with the plaintext registered biometric template. If the comparison is successful, an authentication pass signal is submitted to the operating system to take over the system login process and enable passwordless login.

[0110] Phase 1: Liveness Verification and Feature Extraction Authentication is triggered when the user wakes up the device or clicks the login screen, and the system initiates the authentication process.

[0111] Environmentally Adaptive Image Acquisition: Similar to the registration process, the binocular camera module dynamically adjusts its parameters based on ambient light sensor data and simultaneously captures the current user's face image pair.

[0112] Liveness detection: The AI-accelerated processing module performs liveness detection on the acquired image pairs.

[0113] If the authentication fails: the process terminates immediately and returns an authentication failure message. No further feature comparisons or decryption operations will be performed, effectively preventing spoofing attacks.

[0114] If successful: the process continues.

[0115] Extracting features to be authenticated: The AI-accelerated processing module extracts biometric templates to be authenticated from the image.

[0116] Phase Two: Secure Verification and Identity Determination Initiate decryption request: The system sends a command to the security encryption module to retrieve and decrypt the user's encrypted registration biometric template from the data storage module.

[0117] Security Decryption: The security encryption module performs the following operations: Read the encrypted registration template from the data storage module.

[0118] Inside the encryption chip, a secure key is used to decrypt it into a plaintext registered biometric template. This plaintext template never leaves the security boundaries of the encryption chip.

[0119] Intra-security domain comparison: In the secure environment inside its chip, the security encryption module performs similarity calculations (such as cosine distance calculation) between the decrypted plaintext registered biometric template and the biometric template to be authenticated.

[0120] If the comparison fails (similarity is below the threshold): the process terminates and an authentication failure message is returned.

[0121] If the comparison is successful (similarity above the threshold): the process enters the final stage. At this point, the system simultaneously confirms both the liveness and legal identity of the individual.

[0122] Phase 3: System Takeover and Login Sending a login signal: After receiving a successful authentication signal, the system login module generates a system-recognizable authentication token and submits this token to the operating system kernel through a security interface provided by the operating system (such as Winlogon in Windows or the PAM module in Linux).

[0123] Take over the login process: After the operating system kernel verifies the validity of the token, it interrupts or bypasses the traditional password input interface and directly authorizes the user to access the system desktop.

[0124] Login complete: Users can log in to the system without a password. This achieves highly secure and accurate offline facial recognition and seamless system-level login.

[0125] This invention also provides an edge computing-based face recognition security authentication method applied to the system described in the above embodiments, the method comprising: like Figure 2 The registration process is shown below: The facial images of the person registering are captured using a binocular camera module; The AI-accelerated processing module extracts registered biometric templates and performs liveness detection. If the liveness detection passes, the registered biometric template is encrypted by the security encryption module and then stored in the data storage module. like Figure 3 As shown, the authentication login process is as follows: The system uses a binocular camera module to capture facial images of the person to be authenticated. The AI-accelerated processing module extracts the biometric template to be authenticated and performs liveness detection. If the liveness detection passes, the biometric template to be authenticated is compared with the registered biometric template decrypted from the data storage module. If the comparison is successful, the authentication token is submitted to the authentication framework through the system login module to trigger the operating system to perform a passwordless login action, thereby seamlessly taking over the login process.

[0126] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An edge computing based face recognition security authentication system, characterized in that, The system comprises: a binocular camera module for synchronously collecting two-way face images constituting an image pair; an ambient light sensor module in communication with the binocular camera module for real-time monitoring of ambient light intensity; the binocular camera module is configured to dynamically adjust exposure parameters based on ambient light intensity; an AI acceleration processing module in communication with the binocular camera module and integrated with an NPU chip; the AI acceleration processing module is configured to run a lightweight face recognition model, extract features from face images from the binocular camera module, and generate a biometric feature template; and perform liveness detection based on face image pairs from the binocular camera module; a secure encryption module connected to the AI acceleration processing module and integrated with an encryption chip; the secure encryption module is configured to encrypt the biometric feature template from the AI acceleration processing module and send it to the data storage module, and decrypt and compare the encrypted biometric feature template from the data storage module during authentication; a data storage module connected to the secure encryption module and the AI acceleration processing module for locally storing encrypted biometric feature templates; a system login module connected to the secure encryption module; the system login module is configured to register with the underlying authentication framework of the operating system and submit an authentication token to the authentication framework after face authentication to trigger the operating system to perform password-free login. 2.The edge computing based face recognition security authentication system according to claim 1, characterized in that, The binocular camera module comprises two cameras arranged side by side with a fixed baseline distance, for synchronously exposing and collecting left and right views constituting an image pair under the control of the same trigger signal. 3.The edge computing based face recognition security authentication system according to claim 2, characterized in that, During system initialization, the two cameras of the binocular camera module are double-targeted to obtain camera parameters including focal length, distortion coefficient, and relative position relationship; and a stereo rectification mapping table for real-time image processing is generated based on the camera parameters to align the images captured by the two cameras. 4.The edge computing based face recognition security authentication system according to claim 3, characterized in that, The binocular camera module is integrated with a controller and a fill light, the controller is configured to dynamically adjust exposure parameters based on the ambient light intensity, specifically configured to: receive real-time ambient light intensity data from the ambient light sensor module; based on the ambient light intensity data, determine the target exposure time and the target fill light intensity by querying the pre-defined exposure parameter table, which stores the mapping relationship between different ambient light intensity intervals and camera target exposure time and fill light target fill light intensity; write the target exposure time parameter into the camera sensor register through the I²C control bus, and control the fill light to fill light at the target fill light intensity. 5.The edge computing based face recognition security authentication system according to claim 4, characterized in that, The AI acceleration processing module runs a lightweight face recognition model to extract features from face images from the binocular camera module to generate a biometric feature template, specifically including: performing face detection and key point positioning on a single face image, and performing spatial normalization processing on the detected face image based on the key points; extracting features from the normalized face image through the lightweight face recognition model to generate a fixed-dimension feature vector as the biometric feature template. 6.The edge computing based face recognition security authentication system according to claim 4, characterized in that, The AI acceleration processing module performs live body detection based on a pair of face images, specifically including: performing a stereo matching algorithm on the synchronously captured image pair to generate a depth map of the face region; analyzing the spatial profile features of the depth map to determine whether the captured object is a real live body with a three-dimensional structure. 7.The edge computing based face recognition security authentication system according to claim 6, characterized in that, The secure encryption module encrypts and decrypts the biometric template, specifically including: The secure encryption module receives the plaintext registration biometric template from the AI acceleration processing module, calls the national encryption algorithm engine integrated in the secure encryption module, uses the internally protected data encryption key to encrypt the plaintext registration biometric template, and outputs the generated ciphertext registration biometric template to the data storage module; During authentication, and after the live body detection of the AI acceleration processing module passes, the secure encryption module performs the following steps: obtain the ciphertext registration biometric template from the data storage module; decrypt the ciphertext registration biometric template into a plaintext registration biometric template in a secure environment inside the encryption chip; perform a similarity comparison between the plaintext registration biometric template and the to-be-authenticated biometric template from the AI acceleration processing module in the secure environment; output the comparison result to the system login module. 8.The edge computing based face recognition security authentication system according to claim 7, characterized in that, The system login module is configured to seamlessly take over the login process by: During system deployment, register itself as the corresponding underlying authentication component according to the operating system type; After authentication passes, generate an authentication token that meets the security specifications of the operating system; submit the authentication token to the login management process of the operating system through the registered authentication component interface; wherein the operating system kernel authorizes the user to access the system desktop after verifying the validity of the token, thereby bypassing the password input step. 9.The edge computing based face recognition security authentication system according to claim 8, characterized in that, Registering as the corresponding underlying authentication component according to the operating system type specifically includes: In the Windows system, register as a credential provider; In the Linux system, register as a pluggable authentication module; In the ChinaSoft operating system, register through the provided biometric authentication security interface.

10. A face recognition security authentication method based on edge computing applied to the system of claims 1-9, characterized in that, The method includes: registration process: collect registration face images through the binocular camera module; extract the registration biometric template through the AI acceleration processing module and perform live body detection; if the live body detection passes, store the registration biometric template to the data storage module after encryption by the secure encryption module; authentication login process: collect the to-be-authenticated face image through the binocular camera module; extract the to-be-authenticated biometric template through the AI acceleration processing module and perform live body detection; if the live body detection passes, compare the to-be-authenticated biometric template with the registration biometric template decrypted from the data storage module; if the comparison is successful, submit an authentication token to the authentication framework through the system login module to trigger the operating system to perform a password-free login action, thereby seamlessly taking over the login process.

Images