Cooperative anti-interference access control system based on face recognition

Abstract

The invention discloses a collaborative anti-interference access control system based on face recognition, and belongs to the technical field of access control systems. The cooperative anti-interference access control system comprises a multi-mode sensing acquisition module, an image preprocessing and anti-interference module, a face recognition and living body verification module, a cooperative control and decision module, a remote management and iteration module and an access control execution module. According to the access control system, the working mode can be adjusted in real time according to environmental changes, the influence of various interference factors is effectively resisted, a more stable and reliable face recognition access control service is provided for a user, the complementarity of multi-source data can be fully utilized by the access control system, the collected target image is optimized and processed, and the user experience is improved. The influence of interference factors such as light and noise on the image is reduced, the reliability and the accuracy of data are effectively improved, and the anti-interference capability and the face recognition accuracy of the system in a complex environment are enhanced.

Description

Technical Field

[0001] This invention relates to the field of access control system technology, specifically to a collaborative anti-interference access control system based on facial recognition. Background Technology

[0002] As an advanced means of identity verification and security management, facial recognition access control systems can quickly and accurately identify employees in public places such as office buildings, residential communities, and schools, enabling contactless access, improving office access efficiency, enhancing the security of public areas, and effectively preventing unauthorized entry.

[0003] For example, patent CN212112600U discloses a facial recognition access control system, including an access control subsystem, a visitor subsystem, a reception subsystem, and a central computer room. The access control subsystem, visitor subsystem, and reception subsystem are all connected to the central computer room via a local area network (LAN). The central computer room is also connected to a cloud management platform via the LAN. The access control subsystem includes multiple sets of office area access control systems and residential access control systems. Each of the visitor subsystem, office area access control system, and residential access control system includes a facial recognition camera. The top of the facial recognition camera is equipped with a mounting mechanism that offers multiple installation methods. This allows for wider applicability of the facial recognition camera and provides a larger adjustment range, making adjustment more convenient.

[0004] For example, patent CN112133012B discloses a facial recognition access control system and method for a hospital inpatient building. The system includes: a patient and medical staff data server, a visitor appointment registration server, a facial information database, a main control console, and entrance and exit facial recognition access control gates installed on the visitor passage. The entrance facial recognition access control gate is used to open or close the corresponding access control gate based on the visitor's information and appointment registration information when it is determined that the person passing through is a visitor. The administrator can set the number of visitors allowed to enter the inpatient building at different times through the main control console, and identify and count visitors entering and leaving the inpatient building through the entrance and exit facial recognition access control gates. Then, the entrance facial recognition access control gate is opened or closed according to the settings of the main control console and the visitor appointment status, thereby achieving efficient control of the number of visitors in the hospital inpatient building.

[0005] For example, patent CN114187684B discloses an access control device and management system based on facial recognition. The access control device includes a camera module, a ranging module, a facial recognition module, a communication module, a Bluetooth module, and a door lock module. The facial recognition-based access control device can wake up the facial recognition module through the ranging module to start facial recognition, use the camera module to collect facial information, and if a match is successful, send an opening command to the Bluetooth module through the communication module, so that the Bluetooth module controls the door lock module to drive the motor to open the door. In addition, residents can also open the door via touch screen through the display module or remotely through a remote terminal device. However, existing facial recognition access control systems face multiple anti-interference challenges in practical applications. Abnormal lighting can cause feature capture failure, and occlusion of the target face and changes in facial state can easily cause incomplete feature extraction, ultimately leading to a decrease in recognition rate, insufficient system reliability, and difficulty in adapting to complex and diverse actual scenario requirements.

[0006] To address the aforementioned issues, there is an urgent need for innovative design based on the existing access control system. Summary of the Invention

[0007] The purpose of this invention is to provide a collaborative anti-interference access control system based on face recognition, so as to solve the problems mentioned in the background art, such as insufficient recognition rate and reliability of some existing systems, which are difficult to adapt to the needs of complex and diverse actual scenarios.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a collaborative anti-interference access control system based on face recognition, the collaborative anti-interference access control system comprising: The multimodal perception and acquisition module's core hardware includes an RGB infrared spectral camera, supplementary lighting equipment, environmental sensors, and distance sensors. The module is used to acquire 2D texture, 3D depth, and infrared thermal radiation data of human faces, and simultaneously perceive environmental interference parameters to provide raw data support for anti-interference processing.

[0009] The image preprocessing and anti-interference module incorporates core technologies such as adaptive filtering, backlight compensation, pose correction, blur restoration, and noise filtering algorithms. This module optimizes the acquired raw images, eliminates the effects of abnormal lighting, image blur, and background interference, and outputs high-quality face images, reducing interference pressure on the recognition algorithm.

[0010] The face recognition and liveness verification module incorporates core technologies including multimodal fusion algorithms, ArcFace-Lite feature extraction algorithms, 3D reconstruction, micro-expression detection, and dynamic action verification. The module extracts core facial features and uses 2D, 3D, and infrared multi-dimensional verification to accurately distinguish between real people and images or masks, ensuring the accuracy of identity verification.

[0011] The collaborative control and decision-making module, whose core components include an edge processor, an access controller, and an alarm device, receives facial recognition results, combines them with two-factor authentication to decide whether to allow access, triggers an audible and visual alarm when an abnormal attack is detected, and simultaneously drives an electromagnetic lock to complete the switching control, thus realizing a closed loop of recognition, decision-making, and execution.

[0012] The remote management and iteration module consists of a management platform, a local database, and network transmission components. This module is used to remotely configure user permissions, view access logs, and monitor device status. It continuously improves the system's anti-interference capabilities by incrementally learning and optimizing the algorithm model, and supports the retrieval of locally cached data in the event of a network outage.

[0013] The access control execution module receives instructions from the collaborative control and decision-making module, drives the electromagnetic lock to complete the unlocking and locking actions, and realizes authorized access or prohibition of access.

[0014] Preferably, the environmental sensor includes a light sensor and a temperature and humidity sensor, which collects light intensity and ambient temperature and humidity in real time to provide environmental parameters for the anti-interference algorithm; when the light sensor detects a strong light environment, it automatically activates the camera's HDR mode and reduces the brightness of the fill light; when the light sensor detects a low light environment, it triggers infrared fill light and switches to infrared imaging mode to adaptively acquire images in the environment.

[0015] Preferably, the distance sensor detects the distance between the face and the device, filtering out invalid acquisitions of targets that are too close or too far away, thus preventing image copying attacks; when the distance sensor detects that the distance to the acquisition target exceeds the normal acquisition range, it immediately pauses the recognition and prompts the user to adjust the distance; when the temperature and humidity sensor detects extreme environments, it activates a hardware protection mechanism to prioritize the stability of core data acquisition.

[0016] Preferably, the anti-interference methods of the image preprocessing and anti-interference module include: image preprocessing anti-interference, which eliminates interference noise in single-modal images through basic image processing techniques, removes salt-and-pepper and Gaussian noise through adaptive filtering, solves uneven lighting through backlight compensation, improves motion and out-of-focus blur through blur repair, and corrects facial angle deviation through pose correction, thereby improving image quality from the bottom layer; multimodal fusion anti-interference, which fuses three types of modal information: RGB image, infrared thermal imaging, and 3D depth data. The RGB channel preserves facial texture details, the infrared channel distinguishes the thermal radiation characteristics of live objects from fakes, and the 3D channel provides facial contours. A weight allocation algorithm is used to achieve multimodal feature complementarity; and adaptive algorithm anti-interference, which... Based on real-time data collected by environmental sensors, such as light intensity and noise levels, the algorithm dynamically adjusts its strategies and parameter thresholds. In strong light environments, it automatically enhances backlight compensation intensity; in high-noise scenes, it switches to a strong filtering mode; and for blurred images, it triggers a super-resolution reconstruction process. Through a preset interference type identification model, it automatically matches the optimal processing algorithm, achieving intelligent anti-interference through environmental perception and algorithm adaptation. The collaborative anti-interference mechanism works in conjunction with the front-end multimodal perception and acquisition module to receive environmental interference prediction information and initiate targeted processing in advance. It also interacts with the back-end face recognition and liveness verification module to backtrack and optimize preprocessing parameters based on recognition feedback results. Multiple modules work together to form a closed-loop anti-interference system of acquisition, processing, and recognition.

[0017] Preferably, the image preprocessing and anti-interference module's workflow is as follows: It receives raw multimodal data, including encapsulated data from the perception module via a high-speed data interface, such as RGB images, infrared thermal imaging images, 3D depth point cloud data, and environmental sensor parameters; it initializes data preprocessing, loading pre-trained face detection models and algorithm configuration parameters, initializing memory buffers, storing data by data type, and dynamically allocating thread resources based on device computing power; it analyzes environmental parameters, parsing illumination sensor data, classifying illumination levels by thresholds, calculating noise levels through image variance, and generating environmental interference labels; it performs channel-specific processing, splitting multimodal data into RGB and infrared channels for parallel processing, using an 8-bit color space for the RGB channels and converting the infrared channels to grayscale images, normalizing the dual-channel data to eliminate hardware acquisition differences; and it determines the type of interference, using backlight compensation and... Color temperature calibration; blur interference is addressed through blur repair and sharpening enhancement; pose interference is addressed through key point detection and pose correction; noise interference is addressed through adaptive filtering and noise suppression. Multi-channel data fusion employs a weighted fusion algorithm, preserving color texture in the RGB channel and enhancing facial contours in the infrared channel. Image registration technology aligns the coordinates of the two channels, and the fused data is converted to the YCbCr color space, optimizing the contrast of the luminance component. Background interference removal involves locating the face region using a face detection bounding box, expanding it by 10% as the region of interest (ROI), segmenting the background using the GrabCut algorithm, removing interfering elements such as posters and clutter, and cropping the ROI to a fixed size to meet the input requirements of the recognition algorithm. High-quality face images are output by encapsulating the processed images in a standard format, adding data tags, and transmitting them to the face recognition module via shared memory or a message queue. Simultaneously, the most recent 10 frames of data are cached for anomaly backtracking.

[0018] Preferably, the multimodal fusion algorithm adopts a three-level progressive fusion architecture, achieving deep fusion of multimodal information from the data layer, feature layer to the decision layer. The data layer fusion performs pixel-level alignment of RGB images, infrared thermal imaging images, and 3D depth maps, and calibrates the spatial coordinates of the three types of data using SIFT feature matching technology to ensure that features of the same face region correspond in different modalities. The feature layer fusion designs a multi-branch convolutional neural network, with the RGB branch extracting texture features, the infrared branch extracting thermal radiation features, and the 3D branch extracting depth geometric features, automatically allocating feature weights to each branch through an attention mechanism. The decision layer fusion adopts a weighted voting mechanism, with RGB feature matching score weighted at 0.4, infrared feature matching score weighted at 0.3, and 3D feature matching score weighted at 0.3, and the final matching degree is calculated comprehensively.

[0019] Preferably, in the ArcFace-Lite feature extraction algorithm, the input image is first processed by MTCNN to detect facial key points, and the face is aligned to a standard template through similarity transformation to eliminate the influence of pose and position offset on feature extraction. After alignment, face cropping and normalization are used to improve feature stability. Data augmentation is added during the training phase to allow the model to learn robustness to interference. During the inference phase, blurred images are first reconstructed using super-resolution before being input into the feature extraction network to avoid feature distortion caused by blur. INT8 quantization is used to compress the model weights from 32-bit floating-point to 8-bit integer, and layer fusion is used to reduce memory access frequency and improve the operating efficiency of edge devices.

[0020] Preferably, the 3D reconstruction method is as follows: point cloud generation, converting the depth map into a 3D point cloud containing the X, Y, and Z coordinates of each point; point cloud registration, registering multiple frames of point clouds into a complete face point cloud using the ICP algorithm; mesh reconstruction, converting the point cloud into a 3D mesh model using the Poisson reconstruction algorithm to restore the 3D contour of the face; and texture mapping, mapping the texture information of the RGB image to the 3D mesh to generate a textured 3D face model.

[0021] Preferably, the collaborative strategy of the collaborative control and decision-making module is as follows: receiving the feature matching score, liveness verification result, and attack type label from the face recognition module; issuing lock / unlock commands and alarm trigger commands to the access control execution module; uploading access records and abnormal events to the remote management and iteration module in real time; supporting combined verification of face and password, and face and card swipe, automatically triggered in medium-risk scenarios; automatically switching to local mode when the network connection is lost, relying on the local database to complete recognition and decision-making, and automatically synchronizing data after reconnection.

[0022] Preferably, the algorithm iteration and optimization mechanism in the remote management and iteration module includes: automatically uploading sample data of failed identification and suspected attacks, along with environmental parameters, from the edge; using the FedAvg federated learning framework to aggregate abnormal data from multiple devices in the cloud for model fine-tuning, avoiding full retraining; generating specialized optimization models for scenarios with weak anti-interference capabilities; and generating an update package of less than 10MB after quantization and compression of the optimized model, which is then pushed to the edge via differential update technology. The edge automatically updates the model during idle periods, and the updated model can only be enabled after verification on a test set.

[0023] Compared with the prior art, the beneficial effects of the present invention are: The multimodal perception acquisition module, through the coordination of RGB, infrared, and 3D visual data and environmental parameters, provides a data foundation for subsequent anti-interference processing, reducing the impact of interference factors on data quality from the source. The access control system can adjust its working mode in real time according to environmental changes, effectively resisting the influence of various interference factors and providing users with a more stable and reliable facial recognition access control service.

[0024] The image preprocessing and anti-interference module eliminates the effects of abnormal lighting, image blurring, background interference, etc., and outputs high-quality face images, providing a data foundation for subsequent face recognition and liveness verification modules. This greatly reduces the interference pressure on the recognition algorithm, enabling the system to more accurately identify user identities and effectively improve the reliability and stability of the access control system.

[0025] The face recognition and liveness verification module integrates static and dynamic features to build a comprehensive anti-counterfeiting system, effectively resisting various forgery attacks. The lightweight algorithm design is adapted to edge deployment, ensuring rapid response in access control scenarios and meeting the efficiency requirements of practical applications. The access control system can make full use of the complementarity of multi-source data, effectively improving the reliability and accuracy of data, and enhancing the system's anti-interference ability and face recognition accuracy in complex environments. Attached Figure Description

[0026] Figure 1 This is a system module diagram of the present invention.

[0027] Figure 2 This is a flowchart of the image preprocessing and anti-interference module of the present invention.

[0028] Figure 3 This is a flowchart of the present invention. Detailed Implementation

[0029] 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.

[0030] This application provides a collaborative anti-interference access control system based on face recognition. Its core modules include: a multimodal perception and acquisition module for acquiring 2D texture, 3D depth, and infrared thermal radiation data of the face, and simultaneously sensing environmental interference parameters to provide raw data support for anti-interference processing; an image preprocessing and anti-interference module for optimizing the acquired raw images, eliminating the effects of abnormal lighting, image blurring, and background interference, and outputting high-quality face images to reduce interference pressure on the recognition algorithm; and a face recognition and liveness verification module for extracting core facial features and accurately distinguishing real people from images and masks through 2D, 3D, and infrared multi-dimensional verification. The system is designed to ensure accurate identity verification. The collaborative control and decision-making module receives facial recognition results, combines them with two-factor authentication to determine whether to grant access, triggers an audible and visual alarm when an abnormal attack is detected, and simultaneously drives the electromagnetic lock to complete the opening and closing control, achieving a closed loop of recognition, decision-making, and execution. The remote management and iteration module is used to remotely configure user permissions, view access logs, and monitor device status. It continuously improves the system's anti-interference capability through incremental learning to optimize the algorithm model and supports local cached data retrieval in offline conditions. The access control execution module receives instructions from the collaborative control and decision-making module, drives the electromagnetic lock to complete unlocking and locking actions, and enables authorized or prohibited access.

[0031] Example 1, see Figures 1-3 This application proposes a collaborative anti-interference access control system based on face recognition, wherein the method of using the collaborative anti-interference access control system includes: S1. System initialization, enabling anti-interference mode, completing hardware calibration of cameras, sensors, etc., and loading multi-modal fusion and adaptive anti-interference algorithms.

[0032] In this embodiment, a multimodal perception and acquisition module is included. The core hardware includes an RGB infrared spectral camera, a supplementary lighting device, an environmental sensor, and a distance sensor. The module is used to acquire 2D texture, 3D depth, and infrared thermal radiation data of the face, and simultaneously perceive environmental interference parameters to provide raw data support for anti-interference processing.

[0033] In this embodiment, the environmental sensor includes a light sensor and a temperature and humidity sensor, which collect light intensity and ambient temperature and humidity in real time to provide environmental parameters for the anti-interference algorithm; when the light sensor detects a strong light environment, it automatically activates the camera's HDR mode and reduces the brightness of the fill light; when the light sensor detects a low light environment, it triggers infrared fill light and switches to infrared imaging mode to adaptively acquire images in the environment.

[0034] In this embodiment, the distance sensor detects the distance between the face and the device, filtering out invalid acquisitions of targets that are too close or too far away to avoid image copying attacks; when the distance sensor detects that the distance to the acquisition target exceeds the normal acquisition range, it immediately pauses the acquisition and prompts the user to adjust the distance; when the temperature and humidity sensor detects extreme environments, it activates the hardware protection mechanism to prioritize the stability of core data acquisition.

[0035] It should be noted that by using hardware clock synchronization technology, we ensure that RGB images, infrared images, and 3D depth data are acquired under the same frame timestamp, avoiding feature matching errors caused by timing deviations and ensuring the correlation of multimodal data.

[0036] S2. User face registration: In normal lighting conditions, face the camera and follow the prompts to blink and turn your head. The system simultaneously collects 2D, 3D, and infrared multimodal data and automatically filters out occlusion and posture deviation interference.

[0037] In this embodiment, a face recognition and liveness verification module is included. The core technologies include multimodal fusion algorithm, ArcFace-Lite feature extraction algorithm, 3D reconstruction, micro-expression detection, and dynamic action verification. The module is used to extract core facial features and accurately distinguish between real people and images or masks through 2D, 3D, and infrared multi-dimensional verification, ensuring the accuracy of identity verification.

[0038] In this embodiment, the multimodal fusion algorithm adopts a three-level progressive fusion architecture, achieving deep fusion of multimodal information from the data layer, feature layer to the decision layer. The data layer fusion performs pixel-level alignment of RGB images, infrared thermal images, and 3D depth maps, and calibrates the spatial coordinates of the three types of data using SIFT feature matching technology to ensure that features of the same face region correspond in different modalities. The feature layer fusion designs a multi-branch convolutional neural network, with the RGB branch extracting texture features, the infrared branch extracting thermal radiation features, and the 3D branch extracting depth geometric features. An attention mechanism is used to automatically allocate feature weights to each branch. The decision layer fusion employs a weighted voting mechanism, with RGB feature matching score weighted at 0.4, infrared feature matching score weighted at 0.3, and 3D feature matching score weighted at 0.3, and the final matching degree is calculated comprehensively.

[0039] In this embodiment, the ArcFace-Lite feature extraction algorithm first uses MTCNN to detect facial key points in the input image, and then aligns the face to a standard template through similarity transformation to eliminate the influence of pose and position offset on feature extraction. After alignment, face cropping and normalization are used to improve feature stability. Data augmentation is added during the training phase to allow the model to learn robustness to interference. During the inference phase, blurred images are first reconstructed using super-resolution before being input into the feature extraction network to avoid feature distortion caused by blur. INT8 quantization is used to compress the model weights from 32-bit floating-point to 8-bit integer, and layer fusion is used to reduce memory access frequency and improve the operating efficiency of edge devices.

[0040] In this embodiment, the 3D reconstruction method is as follows: point cloud generation, which converts the depth map into a 3D point cloud containing the X, Y, and Z coordinates of each point; point cloud registration, which uses the ICP algorithm to register multiple frames of point clouds into a complete face point cloud; mesh reconstruction, which uses the Poisson reconstruction algorithm to convert the point cloud into a 3D mesh model to restore the 3D contour of the face; and texture mapping, which maps the texture information of the RGB image to the 3D mesh to generate a textured 3D face model.

[0041] It should be noted that the multimodal fusion algorithm adopts a feature-level fusion strategy, inputting the texture features of RGB images, the depth features of 3D point clouds, and the thermal radiation features of infrared images into the feature fusion network. Key features are weighted through an attention mechanism, outputting a fused high-dimensional face feature vector to compensate for the deficiencies of single modalities and resist attacks from flat images and printed photos. The ArcFace-Lite feature extraction algorithm performs MTCNN face alignment on the preprocessed face image and inputs it into a lightweight backbone network to extract a 512-dimensional feature vector. The ArcFace loss function is used to optimize the classification boundary and improve feature discrimination. The model is quantized and compressed to 8MB to adapt to edge computing power, quickly and accurately extracting core facial features, ensuring recognition efficiency and accuracy. 3D reconstruction is based on 3D structures. The system uses light-collected depth point cloud data to generate a 3D facial mesh model via Poisson reconstruction algorithm. It calculates the depth values ​​of key points such as the tip of the nose and brow bone, compares them with preset depth thresholds for real people, and identifies depth deviations in planar forgeries or 3D masks, thus resisting three-dimensional forgery attacks such as 3D printed masks and silicone headgear. Micro-expression detection uses temporal convolutional networks to analyze continuous frames of facial images, capturing micro-expression features such as blinking and muscle micro-movements. It sets micro-expression frequency thresholds to distinguish between natural human movements and static features of forgeries, identifying attacks without dynamic features such as video re-enactments and static images. Dynamic action verification randomly issues interactive commands, detects the completion rate of user actions using keypoint tracking algorithms, and combines this with action temporal consistency judgment to prevent attackers from using pre-recorded videos for deception, thus countering highly realistic video re-enactment attacks.

[0042] In practice, 3D mask attack detection includes: material difference detection, where the surface roughness of silicone and resin masks differs from that of real skin, which can be identified by analyzing the frequency of surface normal changes in 3D point clouds; depth anomaly detection, where facial features of masks (such as the width of the nose and the height of the cheekbones) deviate slightly from those of real people, and the depth error of key points between the reconstructed model and the real person's 3D template can be used to determine if it is a forgery; and dynamic deformation detection, where real people's facial muscles undergo slight deformations when they move (such as speaking or blinking), while masks do not have this dynamic change, which can be further verified by deformation analysis of continuous frame 3D models.

[0043] The 3D reconstruction is lightweight and optimized with real-time performance guaranteed. Only the face region is reconstructed in 3D, instead of the entire image, reducing the amount of computation by more than 60%. CUDA parallel computing is used to optimize point cloud registration and mesh reconstruction, reducing the reconstruction time from 500ms to 150ms. Key geometric features of the 3D model are extracted, such as the depth, curvature, and normal vectors of 100 key points, instead of the complete model, for anti-spoofing verification, reducing the amount of data to 1% of the original.

[0044] S3, triggered by daily access, the system automatically activates backlight compensation and noise filtering by monitoring light and noise levels in real time through environmental sensors.

[0045] In this embodiment, an image preprocessing and anti-interference module is included. The core technologies include adaptive filtering, backlight compensation, pose correction, blur repair and noise filtering algorithms. The module is used to optimize the acquired raw image, eliminate the effects of abnormal lighting, image blur and background interference, and output high-quality face images to reduce interference pressure on the recognition algorithm.

[0046] In this embodiment, the anti-interference methods of the image preprocessing and anti-interference module include: image preprocessing anti-interference, which eliminates interference noise in single-modal images through basic image processing techniques, removes noise through adaptive filtering, solves uneven lighting through backlight compensation, improves motion and out-of-focus blur through blur repair, and corrects facial angle deviations through pose correction, thereby improving image quality from the bottom layer; multimodal fusion anti-interference, which fuses three types of modal information: RGB image, infrared thermal imaging, and 3D depth data. The RGB channel preserves facial texture details, the infrared channel distinguishes the thermal radiation characteristics of live objects from fakes, and the 3D channel provides facial contours. A weight allocation algorithm is used to achieve multimodal feature complementarity; and adaptive algorithm anti-interference, based on ring... The system uses real-time sensors to collect parameters such as light intensity and noise level, dynamically adjusting algorithm strategies and parameter thresholds. In strong light environments, it automatically enhances backlight compensation intensity; in high-noise scenes, it switches to a strong filtering mode; and for blurred images, it triggers a super-resolution reconstruction process. Through a preset interference type identification model, it automatically matches the optimal processing algorithm to achieve intelligent anti-interference through environmental perception and algorithm adaptation. The collaborative anti-interference mechanism works in conjunction with the front-end multimodal perception and acquisition module to receive environmental interference prediction information and initiate targeted processing in advance. It also interacts with the back-end face recognition and liveness verification module to backtrack and optimize preprocessing parameters based on recognition feedback results. Multiple modules work together to form a closed-loop anti-interference system of acquisition, processing, and recognition.

[0047] In this embodiment, the workflow of the image preprocessing and anti-interference module is as follows: It receives raw multimodal data, including encapsulated data from the perception module via a high-speed data interface, such as RGB images, infrared thermal imaging images, 3D depth point cloud data, and environmental sensor parameters; it initializes data preprocessing by loading pre-trained face detection models and algorithm configuration parameters, initializing memory buffers, storing data according to data type, and dynamically allocating thread resources based on device computing power; it analyzes environmental parameters by parsing illumination sensor data, classifying illumination levels according to thresholds, calculating noise levels through image variance, and generating environmental interference labels; it performs channel-specific processing by splitting multimodal data into RGB and infrared channels for parallel processing, using an 8-bit color space for the RGB channels and converting the infrared channels to grayscale images, and normalizing the dual-channel data to eliminate hardware acquisition differences; and it determines the type of interference, using backlight compensation for illumination interference. For color temperature calibration, blur interference is addressed through blur repair and sharpening enhancement; pose interference is addressed through key point detection and pose correction; and noise interference is addressed through adaptive filtering and noise suppression. Multi-channel data fusion employs a weighted fusion algorithm, preserving color texture in the RGB channel and enhancing facial contours in the infrared channel. Image registration aligns the coordinates of the two channels, and the fused data is converted to the YCbCr color space, optimizing the contrast of the brightness component. Background interference removal involves locating the face region using a face detection bounding box, expanding it by 10% as the region of interest (ROI), segmenting the background using the GrabCut algorithm, removing interfering elements such as posters and clutter, and cropping the ROI to a fixed size to meet the input requirements of the recognition algorithm. High-quality face images are output by encapsulating the processed images in a standard format, adding data tags, and transmitting them to the face recognition module via shared memory or a message queue. Simultaneously, the most recent 10 frames are cached for anomaly backtracking.

[0048] It should be noted that adaptive filtering combines bilateral filtering and guided filtering to eliminate noise while preserving facial edge details. The filter kernel size is dynamically adjusted based on the local variance of the image, increasing the kernel size in noisy areas and decreasing it in edge areas to avoid blurring facial features. Backlight compensation uses the Retinex enhancement algorithm to address the problem of overly dark faces under backlight and strong light. The image is decomposed into illumination and reflection components, and a logarithmic transformation is performed on the illumination component to compress its dynamic range, enhancing facial details in dark areas. The compensation intensity is adaptively adjusted based on illumination data from environmental sensors. Pose correction is based on affine transformations of key points to correct facial tilt or profile angles, detecting 68 facial landmarks. Key points (corners of the eyes, tip of the nose, corners of the mouth, etc.) are used to calculate the facial deflection angle. Affine transformation is used to correct the tilted face to a frontal view, ensuring the consistency of feature extraction. Blur restoration employs wavelet transform and super-resolution reconstruction to improve motion blur and out-of-focus blur. Wavelet decomposition is performed on the blurred image to extract and enhance high-frequency detail components. Combined with the facial prior model and facial structure features, clear facial texture is reconstructed. Noise filtering is based on adaptive median filtering and Gaussian noise suppression to eliminate environmental electromagnetic interference and sensor noise. Salt-and-pepper noise and Gaussian noise types are distinguished, and targeted filtering algorithms are selected. Noise filtering is strengthened in the background area, while details are preserved in the facial area to avoid over-processing.

[0049] S4. Multimodal anti-interference recognition: The system performs posture correction and blur repair on the acquired images, and integrates 3D depth and infrared thermal radiation features to resist photo and mask forgery attacks.

[0050] In this embodiment, a collaborative control and decision-making module is included. The core components include an edge processor, an access controller, and an alarm device. The module is used to receive the face recognition results, combine them with two-factor verification to decide whether to allow access, trigger an audible and visual alarm when an abnormal attack is detected, and simultaneously drive the electromagnetic lock to complete the switch control, thereby realizing a closed loop of recognition, decision-making, and execution.

[0051] It should be noted that the edge processor is the core computing unit of the module, running decision algorithms, storing the local face database, and processing data from multiple module interactions; the access controller receives control commands, drives the electromagnetic lock switch, integrates a Wiegand / RS485 communication interface, and supports two-factor authentication; the alarm device emits a 110dB audible and visual warning in case of an abnormal attack, and supports local and remote triggering; the local database caches authorized face feature data, access logs, and abnormal event records; the communication module enables data interaction with the remote management platform and supports breakpoint resume.

[0052] In this embodiment, the collaborative strategy of the collaborative control and decision-making module is as follows: receiving the feature matching score, liveness verification result, and attack type label from the face recognition module; issuing lock / unlock commands and alarm trigger commands to the access control execution module; uploading access records and abnormal events to the remote management and iteration module in real time; supporting combined verification of face and password, and face and card swiping, automatically triggered in medium-risk scenarios; automatically switching to local mode when the network connection is lost, relying on the local database to complete recognition and decision-making, and automatically synchronizing data after reconnection.

[0053] The collaborative control and decision-making module adopts a weighted scoring decision model to comprehensively calculate the safety score. The calculation formula is as follows: Face matching score, weight ; Weighting for liveness verification score ; Score historical behavior, weighted .

[0054] S5. Result feedback and execution: If the recognition is successful, the access control will be unlocked; if it fails, the system will prompt the user to adjust the position and remove the obstruction. Abnormal attacks will trigger an audible and visual alarm.

[0055] In this embodiment, a remote management and iteration module is included. The core components include a management platform, a local database, and a network transmission component. The module is used to remotely configure user permissions, view access logs, and monitor device status. It continuously improves the system's anti-interference capability by incrementally learning and optimizing the algorithm model, and supports the retrieval of locally cached data in the event of a network outage.

[0056] In this embodiment, an access control execution module is included, which receives instructions from the collaborative control and decision-making module and drives the electromagnetic lock to complete the unlocking and locking actions to enable authorized passage or prohibit passage.

[0057] In this embodiment, the algorithm iteration and optimization mechanism in the remote management and iteration module includes: automatically uploading sample data of failed identification and suspected attacks, along with environmental parameters, from the edge; using the FedAvg federated learning framework to aggregate abnormal data from multiple devices in the cloud for model fine-tuning, avoiding full retraining; generating specialized optimization models for scenarios with weak anti-interference capabilities; and quantizing and compressing the optimized model to generate an update package of less than 10MB, which is then pushed to the edge via differential update technology. The edge automatically updates the model during idle periods, and it can only be enabled after being verified through a test set.

[0058] It should be noted that the management platform adopts a visual management interface, supporting permission configuration, log viewing, and device monitoring; the cloud database stores all user data, access logs, and algorithm model parameters; the local database caches frequently used user data and local access logs, and runs independently when the network is disconnected; the network transmission component enables low-latency data interaction between the cloud and the edge, and supports breakpoint resume; the algorithm iteration engine optimizes the model based on the newly added data, generates lightweight update packages, and pushes them to the edge.

[0059] In practice, administrators use the platform to grant tiered permissions to users, such as allowing ordinary employees access only on weekdays and administrators access 24 / 7. The platform supports batch import and export of user data. Permission change commands are pushed to the edge devices in real time via the MQTT protocol, and the local database is automatically updated synchronously. Real-time data collection includes edge device CPU / memory usage, camera status, network connectivity, and access control opening / closing records. Threshold alarms are set, such as when a device is offline for more than 5 minutes or storage usage exceeds 80%, alarm information is pushed via SMS and the app. The entire access log is stored in the cloud, including personnel ID, time, access result, and anomaly type, and can be filtered and queried by time, personnel, and device. Log data is automatically backed up to cold storage, with a configurable retention period.

[0060] It should be noted that the system security mechanism design includes: data encryption, user facial features are stored using AES-256 encryption, and the transmission process uses the TLS1.3 protocol; access control, hierarchical authorization of administrators, such as super administrators, device administrators, and viewers, and full auditing of operation logs; and model security, with digital signatures added to the algorithm model, and the signature is verified at the edge before loading and updating to prevent malicious model injection.

[0061] Example 2: To better understand the above technical solution, the following will provide a detailed explanation of the specific implementation method using actual cases. To provide efficient and accurate solutions for various practical application scenarios, a collaborative anti-interference access control system based on face recognition is proposed. This collaborative anti-interference access control system includes: The multimodal perception and acquisition module's core hardware includes an RGB infrared spectral camera, supplementary lighting equipment, environmental sensors, and distance sensors. The module is used to acquire 2D texture, 3D depth, and infrared thermal radiation data of human faces, and simultaneously perceive environmental interference parameters to provide raw data support for anti-interference processing.

[0062] The image preprocessing and anti-interference module incorporates core technologies such as adaptive filtering, backlight compensation, pose correction, blur restoration, and noise filtering algorithms. This module optimizes the acquired raw images, eliminates the effects of abnormal lighting, image blur, and background interference, and outputs high-quality face images, reducing interference pressure on the recognition algorithm.

[0063] The face recognition and liveness verification module incorporates core technologies including multimodal fusion algorithms, ArcFace-Lite feature extraction algorithms, 3D reconstruction, micro-expression detection, and dynamic action verification. The module extracts core facial features and uses 2D, 3D, and infrared multi-dimensional verification to accurately distinguish between real people and images or masks, ensuring the accuracy of identity verification.

[0064] The collaborative control and decision-making module, whose core components include an edge processor, an access controller, and an alarm device, receives facial recognition results, combines them with two-factor authentication to decide whether to allow access, triggers an audible and visual alarm when an abnormal attack is detected, and simultaneously drives an electromagnetic lock to complete the switching control, thus realizing a closed loop of recognition, decision-making, and execution.

[0065] The remote management and iteration module consists of a management platform, a local database, and network transmission components. This module is used to remotely configure user permissions, view access logs, and monitor device status. It continuously improves the system's anti-interference capabilities by incrementally learning and optimizing the algorithm model, and supports the retrieval of locally cached data in the event of a network outage.

[0066] The access control execution module receives instructions from the collaborative control and decision-making module, drives the electromagnetic lock to complete the unlocking and locking actions, and realizes authorized access or prohibition of access.

[0067] System deployment and hardware installation: Determine the installation location of the access control system based on actual needs, such as building entrances and exits, doors of important rooms, etc. Choose areas with sufficient light, normal personnel flow and no obvious obstructions to ensure the facial recognition effect. The software configuration and database setup assign appropriate access control permissions to different users or user groups based on their identities and needs. For example, company employees can be set to freely enter and exit the office area during working hours on weekdays; visitors can be set to only access specific reception areas during designated time periods; and administrators have higher permissions, allowing them to enter and exit various areas at any time. Simultaneously, different users' operational permissions for the access control system can be set, such as whether they are allowed to modify system settings or view records. The system associates and binds the entered user facial feature data with the corresponding user identity information to ensure that the system can accurately identify and match each user. After data entry and permission allocation are completed, the data undergoes comprehensive verification to check for issues such as duplicate entries, incomplete information, and unreasonable permission allocation, and erroneous data is corrected promptly.

[0068] During user registration and facial information entry, users stand in front of the camera and complete a series of actions as prompted by the system, such as blinking, slowly turning their heads, and slightly raising and lowering their heads. These actions aim to collect facial features from multiple angles and in all directions to improve the accuracy and reliability of recognition. The system automatically extracts and saves the user's facial feature data and associates it with the identity information entered by the user in step two. When storing, the facial feature data is stored in the system database in an encrypted form to ensure data security and privacy.

[0069] In routine access operations, as a user walks normally towards the access control area, the system automatically activates when the user enters the effective recognition range of the facial recognition camera. The camera begins to capture the user's facial image quickly and clearly. The system immediately processes the captured image, extracts facial features, and compares and analyzes these features with user facial feature data pre-stored in the database. The comparison process employs fast and accurate algorithms to ensure accurate results are obtained in a short time.

[0070] If the system determines that the facial features of the collected face match the facial features of a user in the database to a preset threshold, and the user has the current access control permission, the access control device will immediately issue an instruction to automatically open the access control, such as unlocking the electric lock or opening the gate. At the same time, the system can inform the user of successful access through voice prompts, screen displays, or flashing lights.

[0071] If the comparison result shows a mismatch or the matching degree is lower than the threshold, the system will take corresponding actions according to the pre-set rules. For example, it will inform the user of the recognition failure through voice prompts and screen prompts. At the same time, the system will automatically record the recognition attempt, including the time, location, and captured image information, for subsequent query and analysis. If the recognition fails multiple times in a row, the system can also take further security measures, such as triggering an alarm to notify security personnel, automatically capturing on-site photos and uploading them to the management center.

[0072] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A collaborative anti-interference access control system based on face recognition, characterized in that, The cooperative anti-interference access control system comprises: A multi-modal perception acquisition module, the core hardware includes an RGB infrared spectrum camera, a light supplementing device, an environment sensor and a distance sensor, the module is used for collecting face 2D texture, 3D depth and infrared thermal radiation data, synchronously perceiving environment interference parameters, and providing original data support for anti-interference processing; An image preprocessing and anti-interference module, the core technology includes adaptive filtering, back light compensation, posture correction, blur repair and noise filtering algorithm, the module is used for optimizing the collected original image, eliminating the influence of light abnormality, image blur and background interference, and outputting a high-quality face image, thereby reducing the interference pressure for the recognition algorithm; A face recognition and living body verification module, the core technology includes a multi-modal fusion algorithm, an ArcFace-Lite feature extraction algorithm, 3D reconstruction, micro-expression detection and dynamic action verification, the module is used for extracting face core features, accurately distinguishing between a real person and a picture or a mask through 2D, 3D and infrared multi-dimensional verification, and ensuring the accuracy of identity verification; A cooperative control and decision module, the core components include an edge processor, an access controller and an alarm device, the module is used for receiving face recognition results, deciding whether to release in combination with double-factor verification, triggering sound and light alarm when detecting abnormal attack, and driving an electromagnetic lock to complete switching control, so that recognition, decision and execution are closed loops; A remote management and iteration module, the core components include a management platform, a local database and a network transmission component, the module is used for remotely configuring user permissions, checking access logs and monitoring equipment states, optimizing algorithm models through incremental learning, continuously improving the anti-interference capability of the system, and supporting local cache data calling in a network interruption state; An access control execution module, used for receiving instructions of the cooperative control and decision module, driving an electromagnetic lock to complete unlocking and locking actions, and realizing authorized access or prohibited access.

2. The collaborative anti-interference access control system based on face recognition according to claim 1, characterized in that: The environment sensor includes an illumination sensor and a temperature and humidity sensor, which are used for collecting illumination intensity and environment temperature and humidity in real time, and providing environment parameters for anti-interference algorithms; When the illumination sensor detects a strong light environment, the camera HDR mode is automatically started, and the light supplementing lamp brightness is reduced; when the illumination sensor detects a dark light environment, infrared light supplementing is triggered, and the infrared imaging mode is switched to, and the image is adaptively collected in the environment.

3. The collaborative anti-interference access control system based on face recognition according to claim 2, characterized in that: The distance sensor detects the distance between the face and the equipment, filters invalid collection of a target being too close or too far, and avoids picture copying attack; When the distance sensor identifies that the distance of the collection target exceeds the normal collection range, the recognition is immediately paused and the user is prompted to adjust the distance; when the temperature and humidity sensor monitors an extreme environment, a hardware protection mechanism is started, and the stability of core data collection is preferentially ensured.

4. The collaborative anti-interference access control system based on face recognition according to claim 1, characterized in that: The anti-interference mode of the image preprocessing and anti-interference module comprises: Image preprocessing anti-interference, interference noise in a single modal image is eliminated through basic image processing technology, salt and pepper and Gaussian noise are removed through adaptive filtering, uneven light is solved through back light compensation, motion and defocus blur are improved through blur repair, and face angle deviation is corrected through posture correction, so that the quality is improved from the bottom of the image; Multi-modal fusion anti-interference, fusing RGB image, infrared thermal imaging, and 3D depth data three types of modal information, RGB channel retains face texture details, infrared channel distinguishes the thermal radiation characteristics of living beings and counterfeit objects, 3D channel provides face stereo profile, and a weight distribution algorithm is used to realize the complementary of multi-modal features; Adaptive algorithm anti-interference, based on the parameters such as light intensity and noise level collected by the environment sensor in real time, dynamically adjusting the algorithm strategy and parameter threshold, automatically enhancing the back light compensation intensity in strong light environment, switching to strong filtering mode in high noise scene, triggering super-resolution reconstruction process for blurred images, automatically matching the optimal processing algorithm through the preset interference type recognition model, realizing the intelligent anti-interference of environment perception and algorithm adaptation; Collaborative anti-interference mechanism, linked with the front-end multi-modal perception and collection module, receiving environment interference prediction information, starting targeted processing in advance, interacting with the back-end face recognition and living body verification module, optimizing preprocessing parameters according to the recognition feedback results, multiple modules working collaboratively, forming a closed-loop anti-interference system of collection-processing-identification.

5. The collaborative anti-interference access control system based on face recognition according to claim 4, characterized in that: The image preprocessing and anti-interference module workflow is as follows: Receiving raw multi-modal data, receiving encapsulated data from the perception module through a high-speed data interface, including RGB image, infrared thermal imaging image, 3D depth point cloud data and environment sensor parameters; Data preprocessing initialization, loading pre-trained face detection model and algorithm configuration parameters, initializing memory buffer, storing according to data type, dynamically allocating thread resources according to device computing power; Environment parameter analysis, analyzing light sensor data, dividing light level according to threshold, calculating noise level through image variance, generating environment interference label; Channel processing, multi-modal data is split into RGB channel and infrared channel for parallel processing, RGB channel uses 8-bit color space, infrared channel is converted to grayscale image, and double-channel data is normalized to eliminate hardware collection differences; Judging interference type, light interference adopts back light compensation and color temperature calibration, fuzzy interference adopts fuzzy repair and sharpening enhancement, posture interference adopts key point detection and posture correction, and noise interference adopts adaptive filtering and noise suppression; Multi-channel data fusion, using weight fusion algorithm, RGB channel retains color texture, infrared channel enhances face contour, aligning double-channel image coordinates through image registration technology, converting the fused image to YCbCr color space to optimize brightness component contrast; Background interference elimination, positioning face area through face detection frame, expanding 10% as region of interest, segmenting background through GrabCut algorithm, eliminating poster, debris and other interference elements, cropping ROI area to fixed size to meet the input requirements of recognition algorithm; Output high-quality face image, packaging the processed image according to the standard format, adding data label, transmitting to face recognition module through shared memory or message queue, and buffering the last 10 frames of data for abnormal backtracking.

6. The collaborative anti-interference access control system based on face recognition according to claim 1, characterized in that: The multi-modal fusion algorithm adopts a three-level progressive fusion architecture to realize deep fusion of multi-modal information from data layer, feature layer to decision layer; Data layer fusion, pixel-level alignment of RGB images, infrared thermal imaging images, and 3D depth maps, calibration of the spatial coordinates of the three types of data through SIFT feature matching technology to ensure that the features of the same face region correspond in position in different modalities; Feature layer fusion, design of a multi-branch convolutional neural network, extraction of texture features by the RGB branch, extraction of thermal radiation features by the infrared branch, and extraction of depth geometric features by the 3D branch, automatic allocation of feature weights of each branch through an attention mechanism; Decision layer fusion, adoption of a weighted voting mechanism, RGB feature matching score weight 0.4, infrared feature matching score weight 0.3, and 3D feature matching score weight 0.3, and comprehensive calculation of the final matching degree.

7. The collaborative anti-interference access control system based on face recognition according to claim 1, characterized in that: In the ArcFace-Lite feature extraction algorithm, the input image is first detected for face key points by MTCNN, the face is aligned to a standard template through a similarity transformation to eliminate the influence of posture and position offset on feature extraction, face cropping and normalization are performed after alignment to improve feature stability; Data augmentation is added in the training phase to enable the model to learn robustness to interference, and in the inference phase, a blurred image is first subjected to super-resolution reconstruction and then input into the feature extraction network to avoid feature distortion caused by blurring; The model weights are compressed from 32-bit floating point to 8-bit integer through INT8 quantization, and layer fusion is used to reduce memory access times and improve the running efficiency of edge devices.

8. The collaborative anti-interference access control system based on face recognition according to claim 1, characterized in that: The 3D reconstruction method of the 3D reconstruction comprises: Point cloud generation, conversion of the depth map into a three-dimensional point cloud containing X, Y, and Z coordinates of each point; Point cloud registration, registration of multiple frames of point clouds into a complete face point cloud through an ICP algorithm; Mesh reconstruction, conversion of the point cloud into a three-dimensional mesh model using a Poisson reconstruction algorithm to restore the three-dimensional contour of the face; Texture mapping, mapping of the texture information of the RGB image to the 3D mesh to generate a 3D face model with texture.

9. The collaborative anti-interference access control system based on face recognition according to claim 1, characterized in that: The collaborative strategy of the collaborative control and decision module comprises: Receiving the feature matching score, liveness verification result, and attack type label from the face recognition module; Issuing a lock opening instruction and an alarm triggering instruction to the access control execution module; Real-time uploading of access records and abnormal events to the remote management and iteration module; Supporting combined verification of face and password, face and card swiping, automatic triggering in medium-risk scenarios; Automatic switching to local mode when the network is disconnected, relying on the local database to complete identification and decision-making, and automatically synchronizing data after networking.

10. The collaborative anti-interference access control system based on face recognition according to claim 1, characterized in that: The algorithm iteration optimization mechanism in the remote management and iteration module comprises: Automatic uploading of sample data of failed recognition and suspected attacks by the edge, with environmental parameters; Using the FedAvg federated learning framework to aggregate abnormal data from multiple devices on the cloud to fine-tune the model, avoiding full retraining; generating a special optimization model for anti-interference weak scenarios; The optimized model is quantized and compressed to generate an update package smaller than 10MB, which is pushed to the edge through differential update technology, and the edge automatically updates the model during idle periods. The updated model must be verified by a test set before it can be enabled.

Images