Face recognition method and system based on cloud machine cooperation

Abstract

The invention discloses a face recognition method and system based on cloud machine collaboration, and the method comprises the steps: obtaining data: obtaining to-be-recognized face image data; co-processing: the device end receives preset face feature library update data and algorithm optimization parameters fed back by the cloud server, and the device end and the cloud server execute data synchronization; and data processing: the equipment end performs preprocessing and feature extraction on the face image data, generates a face feature data set with feature identification, and performs comparison and living body verification through a multi-dimensional feature matching model based on the face feature data set and in combination with the preset face feature library update data and a local feature library. Generating a face recognition report containing a matching result and a living body judgment result; and executing operation, analyzing the face recognition report, and performing recognition judgment. The processing requirements of high-data and high-complexity scenes are met, and the problems that existing equipment is poor in flexibility, insufficient in performance, limited in battery life and the like are solved.

Description

Technical Field

[0001] This invention relates to the field of facial recognition technology, and in particular to a facial recognition method and system based on cloud-computer collaboration. Background Technology

[0002] With the rapid integration of artificial intelligence and Internet of Things technologies, facial recognition, as a core branch of biometric technology, has expanded from fixed-scene applications to mobile and portable scenarios.

[0003] In existing technologies, facial recognition devices mostly adopt a local independent processing mode, which has three key problems: First, local computing power and storage resources are limited. When faced with special scenarios such as multi-angle and complex lighting, incomplete feature extraction and low matching accuracy are prone to occur. Especially on portable devices, the battery life is limited, making it difficult to maintain a high-performance processing state for a long time. Second, there is a lack of machine-cloud collaboration mechanism. Local feature library updates rely on manual operation and cannot synchronize newly added or updated identity information in the cloud in real time, resulting in data gaps when recognizing across regions and devices. Third, the ability to handle anomalies is weak. When recognition fails (such as feature mismatch or abnormal liveness detection), only simple logs can be recorded locally. It is impossible to conduct large-scale abnormal data analysis and algorithm optimization iteration through the cloud, making it difficult to adapt to the needs of dynamically changing usage scenarios.

[0004] Especially in scenarios such as mobile asset supervision verification and temporary unlocking authorization management, how to improve the accuracy, real-time performance and data collaboration of facial recognition while ensuring device portability and battery life has become a major bottleneck in current technological development. Summary of the Invention

[0005] To address one or more problems existing in the prior art, this invention provides a face recognition method and system based on cloud-computer collaboration. The technical solution adopted by this invention to solve the above problems is: a face recognition method based on cloud-computer collaboration, the method comprising: data acquisition, acquiring face image data to be recognized; Collaborative processing: Through the cloud-based collaborative processing unit, the device uploads part of the facial image data or its preprocessing requirements to the cloud server. The device receives the preset facial feature library update data and algorithm optimization parameters fed back by the cloud server. The device and the cloud server perform data synchronization. The model is processed collaboratively. The device is equipped with a lightweight scene classification and recognition AI model and a face enhancement AI model, while the cloud server is equipped with a deep scene classification and recognition AI model and a face enhancement AI model. The cloud-machine collaborative processing unit distributes scene classification and recognition tasks and face enhancement tasks to the device and the cloud server. The processing results of the cloud server are fed back to the device. The processing results of the scene classification and recognition task are used as input to the face enhancement task. The processing results of the face enhancement task are integrated into the face image data. Data processing: The device performs preprocessing and feature extraction on the face image data to generate a face feature dataset with feature identifiers. Based on the face feature dataset, combined with the updated data of the preset face feature library and the local feature library, a multi-dimensional feature matching model is used to perform real-time comparison and liveness verification to generate a face recognition report containing matching results and liveness determination results. The system performs the operation, parses the face recognition report and makes a recognition judgment. If the recognition is successful, a preset authorization operation is triggered. If the recognition fails, the current failure data is uploaded to the cloud server and anomaly analysis, prompt signal output and recognition failure log update are performed.

[0006] In some embodiments, the facial image data acquired by the data acquisition includes: facial images captured from different angles, facial images captured under different lighting conditions, and dynamic facial images; The steps for acquiring the face image data are as follows: Step 1, the multi-module image acquisition unit synchronously acquires the face image of the object to be identified. The multi-module image acquisition unit includes a wide-angle camera, a supplementary light camera and an infrared camera. The image data acquired by the multi-module image acquisition unit is recorded as the initial face image set. Step 2: Perform format conversion and resolution unification on the initial face image set, divide the images into valid images, blurred images, and invalid images according to the screening criteria, retain the valid images, and generate a screened face image set; Step 3: The cloud-machine collaborative unit determines the data volume and complexity of the selected face image set, and then performs a threshold comparison of the data volume and complexity. Based on the threshold comparison result, the cloud-processed data and the local-processed data are divided. The cloud server performs preliminary screening, classification and labeling on the cloud-processed data and feeds back the cloud processing results to the device. The device performs classification and labeling on the local-processed data and generates local processing results. The image data of both the cloud-processed results and the local-processed results contain classification labels. The device integrates the cloud-processed results and the local-processed results to generate the face image data.

[0007] In some embodiments, the face feature dataset generation step of the data processing includes: Step 1, reading a single face image from the face image data, and sequentially performing image noise reduction and grayscale correction processing to generate a first processed image; Step 2: Dynamically allocate face region cropping and edge enhancement processing tasks for the first processed image through the cloud-machine collaboration unit. If the device resources are lower than a preset threshold, the device resources include computing power and power. Then, the face region cropping and edge enhancement processing tasks are allocated to the cloud server for processing and generating a second processed image. The device receives the second processed image. If the device resources are higher than or equal to the preset threshold, the face region cropping and edge enhancement processing tasks are allocated to the device for processing and generating a third processed image. Step 3: Load the algorithm optimization parameters into the deep learning-based feature extraction algorithm, and use the feature extraction algorithm to extract facial key feature points from the second processed image and the third processed image to obtain an initial feature point set. The facial key feature points include: corners of the eyes, tip of the nose, corners of the mouth, and contour features. Step 4: Perform feature quantization and dimensionality reduction on the initial feature point set, remove redundant feature information and retain core feature parameters, add a unique feature identifier to each core feature parameter, and then generate and back up the face feature dataset.

[0008] In some embodiments, the generation of the face recognition report by the data processing includes: During the execution of steps 1-3, the device performs local liveness detection. Through dynamic facial motion capture and infrared feature recognition, it determines whether the identified object is a real live body and generates a first liveness determination result. If the local liveness detection is abnormal, the cloud server obtains the abnormal image and detection data and performs a second determination. The cloud server feeds back the second liveness determination result of the second determination to the device. Step 1: Extract feature identifiers and core feature parameters from the face feature dataset, and then generate feature analysis data; Step 2: The local feature library on the device is synchronized with the cloud feature library on the cloud server; Step 3: The multi-dimensional feature matching model on the device performs preliminary feature similarity calculation based on the feature analysis data and the local feature library to obtain preliminary feature similarity. The multi-dimensional feature matching model includes: a local feature matching algorithm and a global feature matching algorithm. If the preliminary feature similarity is within a preset fuzzy range, the cloud server performs a secondary comparison and feeds back the feature similarity score to the device. The accuracy of the secondary comparison is higher than that of the preliminary feature similarity calculation. If the preliminary feature similarity is within a preset definite range, the feature similarity score is generated. Step 4: Integrate the feature similarity score, the first liveness determination result, and the second liveness determination result. If the first liveness determination result and the second liveness determination result are both determined to be real and the feature similarity score is higher than a preset threshold, then the current recognition is determined to be successful; otherwise, the current recognition is determined to be unsuccessful. Generate a face recognition report and synchronize it to the cloud server.

[0009] In some embodiments, the parsing of the face recognition report in the execution operation includes: parsing the judgment result and key data in the face recognition report, generating a feedback instruction, the feedback instruction being used to drive the display to output visual feedback information, the visual feedback information including: recognition progress, matching percentage and judgment result, the visual feedback information being synchronized to the cloud server; The preset authorization operations include: authorization verification, data access authorization, unlocking, and log recording, wherein the log recording is synchronized to the cloud server; The anomaly analysis includes: feature library missing analysis and image acquisition anomaly analysis. The cloud server provides optimization suggestions to the device based on the results of the anomaly analysis.

[0010] In some embodiments, the training data for the lightweight scene classification and recognition AI model and the face enhancement AI model in the model co-processing are low-complexity data, which are photos of a single living person under different backgrounds and facial photos of a single living person with different resolutions, respectively. The photos of a single living person under different backgrounds are divided into multiple groups according to pixel range and grayscale range and input separately for training. The facial photos of a single living person with different resolutions are divided into multiple groups according to pixel range, color range, and age range and input separately for training. The individual live facial photos of different resolutions are photos with scene annotations output by the lightweight scene classification and recognition AI model; The outputs of the lightweight scene classification and recognition AI model and the face enhancement AI model are the lightweight background result of the live object and the lightweight enhanced face image, respectively. The training data for the deep scene classification and recognition AI model and the face enhancement AI model that are processed by the model are medium to high complexity data, which are photos of multiple live objects under different backgrounds and facial photos of multiple live objects with different clarity. The photos of multiple live objects under different backgrounds are divided into multiple groups according to pixel range, grayscale range, and number of live objects and are input separately for training. The facial photos of multiple live objects with different clarity are divided into multiple groups according to pixel range, color range, age range, and number of faces and are input separately for training. The multiple live facial photos of different resolutions are photos with scene annotations output by the deep scene classification and recognition AI model; The low-complexity data and the medium-to-high-complexity data are divided by manual annotation or AI annotation tools; The outputs of the deep scene classification and recognition AI model and the face enhancement AI model are respectively the deep background result of the living object and the deep enhanced face image; The lightweight background result of the live subject and the deep background result of the live subject are used as the processing results of the scene separation and recognition task; Both the lightweight facial enhancement AI model and the deep facial enhancement AI model are equipped with multiple scene processing modes. The lightweight facial enhancement AI model selects the corresponding scene processing mode based on the background result of the lightweight live subject. The deep facial enhancement AI model selects the corresponding scene processing mode based on the background result of the deep facial enhancement.

[0011] Furthermore, in the model collaborative processing, the method of allocating scene classification and recognition tasks and face enhancement tasks to the device and the cloud server through the cloud machine collaborative processing unit includes: the device analyzes and recognizes the initial face image set to obtain the number of live subjects and the number of faces, and compares them with a preset threshold. If the values ​​are within the threshold range, the device performs local processing; if the values ​​are not within the threshold range, the cloud server performs cloud processing.

[0012] In some embodiments, in the model collaborative processing, the method of allocating scene classification and recognition tasks and face enhancement tasks to the device and cloud server through the cloud machine collaborative processing unit includes: allocating tasks according to the amount and complexity of data being processed in the current task.

[0013] And a system applying the above method, comprising: a data acquisition unit, the data acquisition module being used to acquire face image data to be recognized; a cloud-machine collaboration unit, the cloud-machine collaboration unit being used to perform collaborative processing and model collaborative processing; a data processing unit, the data processing unit being used to generate a face recognition report; and an execution unit, the execution unit being used to parse the face recognition report and perform recognition determination, and execute pass and fail operations of the recognition determination.

[0014] Furthermore, the data acquisition unit is equipped with a wide-angle camera, a fill light camera, and an infrared camera, and the data acquisition unit is equipped with a fill light mode, an extreme black light mode, and an AOV mode.

[0015] The technical effects achieved by this invention are as follows: Through the above-mentioned methods and systems, data synchronization and task allocation between the device and the cloud server are realized. The device performs local processing to ensure the real-time performance of basic recognition, while the cloud server utilizes its high computing power to perform collaborative processing, meeting the processing needs of high-data and high-complexity scenarios. The linkage mechanism between the two effectively solves the problems of poor flexibility, insufficient performance, and limited battery life of existing devices. The combination of pre-set facial feature library update data and local feature library avoids the problem of data gaps. The bidirectional data synchronization, data analysis optimization, and optimization feedback between the device and the cloud server enable the entire system to move from deployment to mature use more quickly. Ultimately, while ensuring the portability and battery life of the device, the timeliness, accuracy, and multi-scenario adaptability of facial recognition are improved. Attached Figure Description

[0016] Figure 1 This is a schematic block diagram of the method of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features, and advantages of the present invention more readily understood, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0018] like Figure 1 As shown, the present invention discloses a face recognition method based on cloud-machine collaboration, the method comprising: data acquisition, acquiring face image data to be recognized; Collaborative processing: Through the cloud-based collaborative processing unit, the device uploads part of the facial image data or its preprocessing requirements to the cloud server. The device receives the preset facial feature library update data and algorithm optimization parameters fed back by the cloud server. The device and the cloud server perform data synchronization. Data processing: The device performs preprocessing and feature extraction on the face image data to generate a face feature dataset with feature identifiers. Based on the face feature dataset, combined with the updated data of the preset face feature library and the local feature library, a multi-dimensional feature matching model is used to perform real-time comparison and liveness verification to generate a face recognition report containing matching results and liveness determination results. The model is processed collaboratively. The device is equipped with a lightweight scene classification and recognition AI model and a face enhancement AI model, while the cloud server is equipped with a deep scene classification and recognition AI model and a face enhancement AI model. The cloud-machine collaborative processing unit distributes scene classification and recognition tasks and face enhancement tasks to the device and the cloud server. The processing results of the cloud server are fed back to the device. The processing results of the scene classification and recognition task are used as input to the face enhancement task. The processing results of the face enhancement task are integrated into the face image data. The system performs the operation, parses the face recognition report and makes a recognition judgment. If the recognition is successful, a preset authorization operation is triggered. If the recognition fails, the current failure data is uploaded to the cloud server and anomaly analysis, prompt signal output and recognition failure log update are performed.

[0019] The scene classification and recognition AI model, through collaborative processing, enables rapid scene recognition on both the device and cloud sides, providing processing guidance for subsequent facial enhancement AI models, thereby accelerating processing speed and accuracy. The facial enhancement AI model, also through collaborative processing, enables different levels of image enhancement precision on both the device and cloud sides, and assists in subsequent face recognition, ensuring high accuracy. By separating scene recognition and facial enhancement into two large models, model complexity is reduced, thus lowering hardware requirements on the device. Simultaneously, the processing efficiency and accuracy of scene recognition and facial enhancement are specifically improved. Ultimately, this facilitates highly efficient and accurate face recognition for varying numbers of live subjects, faces, and scenes, including photos with varying degrees of blur, different background objects, daytime, and nighttime conditions.

[0020] Specifically, the facial image data acquired by the data acquisition includes: facial images captured from different angles, facial images captured under different lighting conditions, and dynamic facial images; The steps for acquiring the face image data are as follows: Step 1, the multi-module image acquisition unit synchronously acquires the face image of the object to be identified. The multi-module image acquisition unit includes a wide-angle camera, a supplementary light camera and an infrared camera. The image data acquired by the multi-module image acquisition unit is recorded as the initial face image set. Step 2: Perform format conversion and resolution unification on the initial face image set, divide the images into valid images, blurred images, and invalid images according to the screening criteria, retain the valid images, and generate a screened face image set; Step 3: The cloud-machine collaborative unit determines the data volume and complexity of the selected face image set, and then performs a threshold comparison of the data volume and complexity. Based on the threshold comparison result, the cloud-processed data and the local-processed data are divided. The cloud server performs preliminary screening, classification and labeling on the cloud-processed data and feeds back the cloud processing results to the device. The device performs classification and labeling on the local-processed data and generates local processing results. The image data of both the cloud-processed results and the local-processed results contain classification labels. The device integrates the cloud-processed results and the local-processed results to generate the face image data.

[0021] Specifically, the face feature dataset generation step of the data processing includes: Step 1, reading a single face image from the face image data, and sequentially performing image noise reduction and grayscale correction processing to generate a first processed image; Step 2: Dynamically allocate face region cropping and edge enhancement processing tasks for the first processed image through the cloud-machine collaboration unit. If the device resources are lower than a preset threshold, the device resources include computing power and power. Then, the face region cropping and edge enhancement processing tasks are allocated to the cloud server for processing and generating a second processed image. The device receives the second processed image. If the device resources are higher than or equal to the preset threshold, the face region cropping and edge enhancement processing tasks are allocated to the device for processing and generating a third processed image. Step 3: Load the algorithm optimization parameters into the deep learning-based feature extraction algorithm, and use the feature extraction algorithm to extract facial key feature points from the second processed image and the third processed image to obtain an initial feature point set. The facial key feature points include: corners of the eyes, tip of the nose, corners of the mouth, and contour features. Step 4: Perform feature quantization and dimensionality reduction on the initial feature point set, remove redundant feature information and retain core feature parameters, add a unique feature identifier to each core feature parameter, and then generate and back up the face feature dataset.

[0022] Specifically, the generation of the face recognition report in the data processing includes: during the execution of steps 1-3, the device performs local liveness detection, determines whether the object being identified is a real live person through dynamic face motion capture and infrared feature recognition, and generates a first liveness determination result. If there is an anomaly in the local liveness detection, the cloud server obtains the abnormal image and detection data and performs a second determination. The cloud server then feeds back the second liveness determination result of the second determination to the device. Step 1: Extract feature identifiers and core feature parameters from the face feature dataset, and then generate feature analysis data; Step 2: The local feature library on the device is synchronized with the cloud feature library on the cloud server; Step 3: The multi-dimensional feature matching model on the device performs preliminary feature similarity calculation based on the feature analysis data and the local feature library to obtain preliminary feature similarity. The multi-dimensional feature matching model includes: a local feature matching algorithm and a global feature matching algorithm. If the preliminary feature similarity is within a preset fuzzy range, the cloud server performs a secondary comparison and feeds back the feature similarity score to the device. The accuracy of the secondary comparison is higher than that of the preliminary feature similarity calculation. If the preliminary feature similarity is within a preset definite range, the feature similarity score is generated. Step 4: Integrate the feature similarity score, the first liveness determination result, and the second liveness determination result. If the first liveness determination result and the second liveness determination result are both determined to be real and the feature similarity score is higher than a preset threshold, then the current recognition is determined to be successful; otherwise, the current recognition is determined to be unsuccessful. Generate a face recognition report and synchronize it to the cloud server.

[0023] Specifically, the parsing of the face recognition report in the execution operation includes: parsing the judgment results and key data in the face recognition report, generating feedback instructions, the feedback instructions being used to drive the display to output visual feedback information, the visual feedback information including: recognition progress, matching percentage and judgment results, and the visual feedback information being synchronized to the cloud server; The preset authorization operations include: authorization verification, data access authorization, unlocking, and log recording, wherein the log recording is synchronized to the cloud server; The anomaly analysis includes: feature library missing analysis and image acquisition anomaly analysis. The cloud server provides optimization suggestions to the device based on the results of the anomaly analysis.

[0024] Specifically, the training data for the lightweight scene classification and recognition AI model and the face enhancement AI model that are processed by the model are low-complexity data, which are photos of a single living person under different backgrounds and facial photos of a single living person with different resolutions, respectively. The photos of a single living person under different backgrounds are divided into multiple groups according to pixel range and grayscale range and input separately for training. The facial photos of a single living person with different resolutions are divided into multiple groups according to pixel range, color range, and age range and input separately for training. The individual live facial photos of different resolutions are photos with scene annotations output by the lightweight scene classification and recognition AI model; The outputs of the lightweight scene classification and recognition AI model and the face enhancement AI model are the lightweight background result of the live object and the lightweight enhanced face image, respectively. The training data for the deep scene classification and recognition AI model and the face enhancement AI model that are processed by the model are medium to high complexity data, which are photos of multiple live objects under different backgrounds and facial photos of multiple live objects with different clarity. The photos of multiple live objects under different backgrounds are divided into multiple groups according to pixel range, grayscale range, and number of live objects and are input separately for training. The facial photos of multiple live objects with different clarity are divided into multiple groups according to pixel range, color range, age range, and number of faces and are input separately for training. The multiple live facial photos of different resolutions are photos with scene annotations output by the deep scene classification and recognition AI model; The low-complexity data and the medium-to-high-complexity data are divided by manual annotation or AI annotation tools; The outputs of the deep scene classification and recognition AI model and the face enhancement AI model are respectively the deep background result of the living object and the deep enhanced face image; The lightweight background result of the live subject and the deep background result of the live subject are used as the processing results of the scene separation and recognition task; Both the lightweight facial enhancement AI model and the deep facial enhancement AI model are equipped with multiple scene processing modes. The lightweight facial enhancement AI model selects the corresponding scene processing mode based on the background result of the lightweight live subject. The deep facial enhancement AI model selects the corresponding scene processing mode based on the background result of the deep facial enhancement.

[0025] Specifically, in the model collaborative processing, the method of allocating scene classification and recognition tasks and face enhancement tasks to the device and cloud server through the cloud machine collaborative processing unit includes: the device analyzes and recognizes the initial face image set to obtain the number of live subjects and the number of faces, and compares them with a preset threshold. If the values ​​are within the threshold range, the device performs local processing; if the values ​​are not within the threshold range, the cloud server performs cloud processing.

[0026] Alternatively, in the model collaborative processing, the method of allocating scene classification and recognition tasks and face enhancement tasks to the device and cloud server through the cloud machine collaborative processing unit includes: allocating tasks according to the amount and complexity of data being processed in the current task.

[0027] The system applying the above method includes: a data acquisition unit, which acquires face image data to be recognized; a cloud-machine collaboration unit, which performs collaborative processing and model collaborative processing; a data processing unit, which generates a face recognition report; and an execution unit, which parses the face recognition report, performs recognition judgment, and executes pass and fail operations for the recognition judgment. The data acquisition unit is equipped with a wide-angle camera, a supplementary lighting camera, and an infrared camera. The data acquisition unit is configured with a supplementary lighting mode, an extreme black light mode, and an AOV mode to achieve scene supplementary lighting, auxiliary image enhancement, and ultimately improve the efficiency and accuracy of face recognition.

[0028] The system's hardware architecture includes: Device side: an STM32H743 microcontroller, 128GB of local Flash memory for storing the local facial feature database; 8GB of RAM for temporary image cache; an 8.4-inch low-power touchscreen with a resolution of 1920×1200 and a touch response time of less than 5ms; a high-capacity 15000mAh rechargeable lithium battery supporting fast charging (18W charging power, approximately 4 hours for a full charge), and a power monitoring chip (monitoring accuracy ±1%). The data acquisition unit is equipped with a wide-angle camera (1080P resolution, 120° field of view), a supplementary lighting camera (infrared supplementary lighting, effective supplementary lighting distance 0-3 meters), and an infrared camera (for liveness detection, wavelength 850nm). The cloud-machine collaborative unit's communication module for data interaction includes: a 4G communication module (supporting Cat.4 standard, downlink peak rate 150Mbps, uplink peak rate 50Mbps) supporting standard SIM card access, with automatic network reconnection function to improve wide area network connection stability; a Wi-Fi module (compliant with IEEE 802.11ax protocol, with higher concurrency, lower latency, and better anti-interference capabilities, with a theoretical single-device transmission rate of up to 9.6Gbps); and a data encryption / decryption chip (using AES-256 symmetric encryption algorithm, compliant with international information security standards; capable of real-time encryption / decryption of transmitted data in 4G wide area links and Wi-Fi 6 local area links). Decryption processing supports custom key configuration and uses a data integrity verification mechanism to prevent data from being stolen, tampered with, or forged during transmission.

[0029] The system's software architecture includes: a deep learning-based face feature extraction model (using a ResNet-18 network structure, with a model size of approximately 100MB), a multi-dimensional feature matching model (integrating local and global feature matching algorithms), and a liveness detection algorithm (combining motion capture and infrared feature recognition); the AI ​​model is trained based on public face datasets such as LFW and CASIA-WebFace, with 200 training rounds and a batch size of 64, after which the trained model needs to be deployed to the core control unit; The cloud-machine collaboration software / program performs data synchronization and task allocation between the device and the cloud, including cloud feature library synchronization protocol and task allocation logic code. It simulates various network environments (such as weak network, network disconnection and reconnection) for testing to ensure the stability and reliability of data synchronization.

[0030] Example 1, specific execution: 1. Face data acquisition: When a person enters the recognition area, the infrared sensor (detection distance 0-5 meters, detection angle 110°) detects the presence of the human body, triggering each camera to start working. Among them, the wide-angle camera, the supplementary lighting camera, and the infrared camera work simultaneously to acquire the face image of the object to be identified; the wide-angle camera acquires face images from different angles at a frame rate of 30 frames / second, the supplementary lighting camera automatically adjusts the supplementary lighting intensity according to the ambient light intensity (supplementary lighting is activated when the light intensity is below 300 lux), and the infrared camera acquires dynamic face images for liveness detection at a frame rate of 25 frames / second. The initial set of collected face images undergoes format conversion (converted to JPEG format) and resolution adjustment (to 800×600 pixels). Blurry images (defined by calculating image sharpness indices, such as variance; images below a threshold of 100 are considered blurry) and invalid images (images where the face area occupies less than 30%) are removed. Valid images that meet preset standards are retained, generating a filtered set of face images. Then, classification labels are established according to the acquisition angle (front, left side, right side), lighting conditions (strong light, weak light, normal light), and image type (static, dynamic), generating face image data with classification labels.

[0031] 2. Face image data preprocessing: Median filtering algorithm is used to reduce noise in face images, with a filtering window size of 3×3; grayscale stretching algorithm is used to adjust the grayscale value range of face images to [0, 255] to enhance image contrast; face detection algorithm (based on Haar features and Adaboost classifier) ​​is used to locate face regions, and then the face regions are cropped out, with the cropped image size being 224×224 pixels; Sobel operator is used to perform edge enhancement processing on the cropped face image to highlight facial contour features.

[0032] 3. Use the built-in AI model to extract facial features and perform liveness detection, cloud-machine collaboration, and identity matching to generate recognition results; call the feature extraction model based on the ResNet-18 network structure to extract key facial feature points (including 68 feature points such as the corners of the eyes, the tip of the nose, the corners of the mouth, and contour feature points) from the preprocessed facial image to obtain the initial feature point set; The initial feature point set is subjected to feature quantization (using Euclidean distance quantization method) and dimensionality reduction (using principal component analysis (PCA) algorithm to retain 95% of the variance), redundant feature information is removed, core feature parameters are retained, and a unique feature identifier is added to each core feature parameter to generate a face feature dataset. The cloud-machine collaboration unit determines whether the local feature library version is consistent with the cloud feature library version. If there is a version difference, it prioritizes receiving the incremental feature library update data pushed by the cloud (using a differential synchronization algorithm, with a synchronization delay of less than 1 second) to complete the local feature library synchronization. Based on a facial feature dataset and a synchronized local feature library, a multi-dimensional feature matching model is used to calculate feature similarity. This model integrates local feature matching algorithms (such as SIFT, where a match is considered valid if more than 20 keypoints are matched) and global feature matching algorithms (such as cosine similarity, with a similarity threshold of 0.9) to obtain a preliminary feature similarity score. If the preliminary feature similarity falls within a preset fuzzy range (0.8-0.95), the feature analysis data is automatically encrypted and uploaded to the cloud (upload bandwidth usage less than 2Mbps). The cloud then uses a higher-precision matching algorithm (such as FaceNet) for a secondary comparison, and the final feature similarity score is received from the cloud (cloud processing latency less than 500ms). The local liveness detection program is initiated. It uses dynamic facial motion capture (requiring the subject to complete at least two specified actions such as blinking and head turning, with an accuracy rate greater than 95%) and infrared feature recognition (detecting the infrared thermal distribution of the face; a match greater than 90% with the actual human body's thermal distribution indicates a liveness detection) to determine if the subject is truly alive and generates a liveness detection result. If an anomaly is detected, the abnormal image and detection data are encrypted and uploaded to the cloud for auxiliary judgment (cloud-based auxiliary judgment latency is less than 1 second), and the program receives corrected liveness detection results from the cloud. The final feature similarity score and the liveness determination result (including correction results) are integrated. If the feature similarity score is higher than the preset threshold (0.9) and the liveness determination is real, the recognition is deemed successful. Otherwise, the recognition is deemed unsuccessful, a complete face recognition report is generated (report generation time is less than 200ms), and the report is synchronized to cloud storage (cloud storage latency is less than 500ms).

[0033] 4. The recognition results are displayed on the screen, and corresponding operations or external terminals are driven based on the results; the judgment results and key data (such as matching percentage, liveness detection results, etc.) in the face recognition report are parsed to generate feedback instructions. Based on the feedback instructions, the 8.4-inch low-power touchscreen outputs visual feedback information, including recognition progress (displayed as a progress bar, refresh rate 1Hz), matching percentage (accurate to one decimal place), and text prompts for the judgment results (such as recognition passed, recognition failed). If the recognition result is successful, the authorization operation is triggered according to the preset rules, such as access control unlocking (controlling the access control motor to rotate, the unlocking time is less than 1 second), system login authorization (sending a login token to the external system, the token is valid for 5 minutes), and data access authorization (generating a data access key, the key length is 256 bits). If the recognition result is a failure, a prompt signal is issued through the audio-visual prompt module (red LED flashing at a frequency of 2Hz; buzzer short beep for 0.5 seconds). The recognition failure time (accurate to milliseconds), the reason for failure (such as feature mismatch, liveness detection failure), and related image data are automatically recorded to the local recognition log (log storage capacity is greater than 1000 entries). At the same time, the failure data is encrypted and uploaded to the cloud, where the cloud performs anomaly analysis (such as missing feature library, abnormal image acquisition, etc.) and receives optimization suggestions returned by the cloud (cloud analysis time is less than 5 seconds) for subsequent device parameter adjustments and algorithm iterations.

[0034] Through the implementation of the above steps, efficient and accurate face recognition based on cloud-machine collaboration is achieved. The reasonable scheduling of work hours and scientific parameter settings for each step ensure the stable operation and good performance of the entire system.

[0035] In summary, the above methods and systems achieve data synchronization and task allocation between the device and the cloud server. The device performs local processing to ensure the real-time performance of basic recognition, while the cloud server leverages its high computing power to perform collaborative processing, meeting the processing needs of high-data, high-complexity scenarios. The linkage mechanism between the two effectively solves the problems of poor flexibility, insufficient performance, and limited battery life of existing devices. The combination of pre-set facial feature library updates and local feature libraries avoids data gaps. The bidirectional data synchronization, data analysis optimization, and optimization feedback between the device and the cloud server enable the entire system to move from deployment to mature use more quickly. Ultimately, while ensuring the portability and battery life of the device, the system improves the timeliness, accuracy, and multi-scenario adaptability of facial recognition.

[0036] The embodiments described above are merely illustrative of one or more implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A cloud machine collaboration based face recognition method, characterized in that, The method comprises: data acquisition, acquiring face image data to be identified; Cooperative processing, through a cloud machine cooperative processing unit, the device end uploads part of the face image data or its preprocessing requirements to the cloud server, the device end receives the preset face feature library update data and algorithm optimization parameters fed back by the cloud server, and the device end and the cloud server perform data synchronization; Model cooperative processing, the device end loads a lightweight scene classification and recognition AI model and a face enhancement AI model, the cloud server loads a deep scene classification and recognition AI model and a face enhancement AI model, and the cloud machine cooperative processing unit allocates scene classification and recognition tasks and face enhancement tasks to the device end and the cloud server, wherein the processing result of the cloud server is fed back to the device end, the processing result of the scene classification and recognition task is loaded as input to the face enhancement task, and the processing result of the face enhancement task is integrated into the face image data; Data processing, the device end performs preprocessing and feature extraction on the face image data to generate a face feature data set with feature labels, performs comparison and liveness verification based on the face feature data set, the preset face feature library update data and the local feature library through a multi-dimensional feature matching model, and generates a face recognition report containing a matching result and a liveness determination result; Execution operation, analyze the face recognition report and perform identification determination, if the identification is passed, trigger a preset authorization operation, if the identification fails, upload the current failure data to the cloud server and perform abnormal analysis, prompt signal output and update the identification failure log. 2.The cloud-based machine collaboration face recognition method of claim 1, wherein, The face image data obtained by the data acquisition includes face images collected at different angles, face images collected under different light environments and dynamic face images; The face image data acquisition step comprises: step 1, a multi-module image acquisition unit synchronously collects face images of an object to be identified, the multi-module image acquisition unit comprises a wide-angle camera, a light supplement camera and an infrared camera, and the image data obtained by the multi-module image acquisition unit is recorded as an initial face image set; Step 2, performing format conversion and resolution unification on the initial face image set, dividing valid images, blurred images and invalid images according to the screening standard, retaining the valid images and generating a screened face image set; Step 3, determining the data volume and complexity of the screened face image set through a cloud machine cooperative unit, then performing threshold comparison of the data volume and the complexity, and dividing cloud processing data and local processing data according to the threshold comparison result, the cloud server performs preliminary screening, classification labeling on the cloud processing data and feeds back a cloud processing result to the device end, the device end performs classification labeling on the local processing data and generates a local processing result, the image data of the cloud processing result and the local processing result both contain classification labels, and the device end integrates the cloud processing result and the local processing result and generates the face image data. 3.The cloud-based machine collaboration face recognition method of claim 1, wherein, The face feature data set generation step of the data processing comprises: step 1, reading a single face image from the face image data, and sequentially performing image noise reduction and gray correction processing to generate a first processed image; Step 2, dynamically allocate the face region cropping and edge enhancement processing task for the first processed image through the cloud machine cooperation unit, if the device end resource is lower than the preset threshold, the device end resource includes computing power and power, then the face region cropping and edge enhancement processing task is allocated to the cloud server for processing and generating a second processed image, the device end receives the second processed image, if the device end resource is higher than or equal to the preset threshold, then the face region cropping and edge enhancement processing task is allocated to the device end for processing and generating a third processed image; Step 3, load the algorithm optimization parameters to the feature extraction algorithm based on deep learning, perform face key feature point extraction on the second processed image and the third processed image through the feature extraction algorithm, and obtain an initial feature point set, the face key feature point includes: eye corner, nose tip, mouth corner and contour feature; Step 4, perform feature quantization and dimension reduction on the initial feature point set, remove redundant feature information and retain core feature parameters, then add a unique feature identifier to each core feature parameter, and then generate and backup the face feature data set. 4.The cloud-based machine collaboration face recognition method of claim 1, wherein, The face recognition report generation of the data processing comprises: During the execution of steps 1-3, the device end performs local liveness detection, judges whether the recognition object is a real living body through dynamic face motion capture and infrared feature recognition, and generates a first liveness determination result, if the local liveness detection is abnormal, the cloud server acquires abnormal image and detection data and performs secondary determination, the cloud server feeds back a second liveness determination result of the secondary determination to the device end; Step 1, propose a feature identifier and a core feature parameter from the face feature data set, and then generate feature analysis data; Step 2, the local feature library of the device end and the cloud feature library of the cloud server perform synchronization; Step 3, the multi-dimensional feature matching model of the device end performs preliminary feature similarity calculation according to the feature analysis data and the local feature library and obtains a preliminary feature similarity, the multi-dimensional feature matching model includes: a local feature matching algorithm and a global feature matching algorithm; if the preliminary feature similarity is in a preset fuzzy interval, the cloud server performs secondary comparison and feeds back a feature similarity score to the device end, the accuracy of the secondary comparison is higher than that of the preliminary feature similarity calculation; if the preliminary feature similarity is in a preset clear interval, the feature similarity score is generated; Step 4, integrate the feature similarity score, the first liveness determination result and the second liveness determination result, if the first liveness determination result and the second liveness determination result are determined to be real and the feature similarity score is higher than a preset threshold, it is determined that the current recognition is passed, otherwise it is determined that the current recognition fails, a face recognition report is generated and synchronized to the cloud server. 5.The cloud-based machine collaboration face recognition method of claim 1, wherein, The analyzing the face recognition report of the executing operation comprises: analyzing a judgment result and key data in the face recognition report, generating a feedback instruction, and the feedback instruction is used to drive a display to output visual feedback information, the visual feedback information comprises: identification progress, matching degree percentage and judgment result, and the visual feedback information is synchronized to a cloud server; The preset authorized operation comprises: authorization verification, data access authorization, unlocking and log recording, wherein the log recording is synchronized to the cloud server; The abnormality analysis comprises: feature library missing analysis and image acquisition abnormality analysis, and the cloud server feeds back optimization suggestions to the device end according to a result of the abnormality analysis. 6.The cloud-based machine collaboration face recognition method of claim 1, wherein, The training data of the lightweight scene classification and identification AI model and the face enhancement AI model in the model cooperative processing are low complexity data, which are respectively single living body photos in different backgrounds and single living body face photos with different clarity, wherein the single living body photos in different backgrounds are divided into multiple groups according to pixel range and gray scale range and are input separately for training, and the single living body face photos with different clarity are divided into multiple groups according to pixel range, color range and person age range and are input separately for training. The single living body face photos with different clarity are photos with scene labels output by the lightweight scene classification and identification AI model. The outputs of the lightweight scene classification and identification AI model and the face enhancement AI model are respectively lightweight background results of living bodies and lightweight enhanced face images. The training data of the deepened scene classification and identification AI model and the face enhancement AI model in the model cooperative processing are medium-high complexity data, which are respectively multiple living body photos in different backgrounds and multiple living body face photos with different clarity, wherein the multiple living body photos in different backgrounds are divided into multiple groups according to pixel range, gray scale range and living body quantity and are input separately for training, and the multiple living body face photos with different clarity are divided into multiple groups according to pixel range, color range, person age range and face quantity and are input separately for training. The multiple living body face photos with different clarity are photos with scene labels output by the deepened scene classification and identification AI model. The low complexity data and the medium-high complexity data are divided by manual labeling or AI labeling tools. The outputs of the deepened scene classification and identification AI model and the face enhancement AI model are respectively deepened background results of living bodies and deepened enhanced face images. The lightweight background results of living bodies and the deepened background results of living bodies are processing results of the scene separation identification task. The lightweight face enhancement AI model and the deepened face enhancement AI model are both provided with multiple scene processing modes. The lightweight face enhancement AI model selects a corresponding scene processing mode according to the lightweight background result of the living body. The deepened face enhancement AI model selects a corresponding scene processing mode according to the deepened background result of the living body. 7.The cloud-based machine collaboration face recognition method of claim 6, wherein, In the model cooperative processing, the cloud machine cooperative processing unit assigns the scene classification identification task and the face enhancement task to the device end and the cloud server in the following manner: the device end analyzes and identifies the initial face image set to obtain the number of living bodies and the number of faces, and compares the number of living bodies and the number of faces with a preset threshold value; if the number of living bodies and the number of faces are within the threshold value range, the device end performs local processing; if the number of living bodies and the number of faces are not within the threshold value range, the cloud server performs cloud processing. 8.The cloud-based machine collaboration face recognition method of claim 1, wherein, In the model cooperative processing, the cloud machine cooperative processing unit assigns the scene classification identification task and the face enhancement task to the device end and the cloud server in the following manner: the cloud machine cooperative processing unit assigns the scene classification identification task and the face enhancement task to the device end and the cloud server according to the data volume and complexity of the current task processing.

9. A system applying the method according to any one of claims 1 to 8, characterized in that, The system comprises: a data acquisition unit, the data acquisition module is used for acquiring face image data to be identified; a cloud machine cooperative unit, the cloud machine cooperative unit is used for executing cooperative processing and model cooperative processing; a data processing unit, the data processing unit is used for generating a face recognition report; an execution unit, the execution unit is used for analyzing the face recognition report and performing identification judgment, executing pass and failure operations of the identification judgment.

10. The system of claim 9, wherein, The data acquisition unit is provided with a wide-angle camera, a light supplement camera and an infrared camera, and the data acquisition unit is provided with a light supplement mode, an extremely dark light mode and an AOV mode.

Images