A one-face management method and system for localization of information technology innovation

Abstract

The present invention discloses a one-face pass management method and system for the localization of information technology innovation, which relates to the field of one-face pass, including collecting user facial images and extracting facial feature vectors; calculating the identity differentiation ability of each dimension based on the feature vector set of multiple users, and constructing a difference heat map; according to the user's pass frequency, mapping the pass frequency to its feature compression dimension number through a preset mapping function; selecting several dimensions ranked before the difference heat score as compression dimension indexes, extracting corresponding eigenvalues ​​from the facial feature vector to form a compressed feature vector, and storing it locally as a user compression template; in the authentication process, using the compressed feature vector for similarity comparison, if it fails, switching to full-dimensional comparison for compensatory recognition. On the basis of maintaining the recognition accuracy, the present invention effectively reduces the system computing load and improves the authentication efficiency, and is particularly suitable for scenarios such as localized parks with obvious differences in the distribution of pass frequencies.

Description

Technical Field

[0001] The present invention relates to the field of one-face access, and more specifically, to a one-face access management method and system for the localization of information technology innovation. Background Art

[0002] With the widespread application of facial recognition technology in security, office, and campus management scenarios, facial recognition-based face-to-face management systems have gradually become a mainstream alternative to traditional card swiping and fingerprint authentication. However, existing facial recognition systems mostly rely on high-performance chips and specialized algorithms, placing high demands on computing resources, storage space, and response latency during facial feature extraction, comparison, and management.

[0003] Especially in the localization scenario of ICT (information technology application innovation), for the sake of security, controllability and compatibility with the ICT ecosystem, the system usually adopts domestic chips, self-developed models and local deployment architecture, facing practical challenges such as limited edge computing capabilities, tight computing resources, and low model deployment efficiency. There is an urgent need to design a facial authentication mechanism that takes into account both efficiency and security.

[0004] On the other hand, in real-world scenarios, usage frequency varies significantly among different users. For example, resident employees may use the system multiple times a day, while visitors or external contractors may only use it once every few days or even weeks. Using a comprehensive feature set for all users would not only burden the system but could also cause response delays and even identification congestion when dealing with large numbers of users.

[0005] Therefore, how to dynamically compress the recognition dimension based on the user's frequency of travel, thereby significantly improving recognition efficiency and system response speed while maintaining recognition accuracy, has become a pressing issue for current facial recognition systems. Furthermore, rapidly authenticating high-frequency passersby through compressed feature dimensions also helps reduce the attack window and improve the system's overall security and anti-counterfeiting capabilities.

[0006] To sum up, the existing technology still lacks a facial recognition system that is oriented towards the domestic environment of information technology innovation and can adaptively adjust the recognition resource allocation strategy based on the frequency of traffic, while taking into account efficiency, storage and security. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a face-to-face management method and system for the localization of information technology, so as to solve the problems mentioned in the background technology.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A face-to-face management method for localization of information technology innovation includes the following steps:

[0010] Collecting a user's face image and extracting a face feature vector using a face recognition model. The face feature vector is a multidimensional vector of length N.

[0011] Based on a set of facial feature vectors of multiple users, the identity differentiation ability of each feature dimension in the face recognition task is calculated, and a difference heat map is constructed. The difference heat map is a one-dimensional vector of length N, where each element represents the difference heat score of the corresponding feature dimension in the facial feature vector;

[0012] Count the travel frequency P of each user within a preset time window, and map the user's travel frequency P to its corresponding feature compression dimension K according to the preset mapping function f(P), where K < N, and K decreases as P increases;

[0013] Selecting the top K feature dimensions ranked by difference heat scores from the difference heat map as the compressed dimension index set for the user; using the compressed dimension index set, extracting feature values ​​of corresponding dimensions from the facial feature vector to form a compressed feature vector, and storing it locally as a compressed template for the user;

[0014] During one-face authentication, a compressed feature vector is extracted based on the frequency P1 of the currently recognized facial feature vector appearing in the historical range, and a similarity comparison is performed with all locally stored compressed templates with corresponding frequencies within a certain range before and after P1. If there is a compressed template whose similarity meets the passing conditions, the authentication is completed, otherwise it switches to full-dimensional comparison for compensation matching.

[0015] In some embodiments, the expression of f(P) is: f(P) = max(K min , K max − α·P), where α is the compression adjustment coefficient greater than 0, K min is the minimum number of compressed dimensions, K max is the maximum number of compressed dimensions.

[0016] In some embodiments, constructing a difference heat map includes:

[0017] The facial feature vectors of multiple users are grouped according to their identity labels, and the mean difference and intra-group variance of each feature dimension between different users are counted.

[0018] For the feature dimensions, and calculate their identity discrimination ability scores in face recognition tasks ;

[0019] For all feature dimensions The values ​​of are normalized to form a one-dimensional vector of length N as the difference heat map, where each element represents the difference heat score of the corresponding feature dimension.

[0020] In some embodiments, wherein It is calculated using the ratio of between-class variance to within-class variance, which is defined as follows:

[0021]

[0022] in:

[0023] Indicates the The inter-class variance of dimensional features among different users; b Represents between classes;

[0024] Indicates the The intra-class variance of the dimension feature within the same user; w Indicates within a class.

[0025] In some embodiments, the similarity comparison includes:

[0026] Calculate the Euclidean distance or cosine similarity between the extracted compressed feature vector and the local compressed template;

[0027] Determine whether the similarity meets the preset compression comparison threshold. If so, the authentication is passed; otherwise, switch to full-dimensional comparison;

[0028] In full-dimensional comparison, the complete facial feature vector is used for matching to determine whether it meets the preset full-dimensional comparison threshold.

[0029] In some embodiments, the statistical method of the user traffic frequency P includes:

[0030] In the sliding time window W, record the number of times the user passes n;

[0031] The frequency of passage is defined as P = n / W, where W is in hours, days or sub-time units.

[0032] In some embodiments, the local storage method includes binding each user's compressed template to their identity ID and using an encryption algorithm to store and protect the template data.

[0033] The present invention also discloses a system for implementing the above method, comprising:

[0034] A facial feature extraction module is used to collect a user's facial image and extract a facial feature vector using a facial recognition model. The facial feature vector is a multidimensional vector of length N.

[0035] A difference heat analysis module is used to calculate the identity differentiation ability of each feature dimension in the face recognition task based on a set of facial feature vectors of multiple users, and to construct a difference heat map. The difference heat map is a one-dimensional vector of length N, where each element represents the difference heat score of the corresponding feature dimension in the facial feature vector;

[0036] The traffic frequency statistics module is used to count the traffic frequency P of each user within a preset time window.

[0037] A compression dimension mapping module is used to map the user's travel frequency P to its corresponding feature compression dimension K according to a preset mapping function f(P), where K < N, and K decreases as P increases;

[0038] A compression template generation module is used to select the top K feature dimensions ranked by difference heat scores from the difference heat map as the user's compression dimension index set, and use the compression dimension index set to extract the feature values ​​of the corresponding dimensions from the facial feature vector to form a compressed feature vector, which is locally stored as the user's compression template;

[0039] The one-face authentication module is used to extract the compressed feature vector based on the frequency P1 of the currently recognized facial feature vector appearing in the historical range during one-face authentication, and perform similarity comparison with all locally stored compressed templates with corresponding frequencies within a certain range before and after P1. If there is a compressed template whose similarity meets the passing conditions, the authentication is completed; otherwise, it switches to full-dimensional comparison for compensation matching.

[0040] In some embodiments, the expression of f(P) is: f(P) = max(K min , K max − α·P), where α is the compression adjustment coefficient greater than 0, K min is the minimum number of compressed dimensions, K max is the maximum number of compressed dimensions.

[0041] In some embodiments, the local storage method includes binding each user's compressed template to his / her identity ID and using an encryption algorithm to store and protect the template data.

[0042] This invention combines the ability to distinguish user traffic frequency and facial features to implement adaptive compression of recognition dimensions and a hierarchical comparison authentication mechanism. While maintaining recognition accuracy, it significantly improves the operating efficiency of the face-to-face system on domestic chips and low-computing power platforms, with the following beneficial effects:

[0043] The present invention dynamically adjusts the compression dimension of the feature vector according to the frequency of each user's passage. High-frequency users only need to extract a small number of highly discriminative feature dimensions to complete rapid authentication, significantly reducing the computational overhead of feature comparison. It is particularly suitable for the information technology innovation environment where the computing power of domestic chips is limited.

[0044] The present invention introduces a difference heat map to score the identity discrimination ability of each feature dimension, and only selects the most discriminative dimensions for compression, so as to retain the discrimination information as much as possible while reducing the dimension; when the compression comparison cannot meet the recognition conditions, it automatically switches to full-dimensional compensation matching, effectively improving the robustness and security of the overall recognition.

[0045] The mapping function enables the adaptive response of the compressed dimension to user traffic behavior, allowing the system to flexibly allocate computing resources according to user activity, effectively alleviating the system pressure caused by the expansion of the user scale.

[0046] The method of the present invention is optimized in terms of compressed template generation and local storage, supports local comparison and local update on low-power devices, reduces dependence on the cloud, and improves the system's autonomous controllability and edge deployment capabilities. It is suitable for scenarios such as government and enterprise parks, industrial control factories, and confidential institutions that have strict requirements for information technology compliance.

[0047] The present invention calculates the feature discrimination ability through the inter-class variance and the intra-class variance, and introduces the compression adjustment coefficient for dynamic control. It can flexibly adjust the compression strategy according to different usage environments, is compatible with face models of different scales and recognition requirements, and has good practicality and engineering feasibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the overall flow chart of the present invention;

[0049] Figure 2 It is a flow chart of constructing a difference heat map of the present invention. DETAILED DESCRIPTION

[0050] The specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0051] like Figure 1 As shown, the present invention includes the following steps:

[0052] Collecting a user's face image and extracting a face feature vector using a face recognition model. The face feature vector is a multidimensional vector of length N.

[0053] Based on a set of facial feature vectors of multiple users, the identity differentiation ability of each feature dimension in the face recognition task is calculated, and a difference heat map is constructed. The difference heat map is a one-dimensional vector of length N, where each element represents the difference heat score of the corresponding feature dimension in the facial feature vector;

[0054] Count the travel frequency P of each user within a preset time window, and map the user's travel frequency P to its corresponding feature compression dimension K according to the preset mapping function f(P), where K < N, and K decreases as P increases;

[0055] Selecting the top K feature dimensions ranked by difference heat scores from the difference heat map as the compressed dimension index set for the user; using the compressed dimension index set, extracting feature values ​​of corresponding dimensions from the facial feature vector to form a compressed feature vector, and storing it locally as a compressed template for the user;

[0056] During one-face authentication, a compressed feature vector is extracted based on the frequency P1 of the currently recognized facial feature vector appearing in the historical range, and a similarity comparison is performed with all locally stored compressed templates with corresponding frequencies within a certain range before and after P1. If there is a compressed template whose similarity meets the passing conditions, the authentication is completed, otherwise it switches to full-dimensional comparison for compensation matching.

[0057] More specifically, facial images are captured using a camera or other image acquisition device. After capture, the image is processed using a trained face recognition model to extract a facial feature vector. This feature vector is a multidimensional vector of length N. The specific value of N typically depends on the model design and can be common values ​​such as 128, 256, or 512.

[0058] Facial recognition models are typically based on deep learning techniques, such as convolutional neural networks. Common models include FaceNet and ArcFace. These models are trained on large-scale face datasets (such as MS-Celeb-1M or VGGFace2) to learn how to map facial images into a high-dimensional feature space.

[0059] In this space, feature vectors for the same user are kept as close together as possible, while feature vectors for different users are kept as far apart as possible. The training process involves data preprocessing (such as image cropping, alignment, and normalization), model architecture design (such as multiple convolutional and fully connected layers), loss function selection (such as triplet loss or softmax loss variants), and optimization algorithm application (such as Adam or stochastic gradient descent).

[0060] Through such training, the model is able to generate feature vectors with high discrimination ability, laying the foundation for subsequent steps.

[0061] After obtaining multiple sets of facial feature vectors, the next step is to evaluate the role of each feature dimension in distinguishing different user identities. This step is achieved by constructing a difference heat map. The difference heat map is a one-dimensional vector of length N, where each element represents the difference heat score of the corresponding feature dimension, reflecting the importance of the dimension in identity recognition. Figure 2 As shown, the specific construction process is as follows:

[0062] First, group the facial feature vectors of multiple users according to their identity labels. For example, group all feature vectors of user A into one group, and those of user B into another. Then, for each feature dimension i, calculate the mean difference and within-group variance across different users. The mean difference can be calculated by taking the difference between the within-group mean and the overall mean for each user, while the within-group variance is the degree of fluctuation in feature values ​​collected multiple times for the same user along that dimension.

[0063] Based on these statistics, the identity discrimination ability score of each characteristic dimension is calculated , defined as the between-class variance and intra-class variance The ratio of .

[0064] Here, the between-class variance Indicates the The degree of dispersion of dimensional features between different users, intra-class variance Indicates the stability within the same user. A higher value means it is more useful in distinguishing different users, which is why this formula is used.

[0065] In order to make the difference heat map more intuitive, all The values ​​need to be normalized, for example, by mapping the scores to the [0,1] interval through min-max normalization, or by z-score normalization so that the scores have zero mean and unit variance, and finally obtain the difference heat map.

[0066] In the practical application of One-Face Pass management, user access frequency is a crucial factor, as it directly impacts authentication efficiency and computing resource utilization. To dynamically adapt to the needs of different users, the system counts each user's access frequency P within a preset time window. Specifically, the time window W can be set to 1 day, 1 week, or even longer, depending on the application scenario, and can be in units of hours or days. Within this time window, the number of times a user accesses n is recorded, and the access frequency P is calculated as n / W. For example, if W is 24 hours and a user accesses One-Face Pass 12 times a day, then P = 12 / 24 = 0.5 times / hour.

[0067] Then, through a mapping function f(P) = max(K min , K max - α·P) converts P into feature compression dimension K, where K is the number of feature dimensions to be retained and K is less than N.

[0068] The reason for designing this function is that when P is large (high-frequency users), K max - α·P will become smaller and eventually approach K min , which means retaining fewer feature dimensions to speed up authentication; when P is small (low-frequency users), K tends to K max , retain more dimensions to improve accuracy. Parameter K min , K max The choice of K and α needs to be adjusted according to actual needs, such as min It can be between 20 and 50, K max It can be between half of N and N (for example, when N = 256, Kmax = 128). α is a compression adjustment coefficient greater than 0, and its typical value may be 1 to 10.

[0069] For example, suppose K min = 20, K max = 100, α = 2, and a user P = 40 times / day, then K = max(20, 100 – 2×40) = max(20, 20) = 20;

[0070] For another user, P = 5 times / day, then K = max(20, 100 - 2×5) = max(20, 90) = 90. In this way, high-frequency users only need to compare a small number of dimensions, while low-frequency users retain more dimensions.

[0071] After determining K, the system selects the top K feature dimensions by difference heat score from the difference heat map to form a compressed dimension index set for the user. These dimensions are the most distinguishing features and are able to retain as much key information as possible after compression. Using this index set, the feature values ​​of the corresponding dimensions are extracted from the original facial feature vector to form a compressed feature vector of length K. This compressed feature vector serves as the user's compression template and is stored locally on the device. To ensure data security, the compression template is bound to the user's identity ID and protected using an encryption algorithm (such as AES-256 or SHA-256) to prevent unauthorized access or tampering.

[0072] During the face authentication process, the system extracts the facial feature vector based on the currently collected facial image, and counts the frequency of appearance of the facial feature vector within a preset time range or a historical range or a past period of time, which is recorded as the pass frequency P1. The statistical process of the pass frequency P1 here does not need to know which user the facial feature vector belongs to. In other words, when the system recognizes the facial feature vector, it does not need to know which user the facial feature vector belongs to. It only needs to know the frequency P1 of the facial feature vector appearing in the past historical range or the past period of time, so that the corresponding compression method can be selected. The preset time range can be the past three days or the past week, and the specific length is selected according to actual needs.

[0073] After obtaining the passing frequency P1 of the current facial feature vector, the system selects the compression dimension according to the frequency P1 and compresses the current facial feature vector into a compressed feature vector.

[0074] Then, the system selects all compressed templates with corresponding traffic frequencies within a certain range before and after P1 from the local storage for similarity comparison. The range can be set to ,in Indicates the allowable frequency fluctuation tolerance. In some embodiments, the range can be selected to be slightly larger, such as 3 to 8 times / week, to ensure that a sufficiently large compressed template area can be covered. If the similarity of a certain compressed template meets the pass condition, the authentication is determined to be successful; otherwise, the system automatically switches to the full-dimensional feature vector for compensation comparison to ensure accuracy. Since the compressed feature vector and the compressed template are used for comparison, even if all the compressed templates are compared, the efficiency can be greatly improved due to the reduction in dimension. Through the above method, the present invention realizes the completion of compression strategy matching based on the frequency of appearance of the facial feature vector itself in the historical records without the need for pre-knowledge of the identity, avoiding the logical paradox of "undetermined identity leading to the unavailability of the compressed template". At the same time, since multiple compressed templates within the frequency range are involved in the comparison, there is no need to know in advance which compressed template to select, which also avoids the occurrence of logical paradoxes.

[0075] In a preferred embodiment, a clustering strategy can be used to calculate the frequency P1 of the current facial feature vector. Facial feature vectors that have appeared within a historical range or a past period of time are clustered into the same category and considered to belong to the same person, without knowing the person's specific ID or whether they meet the passing criteria. This clustering does not rely on user identity tags, but instead organizes them based on the structural characteristics of the facial feature vectors themselves. The clustering process has lower computational complexity than the complete face recognition process because it does not involve matching each facial feature vector with all identity templates in the user database. It only categorizes and organizes facial feature vectors within the local range where the current facial feature vector is located.

[0076] Similarity can be calculated using Euclidean distance or cosine similarity. If the calculated similarity S is greater than the preset compression comparison threshold Tc (for example, Tc can be set to 0.9), authentication is considered successful and the user can proceed directly. If S is less than Tc, the system switches to full-dimensional comparison mode, using the complete N-dimensional facial feature vector to match the locally stored complete template (if available), calculate the full-dimensional similarity Sf, and determine whether it exceeds the full-dimensional comparison threshold Sf (for example, Sf can be set to 0.85). If Sf meets the requirements, authentication is successful; otherwise, authentication fails. Full-dimensional comparison serves as a compensation mechanism to ensure accurate user identification even if compression comparison fails.

[0077] To illustrate this process more intuitively, let's take a look at a specific example. Suppose there are two users in the system: User A is a frequent passer who passes through the face pass 20 times a day; User B is a low-frequency passer who passes through twice a day. Let N = 256, K min = 20, K max = 100, α = 2, and the time window W = 1 day. For user A, PA = 20 times / day, KA = max(20, 100 – 2×20) = max(20, 60) = 60; for user B, PB = 2 times / day, KB = max(20, 100 - 2×2) = max(20, 96) = 96.

[0078] The system selects the top 60 feature dimensions from the difference heat map for user A and the top 96 feature dimensions for user B. During authentication, user A's compressed feature vector contains only 60 values, which reduces computational complexity and is suitable for fast authentication. User B's compressed feature vector contains 96 values, retaining more information to ensure accuracy. If the compressed comparison for user A fails, the system performs a full-dimensional comparison using the full 256-dimensional vector as a fallback.

[0079] This invention is particularly suitable for the localized environment of information technology innovation, as the computing power of localized chips may be limited. By dynamically adjusting the feature dimensions according to the frequency of passage, high-frequency users can quickly complete authentication with fewer computing resources, while low-frequency users can maintain higher accuracy. The entire process, from facial feature extraction, difference heat map construction, dimensionality compression to authentication comparison, forms an efficient and secure one-face management process. At the same time, local storage and encryption protection of data further enhance the security of the system, meeting the dual requirements of security and efficiency of localized equipment.

[0080] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A one-face management method for the localization of information technology, characterized by: It includes the following steps: Collect the user's face image, and use the face recognition model to extract the face feature vector, where the face feature vector is a multi-dimensional vector with a length of N; Based on the set of face feature vectors of multiple users, calculate the identity discrimination ability of each feature dimension in the face recognition task, and construct a difference heat map, where the difference heat map is a one-dimensional vector with a length of N, and each element represents the difference heat score of the corresponding feature dimension in the face feature vector; Statistically calculate the passing frequency P of each user within a preset time window, and map the user passing frequency P to its corresponding feature compression dimension number K according to the preset mapping function f(P), where K < N and K decreases as P increases; Select the top K feature dimensions with the highest difference heat scores from the difference heat map as the compression dimension index set of the user; use the compression dimension index set to extract the feature values of the corresponding dimensions from the face feature vector, form a compressed feature vector, and locally store it as the user's compressed template; During face recognition authentication, extract the compressed feature vector based on the passing frequency P1 of the currently recognized face feature vector within the historical range, and perform similarity comparison with all compressed templates within a certain range before and after P1 of the corresponding passing frequency stored locally. If there is a compressed template whose similarity meets the passing condition, the authentication is completed; otherwise, switch to full-dimensional comparison for compensatory matching; Among them, constructing the difference heat map includes: Group the face feature vectors of multiple users according to the user identity label, and separately calculate the mean difference and within-group variance of each feature dimension among different users; For the feature dimensions, and calculate their identity discrimination ability scores in face recognition tasks ; For all feature dimensions The values ​​of are normalized to form a one-dimensional vector of length N as the difference heat map, where each element represents the difference heat score of the corresponding feature dimension.

2. According to the one-face management method for the localization of information technology innovation according to claim 1, it is characterized in that: The expression of f(P) is: f(P)=max(K min ,K max −α·P), where α is a compression adjustment coefficient greater than 0, K min is the minimum number of compressed dimensions, K max is the maximum number of compressed dimensions.

3. The one-face management method for localization of information technology innovation according to claim 1 is characterized in that: Among them, It is calculated using the ratio of between-class variance to within-class variance, which is defined as follows: ， Where: Indicates the The inter-class variance of dimensional features among different users; b Represents between classes; Indicates the The intra-class variance of the dimension feature within the same user; w Indicates within a class.

4. The one-face management method for localization of information technology innovation according to claim 1 is characterized in that: The similarity comparison includes: Calculate the Euclidean distance or cosine similarity between the extracted compressed feature vector and the local compressed template; Judge whether the similarity meets the preset compression comparison threshold. If it meets, the authentication passes; otherwise, switch to full-dimensional comparison; In full-dimensional comparison, use the complete face feature vector for matching and judge whether it meets the preset full-dimensional comparison threshold.

5. The one-face management method for the localization of information technology according to claim 1 is characterized in that: The statistical method of the user passing frequency P includes: Within the sliding time window W, record the number of times n the user passes; The passing frequency is defined as P = n / W, where the unit of W is hours, days or sub-time units.

6. The one-face management method for the localization of information technology according to claim 1 is characterized in that: The local storage method includes binding the compressed template of each user to its identity ID and using an encryption algorithm to store and protect the template data.

7. A system for implementing the one-face management method for localization of information technology innovation as described in claim 1, characterized in that: It includes: A face feature extraction module for collecting the user's face image and using the face recognition model to extract the face feature vector, where the face feature vector is a multi-dimensional vector with a length of N; A difference heat analysis module for calculating the identity discrimination ability of each feature dimension in the face recognition task based on the set of face feature vectors of multiple users and constructing a difference heat map, where the difference heat map is a one-dimensional vector with a length of N, and each element represents the difference heat score of the corresponding feature dimension in the face feature vector; A passing frequency statistics module for statistically calculating the passing frequency P of each user within a preset time window, A compression dimension mapping module, which is used to map the user's passing frequency P to its corresponding characteristic compression dimension number K according to a preset mapping function f(P), where K < N, and K decreases as P increases; A compression template generation module, which is used to select the top K characteristic dimensions with the highest difference heat scores from the difference heat map as the compression dimension index set of the user, and use the compression dimension index set to extract the eigenvalues of the corresponding dimensions from the face feature vectors to form a compressed feature vector, which is locally stored as the user's compression template; A face recognition authentication module, which is used for face recognition authentication. Based on the passing frequency P1 of the currently recognized face feature vector within the historical range, a compressed feature vector is extracted, and similarity comparison is performed with all compression templates corresponding to the passing frequency within a certain range before and after P1 stored locally. If there is a compression template whose similarity meets the passing condition, the authentication is completed; otherwise, it switches to full-dimension comparison for compensatory matching; Among them, constructing the difference heat map includes: Grouping the face feature vectors of multiple users according to user identity labels, and respectively calculating the mean difference and within-group variance of each feature dimension among different users; For the feature dimensions, and calculate their identity discrimination ability scores in face recognition tasks ; For all feature dimensions The values ​​of are normalized to form a one-dimensional vector of length N as the difference heat map, where each element represents the difference heat score of the corresponding feature dimension.

8. The system according to claim 7, characterized in that: The expression of f(P) is: f(P)=max(K min ,K max −α·P), where α is a compression adjustment coefficient greater than 0, K min is the minimum number of compressed dimensions, K max is the maximum number of compressed dimensions.

9. The system according to claim 7, characterized in that: The local storage method includes binding each user's compression template to its identity ID, and using an encryption algorithm to store and protect the template data.

Images