A cloud-edge collaborative-based face recognition system and method

Abstract

This invention relates to the field of facial recognition technology, specifically a cloud-edge collaborative facial recognition system and method. The system comprises at least one edge node and a cloud service platform. The edge node includes modules for facial image acquisition and preprocessing, lightweight adaptive feature extraction, and edge matching. The cloud node includes modules for high-precision feature extraction, cloud-edge collaborative scheduling, and global matching. The method performs facial image preprocessing, adaptive feature compression, and local matching at the edge node, and uses a dynamic decision threshold to determine whether to invoke high-precision recognition from the cloud. The cloud node performs global matching, dynamic scheduling, and synchronous update of feature increments. This invention can dynamically adapt to changes in the edge node's operating environment and scenarios, balancing real-time performance and accuracy, reducing cloud computing load and bandwidth consumption, improving system robustness and scenario adaptability, and is suitable for large-scale deployment of distributed facial recognition in multiple scenarios.

Description

Technical Field

[0001] This invention relates to the field of facial recognition technology, specifically to a cloud-edge collaborative facial recognition system and method. Background Technology

[0002] With the rapid development of artificial intelligence technology, facial recognition technology has been widely applied in various scenarios such as smart parks, traffic checkpoints, commercial complexes, and rail transit, placing higher demands on the real-time performance, accuracy, and concurrent processing capabilities of the system. The cloud-edge collaborative architecture, by offloading some computing tasks to edge nodes, can effectively reduce the computing pressure and data transmission bandwidth consumption of the cloud service platform, solving problems such as high recognition latency, excessive cloud load, and strong network dependence inherent in centralized facial recognition systems. This has become an important research direction in the current field of facial recognition.

[0003] In the prior art, Chinese invention patent application CN117765655A discloses a fast face recognition method and system based on cloud-edge collaboration. This solution is mainly applied to subway turnstile scenarios. Its core technical content is as follows: constructing a cloud-edge collaborative architecture consisting of a cloud-based face recognition center platform and multiple station-level face recognition subsystems, with one face recognition subsystem corresponding to each subway station; after the turnstile collects passenger face data, it sends it to the face recognition subsystem of the corresponding station. The subsystem performs local recognition based on the locally stored face feature values. If the recognition is successful, it directly generates a turnstile opening command; if the recognition fails, it uploads the face data to the cloud platform for secondary recognition; after successful recognition by the cloud, it issues a turnstile control command and simultaneously sends the corresponding face feature values ​​to the edge subsystem for storage; in addition, this solution uses LFU algorithm, time weighting method and other elimination algorithms to periodically refresh the local face feature database at the edge, deleting infrequently used face features and retaining frequently used features, thereby improving the edge recognition speed and reducing the computing pressure on the cloud platform.

[0004] While the technical solutions described in the aforementioned comparative documents achieve basic cloud-edge collaborative face recognition and alleviate the problem of excessive cloud load in centralized recognition systems to some extent, they still have significant technical shortcomings in practical applications: First, the edge of this solution uses a fixed feature extraction and matching mode, without adaptive feature compression capability. It cannot dynamically adjust the feature compression ratio according to the real-time computing load of the edge node, network bandwidth status, and the quality of the input face image. In scenarios such as peak passenger flow and high load, and network environment fluctuations, it is easy to encounter problems such as increased recognition latency and decreased feature discrimination, and it is impossible to achieve a dynamic balance between recognition efficiency and matching accuracy.

[0005] Secondly, the edge recognition of this scheme uses a fixed decision threshold and does not dynamically adjust it based on key factors such as the intensity of scene interference, the coverage of the target population in the local face feature library, and the confidence of the extracted features. In scenarios with strong interference such as complex lighting and face occlusion, the false recognition rate is prone to increase, and in scenarios with low coverage of the local feature library, the missed recognition rate is prone to increase, resulting in insufficient robustness of the system recognition.

[0006] Third, the solution does not specifically optimize the feature extraction model for cloud-edge collaborative scenarios. It only uses conventional classification loss to complete model training, without simultaneously taking into account intra-class feature aggregation constraints and quantization compression loss constraints. This results in insufficient distinguishability of facial features after compression at the edge, directly reducing the accuracy of local matching and increasing the frequency of secondary recognition calls in the cloud, thus failing to fundamentally reduce the computational pressure on the cloud.

[0007] Fourth, the cloud-edge collaborative scheduling capability of this solution is insufficient. It can only achieve incremental updates of the feature library at fixed periods. It cannot dynamically issue appropriate scheduling parameters based on the real-time operating status of each edge node, nor can it iteratively optimize the global feature library in the cloud based on the high-confidence recognition results of the edge. The depth of cloud-edge collaboration is insufficient, and the recognition accuracy and operating efficiency of the entire system cannot achieve continuous closed-loop optimization. Summary of the Invention

[0008] The purpose of this invention is to provide a cloud-edge collaborative face recognition system and method to solve the problems mentioned in the background art, such as the inability of existing cloud-edge collaborative face recognition technologies to balance recognition efficiency and accuracy, insufficient robustness, and lack of depth in cloud-edge collaboration.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A cloud-edge collaborative face recognition system includes at least one edge node and a cloud service platform that communicates with each edge node. The edge node includes a face acquisition and preprocessing module, a lightweight adaptive feature extraction module, a local face feature library, an edge matching module, and an edge collaborative interaction module. The face acquisition and preprocessing module acquires real-time face images and performs preprocessing, outputting standardized face images. The lightweight adaptive feature extraction module extracts features from the standardized face images and adaptively compresses them, outputting edge-compressed face features. The edge matching module matches the edge-compressed face features with features in the local face feature library, outputting local matching results. The edge collaborative interaction module, based on the local matching results, completes feature data interaction and command transmission / reception with the cloud service platform. The cloud service platform includes a global face feature database, a high-precision feature extraction module, a cloud-edge collaborative scheduling module, a global matching module, and a feature update management module. The high-precision feature extraction module receives standardized face images uploaded by edge nodes and outputs high-precision face features. The global matching module performs full matching between the high-precision face features and the global face feature database and outputs the global matching result. The cloud-edge collaborative scheduling module issues scheduling parameters for feature extraction and matching based on the running status data of the edge nodes. The feature update management module performs incremental synchronous updates of the global face feature database and the local face feature databases of each edge node.

[0010] Preferably, the lightweight adaptive feature extraction module incorporates an adaptive feature compression unit. This unit calculates the optimal feature compression ratio α based on the real-time operating status of the edge nodes and the quality of the input face image. The calculation formula is as follows: ; In the formula, To preset the minimum compression ratio, To preset the maximum compression ratio, This represents the current computing load rate of the edge nodes. This represents the current network bandwidth availability of the edge nodes. To standardize the quality score of face images, These are preset weight coefficients corresponding to computing power load, bandwidth availability, and image quality, respectively; the adaptive feature compression unit quantizes and compresses the extracted initial face features according to the calculated optimal feature compression ratio α to generate edge-compressed face features.

[0011] Preferably, the edge matching module has a built-in dynamic threshold judgment unit, which is used to calculate the decision threshold T for local matching based on the scheduling parameters sent from the cloud. The calculation formula is as follows: ; In the formula, As a preset baseline threshold, For environmental interference factors in the current scene, To determine the target population coverage of the local facial feature database, The confidence score for edge-compressed facial features. These are preset adjustment coefficients corresponding to environmental interference, coverage, and feature confidence, respectively. The dynamic threshold judgment unit compares the highest similarity score of the local match with the decision threshold T. If the highest similarity score is greater than or equal to T, it outputs the local match result that is successfully matched. If the highest similarity score is less than T, it triggers the edge collaborative interaction module to upload a standardized face image to the cloud service platform.

[0012] Preferably, the lightweight adaptive feature extraction module uses an improved multi-task joint loss function to complete model training. The expression for the multi-task joint loss function L is: ; In the formula, Cross-entropy classification loss, The center loss is used to reduce the intra-class feature distance. The quantization loss is used to constrain the distance between the compressed features and the original features. These are preset balance weights for center loss and quantization loss, respectively; the lightweight adaptive feature extraction module extracts and compresses facial features based on the trained model.

[0013] Preferably, the feature update management module has a built-in incremental synchronization unit. The incremental synchronization unit is used to count the newly added, modified, and deleted feature data in the global face feature database and generate incremental feature packages. The incremental synchronization unit splits the incremental feature packages into data segments according to the local face feature database version number and network status of each edge node, and sends them to the corresponding edge nodes at a preset period to complete the incremental update of the local face feature database. The incremental synchronization unit is also used to receive locally matched high-confidence face features uploaded by edge nodes and iteratively optimize the corresponding features in the global face feature database.

[0014] On the other hand, the present invention also provides a cloud-edge collaborative face recognition method, applied to the aforementioned cloud-edge collaborative face recognition system, comprising the following steps: S1. Edge nodes acquire real-time face images, and output standardized face images after preprocessing; S2. Edge nodes perform feature extraction and adaptive compression on standardized face images to generate edge-compressed face features; S3. The edge node performs local matching between the compressed face features and the local face feature database to generate a local matching result; S4. The edge node determines whether to upload the standardized face image to the cloud service platform based on the local matching result and the scheduling parameters issued by the cloud. If the determination is yes, proceed to step S5; if the determination is no, directly output the local matching result as the final recognition result. S5. The cloud service platform receives standardized face images, generates high-precision face features through the high-precision feature extraction module, performs full matching of the high-precision face features with the global face feature library, generates global matching results, and sends them to the corresponding edge nodes as the final recognition results. S6. The cloud service platform updates and distributes scheduling parameters based on the operating status data of each edge node, and simultaneously completes incremental synchronous updates of the global face feature database and the local face feature database of each edge node.

[0015] Preferably, in step S2, the feature extraction and adaptive compression of the standardized face image specifically includes: S21. Based on the trained lightweight feature extraction model, perform feature extraction on the standardized face image to generate initial face features; S22. Collect the current computing power load rate ρ and the current network bandwidth availability β of the edge node, calculate the quality score Q of the standardized face image, and substitute it into the optimal feature compression ratio calculation formula to obtain the optimal feature compression ratio α; S23. Based on the optimal feature compression ratio α, the initial face features are quantized and compressed to generate edge-compressed face features; The formula for calculating the optimal feature compression ratio is: ; In the formula, To preset the minimum compression ratio, To preset the maximum compression ratio, These are preset weighting coefficients corresponding to computing load, bandwidth availability, and image quality, respectively.

[0016] Preferably, in step S4, determining whether to upload the standardized face image to the cloud service platform based on the local matching result and the scheduling parameters issued by the cloud specifically includes: S41. Extract the highest similarity score from the local matching results, and calculate the decision threshold T for local matching by combining it with the scheduling parameters issued by the cloud. S42. If the highest similarity score is greater than or equal to the decision threshold T, it is determined that no upload is needed, and the local matching result is directly output as the final recognition result. S43. If the highest similarity score is less than the decision threshold T, it is determined to be an upload, and the standardized face image is uploaded to the cloud service platform; The formula for calculating the decision threshold T is: ; In the formula, As a preset baseline threshold, For environmental interference factors in the current scene, To determine the target population coverage of the local facial feature database, The confidence score for edge-compressed facial features. These are the preset adjustment coefficients corresponding to environmental interference, coverage, and feature confidence, respectively.

[0017] Preferably, in step S2, the lightweight feature extraction model is trained using an improved multi-task joint loss function, the expression of which is: ; In the formula, Cross-entropy classification loss, Loss at the center To quantify the loss, These are preset balance weights for the center loss and quantization loss, respectively; during training, the model parameters are iteratively updated by minimizing the multi-task joint loss function L.

[0018] Preferably, in step S6, the incremental synchronization update of the global face feature database and the local face feature database of each edge node specifically includes: S61. The cloud service platform counts the newly added, modified, and deleted feature data in the global face feature database and generates incremental feature packages with corresponding version numbers; S62. Collect the local face feature library version number and real-time network status of each edge node, split the incremental feature package into data segments, and send them to the corresponding edge nodes according to a preset cycle; S63. Edge nodes receive fragmented data and complete reassembly, update the local face feature database and version number, and upload the high-confidence face features that are successfully matched locally to the cloud service platform. S64. The cloud service platform receives high-confidence facial features and iteratively optimizes the corresponding features in the global facial feature database.

[0019] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention addresses the technical shortcomings of existing cloud-edge collaborative face recognition technologies, where the fixed feature extraction and compression mode at the edge cannot adapt to dynamic operating environments and is difficult to balance recognition efficiency and matching accuracy. By using a lightweight adaptive feature extraction module at the edge, this invention constructs a dynamic calculation model for the optimal feature compression ratio that integrates edge computing load rate, network bandwidth availability, and face image quality. The feature compression ratio can be dynamically adjusted according to the real-time operating status of the edge and the quality of the input image. Under high computing load and weak network conditions, the feature dimension and data transmission volume are reduced, while under low load and good network conditions, the feature discrimination is guaranteed, thus achieving a dynamic balance between recognition efficiency and matching accuracy. At the same time, to address the accuracy loss caused by feature compression at the edge, this invention designs a multi-task joint loss function that integrates cross-entropy classification loss, center loss, and quantization loss. During the training of the lightweight model, it simultaneously constrains feature classification accuracy, intra-class aggregation degree, and feature consistency before and after compression, so that the compressed edge face features still maintain high discrimination, effectively improving the accuracy of local matching, reducing unnecessary secondary recognition calls in the cloud, fundamentally reducing the computing pressure and network bandwidth occupation of the cloud service platform, and improving the overall real-time response of the system.

[0020] (2) In view of the technical defects of insufficient robustness of fixed matching decision threshold at the edge and weak cloud-edge collaborative scheduling capability in the prior art, the present invention constructs a dynamic decision threshold calculation model that integrates scene environment interference factors, local face feature library target population coverage and edge compressed face feature confidence through a dynamic threshold judgment unit set at the edge. This realizes the dynamic adaptation of local matching decision logic with application scenarios and system status, solves the technical problems of increased false recognition rate in complex interference scenarios and increased missed recognition rate in low coverage scenarios with fixed threshold, and effectively improves the robustness and accuracy of face recognition in different application scenarios. Meanwhile, this invention, through a cloud-edge collaborative scheduling module set up in the cloud, can collect real-time operational status data of each edge node and issue differentiated scheduling parameters, realizing dynamic optimization and configuration of cloud-edge computing and storage resources. Through the incremental synchronization unit of the feature update management module, it can achieve accurate incremental synchronization between the global face feature library and the local face feature library at the edge. At the same time, based on the high-confidence face features uploaded from the edge, it iteratively optimizes the global face feature library, constructing a closed-loop mechanism for continuous optimization of system recognition accuracy. This solves the problems of insufficient cloud-edge collaboration depth and inability to continuously iterate and optimize system performance in existing technologies, and improves the long-term operational stability and recognition accuracy of the entire system.

[0021] (3) The modular and scalable cloud-edge collaborative face recognition system and method constructed in this invention breaks through the limitation of existing technologies that are only adapted to a single rail transit gate scenario. It can flexibly adapt to various distributed face recognition scenarios such as smart parks, urban traffic checkpoints, and commercial complexes through differentiated configuration of scheduling parameters. It is compatible with edge nodes with different computing power levels and network conditions, and has extremely strong scenario adaptability and system scalability. This invention reasonably sinks the core computing tasks of face recognition to the edge and optimizes the computing efficiency of the edge through a dynamic adaptive mechanism. Without increasing the hardware cost of the edge, it achieves high real-time performance and high accuracy of distributed face recognition, while significantly reducing the hardware investment and operating costs of the cloud service platform, and has significant industrial applicability. The technical solution of this invention fills the technical gaps in existing cloud-edge collaborative face recognition technologies in dynamic adaptive optimization, full-link accuracy-efficiency balance control, and closed-loop accuracy iterative optimization, and provides a complete, stable, and efficient technical solution for face recognition applications in large-scale, multi-node, and complex scenarios. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0023] Figure 1 This is a block diagram of the cloud-edge collaborative face recognition system of the present invention; Figure 2 This is a detailed block diagram of the edge node composition in this invention; Figure 3 This is a block diagram of the cloud service platform in this invention; Figure 4 This is a flowchart of the cloud-edge collaborative face recognition method of the present invention; Figure 5 A detailed flowchart of the key steps of this invention is provided. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figures 1-5 As shown, the cloud-edge collaborative face recognition system of the present invention includes at least one edge node and a cloud service platform that establishes a two-way connection with each edge node through a wireless or wired communication network.

[0026] Edge nodes are deployed at the front end of face recognition scenarios, responsible for real-time acquisition, preprocessing, local feature extraction and matching of face images, while also completing collaborative interaction with the cloud. Edge nodes include a face acquisition and preprocessing module, a lightweight adaptive feature extraction module, a local face feature library, an edge matching module, and an edge collaborative interaction module.

[0027] The face capture and preprocessing module acquires face images from the real-time video stream using a high-definition camera unit. It sequentially performs face detection, key point localization, face alignment, and normalization, outputting standardized face images with uniform size and pixel specifications. Face detection employs a lightweight neural network-based face detection model for rapid localization of face regions within the image. Key point localization extracts 68 feature points of the facial features, performing face pose correction and alignment based on these feature points. Finally, the face image is normalized to a pre-defined standardized face image, and a quality score is calculated for the standardized face image. Quality Score The value ranges from 0 to 1. A higher score indicates better clarity, lighting uniformity, and pose compliance of the face image.

[0028] The lightweight adaptive feature extraction module integrates a lightweight feature extraction network and an adaptive feature compression unit to extract features from standardized face images and adaptively compress them, outputting edge-compressed face features. The lightweight feature extraction network employs an improved MobileNetV3 backbone network, reducing network parameters and computational cost through depthwise separable convolutions, adapting to low-computational-power environments at edge nodes. The lightweight feature extraction network uses an improved multi-task joint loss function for model training. The expression is: ; In the formula, Cross-entropy is the classification loss used to constrain the classification accuracy of facial features. The center loss is used to reduce the intra-class feature distance of faces with the same identity and improve the inter-class feature discrimination. To quantify the loss, a constraint is used to set the Euclidean distance between the compressed features and the original high-dimensional features, thereby reducing the accuracy loss caused by feature compression. and These are preset balancing weights for the center loss and quantization loss, each ranging from 0 to 1. During model training, the multi-task joint loss function is minimized using the gradient descent algorithm. This allows for iterative updates of network parameters, enabling the trained lightweight feature extraction network to simultaneously meet the requirements of feature extraction accuracy and lightweight deployment at the edge.

[0029] The adaptive feature compression unit is used to calculate the optimal feature compression ratio based on the real-time running status of the edge nodes and the quality of the input face image. Based on the optimal feature compression ratio The initial face features output by the lightweight feature extraction network are quantized and compressed to generate edge-compressed face features. The optimal feature compression ratio is then determined. The calculation formula is: ; In the formula, To preset the minimum compression ratio, To preset the maximum compression ratio, The range of values ​​is to . This represents the current computing load rate of the edge node, with a value ranging from 0 to 1. This represents the current network bandwidth availability of the edge node, with a value ranging from 0 to 1. The quality score is used to standardize the face image, with a value ranging from 0 to 1. , , These are preset weighting coefficients corresponding to computing load, bandwidth availability, and image quality, respectively, and satisfy the following conditions: .

[0030] The local face feature database stores target face feature data for corresponding scenarios. Each feature data entry includes a unique face identifier, compressed face features, feature version number, and valid time range information. The data in the local face feature database is updated incrementally by the cloud service platform to maintain version consistency with the global face feature database in the cloud.

[0031] The edge matching module incorporates a feature matching unit and a dynamic threshold determination unit. It matches edge-compressed facial features with features in the local facial feature database and outputs the local matching result. The feature matching unit uses a cosine similarity algorithm to calculate the similarity score between the edge-compressed facial features and each feature in the local facial feature database, outputting the match with the highest similarity score and its corresponding highest similarity score. The dynamic threshold determination unit calculates the decision threshold for local matching based on scheduling parameters sent from the cloud. Decision threshold The calculation formula is: ; In the formula, The preset baseline threshold value ranges from 0 to 1. This represents the environmental interference factor of the current scene, with a value ranging from 0 to 1. The higher the value, the stronger the interference factors such as lighting, occlusion, and pose in the current scene. This represents the coverage rate of the target population in the local facial feature database, ranging from 0 to 1. A higher value indicates a higher proportion of the target population covered by the local facial feature database. The confidence score is the edge-compressed facial feature score, ranging from 0 to 1. A higher value indicates a stronger reliability of the compressed feature. , , These are preset adjustment coefficients corresponding to environmental interference, coverage, and feature confidence, respectively. The dynamic threshold judgment unit compares the highest similarity score with the decision threshold. Compare the results; if the highest similarity score is greater than or equal to... Output the successful local matching results, which include the face's unique identifier and the matching similarity score. If the highest similarity score is less than... It outputs the local matching results that are not confirmed, and at the same time triggers the edge collaborative interaction module to upload standardized face images to the cloud service platform.

[0032] The edge collaborative interaction module is used to complete the interaction of feature data and command transmission and reception with the cloud service platform based on the local matching results. When the local matching result is "match not confirmed," the edge collaborative interaction module uploads the standardized face image to the cloud service platform and simultaneously receives the global matching result issued by the cloud service platform. The edge collaborative interaction module is also used to receive scheduling parameters and incremental feature packets issued by the cloud service platform, and upload the running status data of the edge nodes and the high-confidence face features of locally matched high-confidence faces to the cloud service platform.

[0033] The cloud service platform is deployed on a cloud server cluster and is responsible for the management of global facial features, high-precision facial feature extraction and full matching, cloud-edge collaborative scheduling, and incremental feature updates. The cloud service platform includes a global facial feature library, a high-precision feature extraction module, a cloud-edge collaborative scheduling module, a global matching module, and a feature update management module.

[0034] The global face feature library stores high-precision face feature data for all scenarios. Each feature data entry includes a unique face identifier, high-precision face features, feature version number, valid time range, and authorized scenario information. The global face feature library supports the storage and rapid retrieval of tens of millions of face features and adopts a distributed storage architecture to ensure the stability and efficiency of data read and write operations.

[0035] The high-precision feature extraction module adopts a high-precision face recognition model based on deep residual networks. It receives standardized face images uploaded by edge nodes and extracts and generates high-precision face features with higher dimensions and stronger discrimination to ensure the accuracy of face recognition in complex scenarios.

[0036] The global matching module uses a full-scale matching algorithm based on a vector retrieval engine to quickly match high-precision facial features with all features in the global facial feature database and output the global matching results. The global matching results include the unique identifier of the successfully matched face and the matching similarity, or the recognition results of the unmatched face.

[0037] The cloud-edge collaborative scheduling module is used to collect real-time operational status data from each edge node. This data includes computing load rate, network bandwidth availability, environmental interference factors, local matching accuracy, and identification request frequency. Based on this operational status data, scheduling parameters for feature extraction and matching are generated and distributed. These parameters include weighting coefficients. , , adjustment coefficient , , Preset baseline threshold The feature synchronization period information. The cloud-edge collaborative scheduling module can issue differentiated scheduling parameters based on the scenario differences and operating status of different edge nodes, so as to realize the dynamic optimization and configuration of cloud-edge resources.

[0038] The feature update management module incorporates an incremental synchronization unit to perform incremental synchronization updates between the global face feature database and the local face feature databases of each edge node. The incremental synchronization unit continuously monitors newly added, modified, and deleted feature data in the global face feature database, generating incremental feature packages with corresponding version numbers. Based on the version number of the local face feature database of each edge node and the real-time network status, the incremental feature package is split into data segments and distributed to the corresponding edge nodes at preset intervals. The edge nodes receive the segmented data, reassemble them, and then update their local face feature database and version number. The incremental synchronization unit also receives locally matched high-confidence face features uploaded by edge nodes and iteratively optimizes the corresponding features in the global face feature database to improve the scene adaptability and recognition accuracy of the features.

[0039] The cloud-edge collaborative face recognition method of the present invention, applied to the above system, specifically includes the following steps: S1. Edge nodes acquire real-time face images through the face acquisition and preprocessing module, sequentially performing face detection, key point localization, face alignment, and normalization preprocessing to output standardized face images. Simultaneously, the quality score of the standardized face images is calculated. .

[0040] S2. Edge nodes use a lightweight adaptive feature extraction module to extract features from standardized face images and adaptively compress them to generate edge-compressed face features, specifically including: S21. Based on the trained lightweight feature extraction model, feature extraction is performed on the standardized face image to generate initial high-dimensional face features.

[0041] S22. Real-time collection of current computing load rate of edge nodes Current network bandwidth availability Combined with the quality score of standardized face images Substituting into the formula for calculating the optimal feature compression ratio, we obtain the optimal feature compression ratio. .

[0042] S23. Based on the optimal feature compression ratio The initial high-dimensional facial features are quantized, compressed, and dimensionality reduced to generate edge-compressed facial features. Simultaneously, the confidence score of the edge-compressed facial features is calculated. .

[0043] S3. The edge node uses the edge matching module to perform local matching between the compressed face features and the local face feature database, generating a local matching result. The cosine similarity algorithm is used to calculate the similarity score between the compressed face features and each feature in the local face feature database. The highest similarity score and its corresponding match are then selected to generate the local matching result.

[0044] S4. The edge node, based on the local matching result and the scheduling parameters issued by the cloud, determines whether to upload the standardized face image to the cloud service platform. If the determination is yes, proceed to step S5. If the determination is no, directly output the local matching result as the final recognition result. Specifically, this includes: S41. Extract the highest similarity score from the local matching results, and combine it with the scheduling parameters issued by the cloud and the environmental interference factors of the current scene. Target population coverage of the local facial feature database Confidence score of edge-compressed facial features Substituting into the decision threshold calculation formula, we obtain the decision threshold for local matching. .

[0045] S42. If the highest similarity score is greater than or equal to the decision threshold If the result is determined not to be uploaded, the local matching result will be directly output as the final recognition result.

[0046] S43. If the highest similarity score is less than the decision threshold If the image is determined to be an upload, the standardized facial image is uploaded to the cloud service platform through the edge collaborative interaction module.

[0047] The S5 cloud service platform receives standardized face images, generates high-precision face features through a high-precision feature extraction module, performs full matching between the high-precision face features and the global face feature library, generates a global matching result, and sends it to the corresponding edge nodes as the final recognition result.

[0048] S6. The cloud service platform updates and distributes scheduling parameters based on the operational status data of each edge node, and simultaneously performs incremental synchronization updates between the global face feature database and the local face feature database of each edge node. Specifically, this includes: S61. The cloud service platform counts the newly added, modified, and deleted feature data in the global face feature database and generates incremental feature packages with corresponding version numbers.

[0049] S62. Collect the local face feature library version number and real-time network status of each edge node, split the incremental feature package into data segments, and distribute them to the corresponding edge nodes according to a preset cycle.

[0050] S63. Edge nodes receive fragmented data and complete reassembly, update the local face feature database and version number, and upload the locally matched high-confidence face features to the cloud service platform.

[0051] S64. The cloud service platform receives high-confidence facial features and iteratively optimizes the corresponding features in the global facial feature database.

[0052] Example 1: Smart Park Access Control with Facial Recognition This embodiment is applied to a smart park access control facial recognition scenario. The system includes eight edge nodes deployed at various access control gates throughout the park, and a private cloud service platform connected to each edge node. The edge nodes utilize embedded computing devices equipped with high-definition face capture cameras to meet the real-time face recognition requirements of the access control gates. The cloud service platform is deployed on the park's private cloud server, providing global face feature management, high-precision recognition, and cloud-edge collaborative scheduling functions.

[0053] In this embodiment, the preset parameters are configured as follows: In the multi-task joint loss function of lightweight feature extraction networks, The value is 0.05. The value is set to 0.1. The model is trained based on a public face dataset and a face dataset of people inside the park. After training, the model is deployed on edge nodes. The initial face feature dimension is 512.

[0054] In the formula for calculating the optimal feature compression ratio, The value is 0.2. The value is 1.0. The value is 0.4. The value is 0.3. The value is 0.3, which satisfies the condition. .

[0055] In the formula for calculating the decision threshold, The value is 0.75. The value is 0.1. The value is 0.15. The value is 0.05.

[0056] The local facial feature database stores the facial features of employees within the park, achieving a target population coverage rate. The value is 0.95. The global facial feature database stores the facial features of all authorized personnel in the park, including full-authorization data of current employees, temporary visitors, and property staff.

[0057] The face recognition process in this embodiment is as follows: The first step involves the edge node acquiring real-time face images in front of the gate via a face acquisition and preprocessing module. This module performs face detection, key point localization, face alignment, and normalization, outputting a standardized face image of size 112×112. The quality score of this standardized face image is then calculated. .

[0058] The second step involves edge nodes collecting current device operating status data to obtain the current computing load rate. Current network bandwidth availability Substituting into the formula for calculating the optimal feature compression ratio, the calculation process is as follows: ; ; ; ; Based on the calculated optimal feature compression ratio The initial 512-dimensional facial features were quantized and compressed to generate 428-dimensional edge-compressed facial features, and the confidence score of these edge-compressed facial features was calculated. .

[0059] The third step involves the edge node performing cosine similarity matching between the compressed facial features and the local facial feature database, resulting in a maximum similarity score of 0.88, which corresponds to the matching item Zhang San, an employee of the park.

[0060] The fourth step is to collect environmental interference factors in the current scene. Combined with the target population coverage of the local facial feature database Confidence score of edge-compressed facial features Substituting into the decision threshold calculation formula, the calculation process is as follows: ; Compare the highest similarity score of 0.88 with the decision threshold. If the highest similarity score is greater than the decision threshold, it is determined that the matching result is successful and will not be uploaded to the cloud. The face recognition is then completed and the gate will open.

[0061] Fifth, the cloud service platform collects real-time operational status data from each edge node, generating an incremental feature package every 24 hours to synchronously update the local face feature database of each edge node. Simultaneously, based on the high-confidence matching features uploaded by the edge nodes, iteratively optimizes the corresponding feature data in the global face feature database, improving the accuracy of subsequent recognition.

[0062] Example 2: Facial Recognition Scenario at Urban Road Traffic Checkpoints This embodiment is applied to a facial recognition scenario at urban road traffic checkpoints. The system includes 20 edge nodes deployed at checkpoints on main and secondary urban roads, and a cloud service platform for urban traffic connected to each edge node via a 5G communication network. The edge nodes are integrated into the intelligent capture cameras at the checkpoints, possessing low power consumption and wide temperature range operation capabilities, adapting to the facial capture and recognition needs in complex outdoor environments. The cloud service platform is deployed on an urban traffic cloud server cluster, supporting full-scale matching, high-precision recognition, and cross-node collaborative scheduling functions for tens of millions of facial features.

[0063] In this embodiment, the preset parameters are configured as follows: In the multi-task joint loss function of lightweight feature extraction networks, The value is 0.08. The value is 0.12. The model is trained based on a public outdoor face dataset and an urban checkpoint face capture dataset. After training, the model is deployed on edge nodes. The initial face feature dimension is 512.

[0064] In the formula for calculating the optimal feature compression ratio, The value is 0.1. The value is 0.8. The value is 0.35. The value is 0.45. The value is 0.2, which satisfies... .

[0065] In the formula for calculating the decision threshold, The value is 0.7. The value is 0.15. The value is 0.2. The value is 0.08.

[0066] The local facial feature database stores the facial features of key personnel under control, with a target population coverage rate. The value is set to 0.3. The global facial feature database stores the facial features of all personnel under the full control authorized by the public security system, with a data volume reaching tens of millions.

[0067] The face recognition process in this embodiment is as follows: The first step involves the edge node capturing facial images of occupants inside passing vehicles using a face acquisition and preprocessing module. This process includes face detection, key point localization, face alignment, and normalization, resulting in a standardized face image. The quality score of this standardized face image is then calculated. .

[0068] The second step involves edge nodes collecting current device operating status data to obtain the current computing load rate. Current 5G network bandwidth availability Substituting into the formula for calculating the optimal feature compression ratio, the calculation process is as follows: ; Based on the calculated optimal feature compression ratio The initial 512-dimensional facial features were quantized and compressed to generate 161-dimensional edge-compressed facial features, and the confidence score of these edge-compressed facial features was calculated. .

[0069] The third step involves the edge node performing cosine similarity matching between the compressed facial features and the local facial feature database, resulting in a maximum similarity score of 0.68, which corresponds to the matching item "management personnel Li Si".

[0070] The fourth step is to collect environmental interference factors in the current outdoor scene. Combined with the target population coverage of the local facial feature database Confidence score of edge-compressed facial features Substituting into the decision threshold calculation formula, the calculation process is as follows: ; Compare the highest similarity score of 0.68 with the decision threshold. If the highest similarity score is less than the decision threshold, it is determined that the image should be uploaded to the cloud, and the standardized face image will be uploaded to the city traffic cloud platform cloud service platform via 5G network.

[0071] The fifth step involves the cloud service platform receiving standardized facial images and generating 2048-dimensional high-precision facial features through a high-precision feature extraction module. These high-precision facial features are then matched against the global facial feature database using a full vector search, yielding a maximum similarity score of 0.93. The match result is identified as the control personnel, Li Si. The global matching result is then sent to the corresponding edge nodes as the final identification result and simultaneously synchronized with the public security control system.

[0072] The sixth step involves the cloud service platform collecting real-time network status and computing load data from each edge node, dynamically adjusting scheduling parameters, and distributing and synchronizing incremental feature packets every hour to ensure the real-time availability of the local face feature database on each edge node. Simultaneously, based on matching data uploaded from each checkpoint, cross-node face trajectory tracking and global feature optimization are performed.

[0073] Example 3: Facial Recognition Scenario for Passenger Flow in Commercial Complexes This embodiment is applied to a facial recognition scenario for customer flow in a large commercial complex. The system includes 35 edge nodes deployed at various entrances and exits, floor passages, and shop entrances of the commercial complex, as well as a cloud service platform for commercial operations that communicates with each edge node. The edge nodes use intelligent customer flow analysis cameras, adapted to the needs of facial capture and recognition under high traffic, multiple poses, and complex lighting conditions. The cloud service platform is deployed on a commercial operations cloud server, supporting functions such as member facial feature management, customer flow statistics, precise marketing matching, and dynamic scheduling of cloud and edge resources.

[0074] In this embodiment, the preset parameters are configured as follows: In the multi-task joint loss function of lightweight feature extraction networks, The value is 0.06. The value is 0.09. The model is trained based on a publicly available commercial face dataset and a complex member face dataset. After training, the model is deployed on edge nodes. The initial face feature dimension is 512.

[0075] In the formula for calculating the optimal feature compression ratio, The value is 0.15. The value is 0.9. The value is 0.5. The value is 0.25. The value is 0.25, which satisfies the condition. .

[0076] In the formula for calculating the decision threshold, The value is 0.72. The value is 0.12. The value is 0.18. The value is 0.06.

[0077] The local facial feature database stores the facial features of high-level members of the complex, achieving a high coverage rate among the target population. The value is set to 0.6. The global facial feature database stores the facial features of all registered members of the complex, and also associates them with members' consumption preferences and membership level information.

[0078] The face recognition process in this embodiment is as follows: The first step involves the edge node acquiring facial images of passengers in the floor corridor through the face acquisition and preprocessing module. This process includes face detection, key point localization, face alignment, and normalization, resulting in a standardized facial image. The quality score of this standardized facial image is then calculated. .

[0079] The second step involves edge nodes collecting current device operating status data to obtain the computing load rate during peak passenger flow periods. Current network bandwidth availability Substituting into the formula for calculating the optimal feature compression ratio, the calculation process is as follows: ; Based on the calculated optimal feature compression ratio The initial 512-dimensional facial features were quantized and compressed to generate 238-dimensional edge-compressed facial features, and the confidence score of these edge-compressed facial features was calculated. .

[0080] The third step involves the edge node performing cosine similarity matching between the compressed face features and the local face feature database, resulting in a maximum similarity score of 0.76, which corresponds to the match for the member Wang Wu.

[0081] The fourth step is to collect environmental interference factors in the current business scenario. Combined with the target population coverage of the local facial feature database Confidence score of edge-compressed facial features Substituting into the decision threshold calculation formula, the calculation process is as follows: ; Compare the highest similarity score of 0.76 with the decision threshold. If the highest similarity score is less than the decision threshold, it is determined that the image should be uploaded to the cloud, and the standardized face image is uploaded to the commercial operation cloud platform.

[0082] The fifth step involves the cloud service platform receiving standardized facial images and generating 1024-dimensional high-precision facial features through a high-precision feature extraction module. These high-precision facial features are then fully matched against the global facial feature database, yielding a maximum similarity score of 0.91. The match result is identified as Wang Wu, a member of the complex. Simultaneously, the platform retrieves the member's level and consumption preference information and sends the global matching result along with the member information to the corresponding edge nodes as the final recognition result. The edge nodes then simultaneously push the corresponding marketing information to the on-site display screen.

[0083] The sixth step involves the cloud service platform dynamically adjusting scheduling parameters based on peak passenger flow and computing load changes at each edge node. During off-peak hours, full feature synchronization is completed, while during peak hours, only core incremental feature packages are distributed, balancing edge recognition efficiency with cloud resource consumption. Simultaneously, based on the full recognition data, passenger flow heatmap analysis and member profile iteration optimization are performed to improve business operation efficiency.

[0084] This invention addresses the shortcomings of existing cloud-edge collaborative face recognition technologies, where fixed feature extraction and compression modes at the edge end cannot adapt to dynamic operating environments and struggle to balance recognition efficiency and matching accuracy. By employing a lightweight adaptive feature extraction module at the edge end, it constructs a dynamic calculation model for the optimal feature compression ratio, integrating edge computing load, network bandwidth availability, and face image quality. This model dynamically adjusts the feature compression ratio based on the real-time operating status of the edge end and the quality of the input image. Under high computing load and weak network conditions, it reduces feature dimensionality and data transmission volume; under low load and good network conditions, it maintains feature discriminative power, achieving a dynamic balance between recognition efficiency and matching accuracy. Furthermore, to address the accuracy loss caused by edge feature compression, this invention designs a multi-task joint loss function that integrates cross-entropy classification loss, center loss, and quantization loss. During the training of the lightweight model, it simultaneously constrains feature classification accuracy, intra-class aggregation, and feature consistency before and after compression, ensuring that compressed edge face features still maintain high discriminative power. This effectively improves the accuracy of local matching, reduces unnecessary secondary recognition calls in the cloud, fundamentally reduces the computational pressure and network bandwidth consumption of the cloud service platform, and improves the overall real-time responsiveness of the system.

[0085] This invention addresses the shortcomings of existing technologies, such as insufficient robustness of fixed matching decision thresholds at the edge and weak cloud-edge collaborative scheduling capabilities. By constructing a dynamic decision threshold calculation model that integrates scene environment interference factors, target population coverage of the local face feature library, and confidence of edge-compressed face features through a dynamic threshold judgment unit set at the edge, this invention achieves dynamic adaptation of local matching decision logic to application scenarios and system states. It solves the technical problems of increased false recognition rate in complex interference scenarios and increased missed recognition rate in low coverage scenarios with fixed thresholds, effectively improving the robustness and accuracy of face recognition in different application scenarios. Meanwhile, this invention, through a cloud-edge collaborative scheduling module set up in the cloud, can collect real-time operational status data of each edge node and issue differentiated scheduling parameters, realizing dynamic optimization and configuration of cloud-edge computing and storage resources. Through the incremental synchronization unit of the feature update management module, it can achieve accurate incremental synchronization between the global face feature library and the local face feature library at the edge. At the same time, based on the high-confidence face features uploaded from the edge, it iteratively optimizes the global face feature library, constructing a closed-loop mechanism for continuous optimization of system recognition accuracy. This solves the problems of insufficient cloud-edge collaboration depth and inability to continuously iterate and optimize system performance in existing technologies, and improves the long-term operational stability and recognition accuracy of the entire system.

[0086] This invention constructs a modular and scalable cloud-edge collaborative face recognition system and method, breaking through the limitations of existing technologies that are only applicable to a single rail transit gate scenario. Through differentiated configuration of scheduling parameters, it can flexibly adapt to various distributed face recognition scenarios such as smart parks, urban traffic checkpoints, and commercial complexes. It is compatible with edge nodes of different computing power levels and network conditions, possessing strong scenario adaptability and system scalability. This invention rationally offloads the core computing tasks of face recognition to the edge and optimizes edge computing efficiency through a dynamic adaptive mechanism. Without increasing edge hardware costs, it achieves high real-time performance and high accuracy in distributed face recognition, while significantly reducing the hardware investment and operating costs of the cloud service platform, demonstrating significant industrial applicability. The technical solution of this invention fills the technical gaps in existing cloud-edge collaborative face recognition technologies in dynamic adaptive optimization, end-to-end accuracy-efficiency balance control, and closed-loop accuracy iterative optimization, providing a complete, stable, and efficient technical solution for face recognition applications in large-scale, multi-node, and complex scenarios.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A face recognition system based on cloud-edge collaboration, characterized in that, It includes at least one edge node and a cloud service platform that communicates with each edge node. The edge node includes a face acquisition and preprocessing module, a lightweight adaptive feature extraction module, a local face feature library, an edge matching module, and an edge collaborative interaction module. The face acquisition and preprocessing module is used to acquire real-time face images and perform preprocessing to output standardized face images. The lightweight adaptive feature extraction module is used to extract features and adaptively compress the standardized face images to output edge-compressed face features. The edge matching module is used to match the edge-compressed facial features with the features in the local facial feature database and output the local matching result; the edge collaborative interaction module is used to complete the feature data interaction and command transmission and reception with the cloud service platform based on the local matching result. The cloud service platform includes a global face feature database, a high-precision feature extraction module, a cloud-edge collaborative scheduling module, a global matching module, and a feature update management module. The high-precision feature extraction module receives standardized face images uploaded by edge nodes and outputs high-precision face features. The global matching module performs full matching between the high-precision face features and the global face feature database and outputs the global matching result. The cloud-edge collaborative scheduling module issues scheduling parameters for feature extraction and matching based on the running status data of the edge nodes. The feature update management module performs incremental synchronous updates of the global face feature database and the local face feature databases of each edge node.

2. The cloud-edge collaborative face recognition system according to claim 1, characterized in that, The lightweight adaptive feature extraction module incorporates an adaptive feature compression unit. This unit calculates the optimal feature compression ratio α based on the real-time operating status of the edge nodes and the quality of the input face image. The calculation formula is as follows: ； In the formula, To preset the minimum compression ratio, To preset the maximum compression ratio, This represents the current computing load rate of the edge nodes. This represents the current network bandwidth availability of the edge nodes. To standardize the quality score of face images, These are preset weight coefficients corresponding to computing power load, bandwidth availability, and image quality, respectively; the adaptive feature compression unit quantizes and compresses the extracted initial face features according to the calculated optimal feature compression ratio α to generate edge-compressed face features.

3. The cloud-edge collaborative face recognition system according to claim 1, characterized in that, The edge matching module has a built-in dynamic threshold judgment unit. This dynamic threshold judgment unit is used to calculate the decision threshold T for local matching based on the scheduling parameters sent from the cloud. The calculation formula is as follows: ； In the formula, As a preset baseline threshold, For environmental interference factors in the current scene, To determine the target population coverage of the local facial feature database, The confidence score for edge-compressed facial features. These are preset adjustment coefficients corresponding to environmental interference, coverage, and feature confidence, respectively. The dynamic threshold judgment unit compares the highest similarity score of the local match with the decision threshold T. If the highest similarity score is greater than or equal to T, it outputs the local match result that is successfully matched. If the highest similarity score is less than T, it triggers the edge collaborative interaction module to upload a standardized face image to the cloud service platform.

4. The cloud-edge collaborative face recognition system according to claim 1, characterized in that, The lightweight adaptive feature extraction module uses an improved multi-task joint loss function to complete model training. The expression of the multi-task joint loss function L is as follows: ； In the formula, Cross-entropy classification loss, The center loss is used to reduce the intra-class feature distance. The quantization loss is used to constrain the distance between the compressed features and the original features. These are preset balance weights for center loss and quantization loss, respectively; the lightweight adaptive feature extraction module extracts and compresses facial features based on the trained model.

5. The cloud-edge collaborative face recognition system according to claim 1, characterized in that, The feature update management module has a built-in incremental synchronization unit, which is used to count the newly added, modified and deleted feature data in the global face feature database and generate incremental feature packages. The incremental synchronization unit splits the incremental feature package into data segments based on the local face feature library version number and network status of each edge node, and sends them to the corresponding edge nodes at a preset period to complete the incremental update of the local face feature library. The incremental synchronization unit is also used to receive locally matched high-confidence face features uploaded by edge nodes and iteratively optimize the corresponding features in the global face feature library.

6. A face recognition method based on cloud-edge collaboration, characterized in that, The face recognition system based on cloud-edge collaboration as described in any one of claims 1-5 includes the following steps: S1. Edge nodes acquire real-time face images, and output standardized face images after preprocessing; S2. Edge nodes perform feature extraction and adaptive compression on standardized face images to generate edge-compressed face features; S3. The edge node performs local matching between the compressed face features and the local face feature database to generate a local matching result; S4. The edge node determines whether to upload the standardized face image to the cloud service platform based on the local matching result and the scheduling parameters issued by the cloud. If the determination is yes, proceed to step S5; if the determination is no, directly output the local matching result as the final recognition result. S5. The cloud service platform receives standardized face images, generates high-precision face features through the high-precision feature extraction module, performs full matching of the high-precision face features with the global face feature library, generates global matching results, and sends them to the corresponding edge nodes as the final recognition results. S6. The cloud service platform updates and distributes scheduling parameters based on the operating status data of each edge node, and simultaneously completes incremental synchronous updates of the global face feature database and the local face feature database of each edge node.

7. The face recognition method based on cloud-edge collaboration according to claim 6, characterized in that, In step S2, the feature extraction and adaptive compression of the standardized face image specifically includes: S21. Based on the trained lightweight feature extraction model, perform feature extraction on the standardized face image to generate initial face features; S22. Collect the current computing power load rate ρ and the current network bandwidth availability β of the edge node, calculate the quality score Q of the standardized face image, and substitute it into the optimal feature compression ratio calculation formula to obtain the optimal feature compression ratio α; S23. Based on the optimal feature compression ratio α, the initial face features are quantized and compressed to generate edge-compressed face features; The formula for calculating the optimal feature compression ratio is: ； In the formula, To preset the minimum compression ratio, To preset the maximum compression ratio, These are preset weighting coefficients corresponding to computing load, bandwidth availability, and image quality, respectively.

8. The face recognition method based on cloud-edge collaboration according to claim 6, characterized in that, In step S4, determining whether to upload the standardized face image to the cloud service platform based on the local matching results and the scheduling parameters issued by the cloud specifically includes: S41. Extract the highest similarity score from the local matching results, and calculate the decision threshold T for local matching by combining it with the scheduling parameters issued by the cloud. S42. If the highest similarity score is greater than or equal to the decision threshold T, it is determined that no upload is needed, and the local matching result is directly output as the final recognition result. S43. If the highest similarity score is less than the decision threshold T, it is determined to be an upload, and the standardized face image is uploaded to the cloud service platform; The formula for calculating the decision threshold T is: ； In the formula, As a preset baseline threshold, For environmental interference factors in the current scene, To determine the target population coverage of the local facial feature database, The confidence score for edge-compressed facial features. These are the preset adjustment coefficients corresponding to environmental interference, coverage, and feature confidence, respectively.

9. The face recognition method based on cloud-edge collaboration according to claim 6, characterized in that, In step S2, the lightweight feature extraction model is trained using an improved multi-task joint loss function, the expression of which is: ； In the formula, Cross-entropy classification loss, Loss at the center To quantify the loss, These are the preset balance weights for center loss and quantization loss, respectively; During training, the model parameters are iteratively updated by minimizing the multi-task joint loss function L.

10. The face recognition method based on cloud-edge collaboration according to claim 6, characterized in that, In step S6, the incremental synchronization update of the global face feature database and the local face feature database of each edge node specifically includes: S61. The cloud service platform counts the newly added, modified, and deleted feature data in the global face feature database and generates incremental feature packages with corresponding version numbers; S62. Collect the local face feature library version number and real-time network status of each edge node, split the incremental feature package into data segments, and send them to the corresponding edge nodes according to a preset cycle; S63. Edge nodes receive fragmented data and complete reassembly, update the local face feature database and version number, and upload the high-confidence face features that are successfully matched locally to the cloud service platform. S64. The cloud service platform receives high-confidence facial features and iteratively optimizes the corresponding features in the global facial feature database.

Images