Entrance guard equipment face feature map replacement method based on event driving
By using the self-service detection and source recognition of face feature maps using the self-set deep learning discriminator and identity feature vector extraction module in the access control device, the problem of access control device verification failure caused by changes in face features of college students is solved, and high accuracy is achieved The process of changing face feature maps for sexual and security.
Patent Information
- Application Number
- CN202411285375.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-13
- Publication Date
- 2024-12-20
AI Technical Summary
Changes in facial features of college students have led to failure of verification of access control equipment. The existing technology lacks self-service detection of facial feature map replacement function on mobile terminals, which may cause system errors and security risks, and it is necessary to improve accuracy and autonomous error correction capabilities.
Using an event-driven method, the self-built deep learning discriminator Z1 is used to determine whether the face feature map uploaded by the user is a face. The identity feature vector extraction module and the similar feature calculation module are used to identify the source of the person, and the encryption and dispatch process is replaced. Face feature map ensures high security and accuracy of the access control system.
Improve the accuracy of the facial feature map replacement process, reduce the probability of uploading non-face feature maps, reduce system errors and waste of storage resources, and ensure the security and user experience of the access control system, especially in large-scale people in colleges and universities. When face transmission is required.
Smart Images
Figure CN119169730A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent access control, and in particular to a method for replacing a facial feature map of an access control device based on event-driven. Background Art
[0002] Colleges and universities are places where people gather relatively frequently. For the safety of students, most colleges and universities have established access control equipment, and teachers and students enter campuses and dormitories through facial verification.
[0003] The age of college students is generally around 18-25, and their facial image features are prone to change. The facial feature map entered for the first time can pass the verification of the access control device, but after a period of time, when the facial features change, the comparison between the real-time facial feature map and the existing feature map in the system will be unsuccessful, and therefore they will not be able to pass the access control device. At this time, the facial image feature information in the access control system needs to be updated.
[0004] In the existing technology, the facial feature image replacement function of the access control device located on the mobile terminal only exists in the collection of photos and cannot self-detect whether it is a facial information image. If a non-face image or a face image of a person other than the person is uploaded, it may cause system errors or security risks, and requires high manual intervention. Therefore, while maintaining the mobile terminal to replace the access control face photo, it maintains a high degree of accuracy and a certain system self-debugging function, ensuring the security of the access control system while facilitating the travel of university personnel. It is worth studying a feasible solution for it. Summary of the invention
[0005] Purpose of the invention:
[0006] In order to solve the problems mentioned in the background technology, the present invention provides a method for replacing facial feature maps of access control devices based on event-driven. The facial feature maps to be replaced uploaded by the user's mobile client are judged by a self-established deep learning discriminator Z1, and non-face photos are eliminated. Subsequently, the facial feature maps are used to identify the source of the person through the identity feature vector extraction module and the similar feature calculation module. Finally, the face is replaced and injected into the face machine through the encrypted sending process. While further improving the accuracy of the facial feature map replacement process of the personal mobile client, the high security requirements of the access control system are maintained.
[0007] Technical solution:
[0008] The present invention discloses a method for replacing a facial feature map of a door access control device based on event-driven operation, the method comprising the following steps:
[0009] S1: The user writes and submits a facial feature map replacement application on the user end, and the access control service management system backend automatically agrees to the request and sends a collection information notification;
[0010] S2: Enter the review process, the user end reviews and submits the facial feature map for facial feature extraction;
[0011] S2.1: Based on the FaceNet model, we optimized our own F9 model and used the Inception network architecture to build our own deep learning discriminator Z1.
[0012] S2.2: Use the self-developed deep learning discriminator Z1 to determine whether the face feature map uploaded by the user contains a face. If so, save it and submit it to S3. If not, issue a re-upload instruction.
[0013] S3: After the user passes the review, the backend of the access control service management system is notified through the http protocol, and the identity feature vector extraction module and the similar feature calculation module are used to determine whether it is the user. If it is, the original feature map is replaced, otherwise a re-upload instruction is issued;
[0014] S4: After the review is completed, the review process ends, and the access control service management system backend calls the task delivery process to generate a feature map and deliver it to the access control device.
[0015] Furthermore, the operation process of the deep learning discriminator Z1 described in S2.1 is as follows:
[0016] The input initial feature map contains 5023*3 vertex feature information, that is, the three-dimensional coordinates of the vertices. The relationship information between the vertices has been initialized when the network is defined. Multiple ChebConv convolution kernels with k=6 are used in the discriminator to extract features. The LeakyReLU activation function is used after each convolution, and the negative_slope parameter is set to 0.2 to retain some negative information. The number of nodes is reduced by downsampling. In the last two layers of the discriminator, the F9 model optimized and improved based on the FaceNet model is used to further extract features. Multiple convolution layers, pooling layers and activation functions are used to extract image features layer by layer to obtain the final feature vector. The extracted feature vector is compared with the face feature vector. If the recognized face feature vector is as close as possible in the vector space, the feature vector distance calculation formula is as follows:
[0017]
[0018] Among them, f 1,i and f 2,i They are f 1 and f 2 The component in the i-th dimension.
[0019] Furthermore, the F9 model structure is as follows:
[0020] The Inception network architecture is used, which consists of multiple convolutional layers, batch normalization layers, non-linear activation functions, and pooling layers. The specific structure is input layer, deep convolutional neural network, embedding vector layer, and triplet loss.
[0021] Furthermore, the identity feature vector extraction module described in S3 specifically extracts a feature vector from the feature information of the original face feature map through a frame generation technology acquisition module as a generated frame image, and then extracts the current feature vector by extracting the replacement face information uploaded by the user to generate an identity feature vector of the frame; and then makes a judgment through a similar feature vector calculation module by comparing the distance between the embedding vectors of the two faces, and then judges whether the original face feature map and the face feature map to be replaced belong to the same person.
[0022] Furthermore, the similarity feature calculation module is provided with a similarity feature calculation formula, which calculates the similarity between the identity feature vector of the generated frame and the original feature vector by inputting the similarity between the two, and the formula is as follows:
[0023]
[0024] Among them, f 1 : The feature vector of the first face image, f 2 : The feature vector of the second face image, f 1,i : The i-th component of vector f1, f 2,i : the i-th component of vector f2, d: the dimension of the feature vector; each component of the feature vector has a different impact on the distance, and a weight w is introduced for each component i , w i It is set to 0.36 based on the feature desirability.
[0025] Furthermore, the specific steps of S4 are:
[0026] After the review process has reviewed all applications, the review process ends, and the access control service management system calls the task delivery process in the background, generates a feature map and delivers it to the access control device, reads each task and delivers it to the corresponding face recognition machine, and modifies the completion status value of the task according to whether the task is successfully delivered. The feature face identity information vector of the new face extracted from the pre-acquired new face information image is injected into the replacement face representation to obtain the face features after identity replacement;
[0027] After the issuance process completes the issuance of all tasks, the application review process is started again to prevent users from submitting new applications during the task issuance period. The middleware interface of the http protocol post method is used to realize the linkage of systems with different architectures and cross-platforms, and automatically complete the application, review and issuance of feature map changes. The improved HTTP / 3.1 protocol is used to reduce the number of TCP connections through multiplexing, optimize the transmission of data frames, reduce delays, and use the TLS encryption algorithm to ensure that POST requests are encrypted during transmission to prevent sensitive data leakage. Asymmetric encryption is used to protect data.
[0028] Furthermore, the data sent to each face recognition machine is encrypted by the IoT control chip, and the HSM hardware security module and TPM trusted platform module are used to protect the data using the M1 encryption algorithm. The specific calculation formula is as follows:
[0029] C=AES-Encrypt(M,K)
[0030] Among them, M is the plaintext block, K is the key, and C is the encrypted ciphertext block.
[0031] Furthermore, the present invention discloses a computer storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the event-driven access control device facial feature map replacement method as described in any one of claims 1 to 7 are implemented.
[0032] Beneficial effects:
[0033] 1. The present invention uses a self-designed deep learning discriminator Z1, multiple ChebConv convolution kernels with k=6, and a series of F9 models optimized and improved based on the FaceNet model to further extract features. Compared with the existing technology, the present invention has a more accurate facial feature map recognition capability, reduces the upload of non-facial feature maps, reduces the probability of bugs in the access control system, and reduces the meaningless waste of storage resources.
[0034] 2. The present invention utilizes the identity feature vector extraction module and the similar feature calculation module to identify the source of the facial feature map, further screen and ensure that the facial feature map uploaded by the mobile user end is correct, and maintain the robustness of the system.
[0035] 3. The present invention adopts an encrypted sending process and a TLS encryption algorithm based on the encryption of the IoT control chip to ensure that the POST request is encrypted during the transmission process, and uses an improved HTTP / 3.1 protocol to reduce the number of TCP connections and improve data transmission efficiency through multiplexing. While ensuring the privacy of user uploaded face photos, it reduces the delay in face replacement in the access control system and improves the user experience in scenarios where a large number of face transmissions are required, such as the start of the school year in colleges and universities and enrollment after the holidays. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is the overall flow chart of the present invention;
[0037] Figure 2 It is a schematic diagram of the review process of the present invention;
[0038] Figure 3 This is a schematic diagram of the issuing process of the present invention;
[0039] Figure 4 This is the structure diagram of the self-developed deep learning discriminator Z1 of the present invention;
[0040] Figure 5 This is a schematic diagram of the F9 model structure of the present invention. DETAILED DESCRIPTION
[0041] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.
[0042] In this embodiment, the backend of the access control service management system is written based on the Windows platform, and the mobile APP is written based on the Android platform.
[0043] This embodiment provides a method for replacing the facial feature map of a door access device based on event-driven. The overall process is as follows: Figure 1 As shown, the steps are as follows:
[0044] S1: Figure 2 As shown, the user writes and submits the facial feature map replacement application on the user end (mobile APP or web page), and the access control service management system backend automatically agrees to the request and sends a collection information notification, and notifies the review process;
[0045] S2: Figure 3 As shown, the review process begins, and the user reviews and submits the facial feature map for facial feature extraction;
[0046] S2.1: Based on the self-designed F9 model optimized and improved by the FaceNet model, using the Inception network architecture, we built our own deep learning discriminator Z1;
[0047] S2.2: Use the self-developed deep learning discriminator Z1 to determine whether the face feature map uploaded by the user contains a face. If so, save it and submit it to S3. If not, issue a re-upload instruction.
[0048] like Figure 4 As shown, the structure of the deep learning discriminator Z1 is as follows:
[0049] Used to discriminate the input mesh. The input mesh contains 5023*3 vertex feature information, that is, the three-dimensional coordinates of the vertices. The relationship information between the vertices has been initialized when the network is defined. Multiple ChebConv convolution kernels with k=6 are used in the discriminator to extract features. The LeakyReLU activation function is used after each convolution, and its negative_slope parameter is set to 0.2 to retain some negative information. In addition, the network reduces the number of nodes by downsampling. In the last two layers, the F9 model optimized and improved based on the FaceNet model is used to further extract features and enhance the network's ability to aggregate information. It extracts image features layer by layer through multiple convolutional layers, pooling layers, and activation functions, and finally obtains a compact but rich feature vector.
[0050] F9 structure diagram Figure 5 As shown: The F9 model uses a deep convolutional neural network to extract facial features. It uses the Inception network architecture, which consists of multiple convolutional layers, batch normalization layers, nonlinear activation functions (such as ReLU), and pooling layers. These layers gradually extract and compress the spatial features in the face image. Overall structure: input layer → deep convolutional neural network (CNN) → embedding vector layer → triplet loss.
[0051] The extracted feature vector is compared with the face feature vector. If the recognized feature vector is a face feature vector, it is as close as possible in the vector space. The feature vector distance calculation formula is as follows:
[0052]
[0053] Among them, f 1,i and f 2,i They are f 1 and f 2 The component in the i-th dimension.
[0054] After detecting a face, the backend of the access control service management system will perform face alignment to standardize the process. Face alignment means rotating and scaling the face to a uniform perspective to reduce the impact of changes in posture, angle, expression, etc.
[0055] S3: After the APP or web front-end passes the review, the back-end of the access control service management system is notified through the http protocol that a user has submitted a face replacement application. After receiving the APP notification, the back-end immediately runs the review process and uses the identity feature vector extraction module and the similarity feature calculation module to determine whether it is the person himself. The review program automatically reviews each application and generates corresponding dispatch tasks based on the number of face recognition machines. If there are 20 face recognition machines, one application that passes the review will generate 20 dispatch tasks. There is no need for the user to authenticate at the face recognition machine.
[0056] Specifically, the identity feature vector extraction module: In order to ensure that the generated facial feature information is the identity feature information of the source person, the self-developed improved frame generation technology is used here. The feature vector is extracted from the original feature information through the frame generation technology acquisition module as the generated frame image, and then the face information uploaded by the user is extracted through the identity feature vector module, the current feature vector is extracted, and the identity feature vector of the frame is generated; and then the similar feature vector calculation module is used for judgment. By comparing the distance between the embedding vectors of two faces, it is judged whether the two faces belong to the same person.
[0057] The similarity feature calculation module is equipped with a similarity feature calculation formula, which calculates the similarity between the identity feature vector of the generated frame and the original feature vector by inputting them. The formula is as follows:
[0058]
[0059] Among them, f 1 : The feature vector of the first face image, f 2 : The feature vector of the second face image, f 1,i : The i-th component of vector f1, f 2,i : the i-th component of vector f2, d: the dimension of the feature vector; each component of the feature vector has a different impact on the distance, and a weight w is introduced for each component i , w i It is set to 0.36 based on the feature desirability.
[0060] S4: After the review is completed, the review process ends. The access control service management system backend calls the task delivery process to generate a feature map and deliver it to the access control device. The specific steps are as follows:
[0061] After the review process has reviewed all applications, the review process ends, and the access control service management system calls the task delivery process in the background, generates a feature map and delivers it to the access control device, reads each task and delivers it to the corresponding face recognition machine, and modifies the completion status value of the task according to whether the task is successfully delivered. The feature face identity information vector of the new face extracted from the pre-acquired new face information image is injected into the replacement face representation to obtain the face features after identity replacement;
[0062] After the dispatch process completes the dispatch of all tasks, the application review process is started again to prevent users from submitting new applications during the task dispatch period. The middleware interface of the http protocol post method is used to realize the linkage of systems with different architectures and cross-platforms, and automatically complete the application, review and dispatch of feature map changes. The improved HTTP / 3.1 protocol is used to reduce the number of TCP connections through multiplexing, optimize the transmission of data frames, reduce delays, and use the TLS encryption algorithm to ensure that POST requests are encrypted during transmission to prevent sensitive data leakage. Asymmetric encryption is used to protect data. If the dispatch task queue is empty, the operation ends.
[0063] In the process of sending, the data sent to each face recognition machine is encrypted by the IoT control chip, and the HSM hardware security module, TPM trusted platform module and M1 encryption algorithm are used for data protection. The specific calculation formula is as follows:
[0064] C=AES-Encrypt(M,K)
[0065] Among them, M is the plaintext block, K is the key, and C is the encrypted ciphertext block.
[0066] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for replacing a facial feature map of an access control device based on event-driven operation, characterized in that: The method comprises the following steps: S1: The user writes and submits a facial feature map replacement application on the user end, and the access control service management system backend automatically agrees to the request and sends a collection information notification; S2: Enter the review process, the user end reviews and submits the facial feature map for facial feature extraction; S2.1: Based on the FaceNet model, we optimized our own F9 model and used the Inception network architecture to build our own deep learning discriminator Z1. S2.2: Use the self-developed deep learning discriminator Z1 to determine whether the face feature map uploaded by the user contains a face. If so, save it and submit it to S3. If not, issue a re-upload instruction. S3: After the user passes the review, the backend of the access control service management system is notified through the http protocol, and the identity feature vector extraction module and the similar feature calculation module are used to determine whether it is the user. If it is, the original feature map is replaced, otherwise a re-upload instruction is issued; S4: After the review is completed, the review process ends, and the access control service management system backend calls the task delivery process to generate a feature map and deliver it to the access control device.
2. The method for replacing a facial feature map of a door access control device based on event-driven operation according to claim 1, characterized in that: The operation process of the deep learning discriminator Z1 described in S2.1 is as follows: The input initial feature map contains 5023*3 vertex feature information, that is, the three-dimensional coordinates of the vertices. The relationship information between the vertices has been initialized when the network is defined. Multiple ChebConv convolution kernels with k=6 are used in the discriminator to extract features. The LeakyReLU activation function is used after each convolution, and the negative_slope parameter is set to 0.2 to retain some negative information. The number of nodes is reduced by downsampling. In the last two layers of the discriminator, the F9 model optimized and improved based on the FaceNet model is used to further extract features. Multiple convolution layers, pooling layers and activation functions are used to extract image features layer by layer to obtain the final feature vector. The extracted feature vector is compared with the face feature vector. If the recognized face feature vector is as close as possible in the vector space, the feature vector distance calculation formula is as follows: Among them, f 1,i and f 2,i are the components of f1 and f2 in the i-th dimension respectively.
3. The method for replacing a facial feature map of a door access control device based on event-driven operation according to claim 2, characterized in that: The F9 model structure is as follows: The Inception network architecture is used, which consists of multiple convolutional layers, batch normalization layers, non-linear activation functions, and pooling layers. The specific structure is input layer, deep convolutional neural network, embedding vector layer, and triplet loss.
4. The method for replacing a facial feature map of a door access control device based on event-driven operation according to claim 1, characterized in that: The identity feature vector extraction module described in S3 specifically extracts feature vectors from feature information of the original face feature map through a frame generation technology acquisition module as a generated frame image, and then extracts the current feature vector by extracting the replacement face information uploaded by the user to generate an identity feature vector for the frame; and then makes a judgment through a similar feature vector calculation module by comparing the distance between the embedding vectors of the two faces, and then judging whether the original face feature map and the face feature map to be replaced belong to the same person.
5. The method for replacing the facial feature map of access control equipment based on event-driven according to claim 1 or 4, characterized in that: The similarity feature calculation module is provided with a similarity feature calculation formula, which calculates the similarity between the identity feature vector of the generated frame and the original feature vector by inputting the same, and the formula is as follows: Where, f1: the feature vector of the first face image, f2: the feature vector of the second face image, f 1,i : The i-th component of vector f1, f 2,i : the i-th component of vector f2, d: the dimension of the feature vector; each component of the feature vector has a different impact on the distance, and a weight w is introduced for each component i , w i It is set to 0.36 based on the feature desirability.
6. The method for replacing a facial feature map of a door access control device based on event-driven operation according to claim 1, characterized in that: The specific steps of S4 are: After the review process has reviewed all applications, the review process ends, and the access control service management system calls the task delivery process in the background, generates a feature map and delivers it to the access control device, reads each task and delivers it to the corresponding face recognition machine, and modifies the completion status value of the task according to whether the task is successfully delivered. The feature face identity information vector of the new face extracted from the pre-acquired new face information image is injected into the replacement face representation to obtain the face features after identity replacement; After the issuance process completes the issuance of all tasks, the application review process is started again to prevent users from submitting new applications during the task issuance period. The middleware interface of the http protocol post method is used to realize the linkage of systems with different architectures and cross-platforms, and automatically complete the application, review and issuance of feature map changes. The improved HTTP / 3.1 protocol is used to reduce the number of TCP connections through multiplexing, optimize the transmission of data frames, reduce delays, and use the TLS encryption algorithm to ensure that POST requests are encrypted during transmission to prevent sensitive data leakage. Asymmetric encryption is used to protect data.
7. The method for replacing a facial feature map of a door access control device based on event-driven operation according to claim 6, characterized in that: The data sent to each face recognition machine is encrypted by the IoT control chip, and the HSM hardware security module and TPM trusted platform module are used to protect the data using the M1 encryption algorithm. The specific calculation formula is as follows: C=AES-Encrypt(M,K) Among them, M is the plaintext block, K is the key, and C is the encrypted ciphertext block.
8. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for replacing a facial feature map of a door access control device based on event-driven operation are implemented as described in any one of claims 1 to 7.