5g recorder face recognition front filter method and system

Abstract

The application relates to the field of face recognition, in particular to a 5G recorder face recognition pre-filtering method and system, which comprises an image capturing module, a face feature extraction module, a face characteristic dimension reduction and normalization module, a face library module, a face filtering tree module, a face pre-filtering module, a face characteristic block module and a face hierarchical comparison module. The face feature filtering tree method is used to realize face matching pre-filtering of 5G recorders and other equipment. The face library is blocked according to normalized face characteristics, a hierarchical method is realized according to face feature matching strings for block matching, a matching face is found as much as possible, and the calculation amount of a server is reduced.

Description

A 5G recorder face recognition pre-filtering method and system Technical Field

[0001] This invention relates to the field of facial recognition, specifically to a pre-filtering method and system for facial recognition in a 5G recorder. Background Technology

[0002] With people paying increasing attention to social environmental security, devices or systems for detecting outsiders are being installed in workplaces, residences, parking lots, and other environments. The most convenient and commonly used method is to identify and distinguish people coming and going through facial recognition to ensure a "no strangers allowed" social environment and further safeguard the security of various social environments.

[0003] In existing facial recognition systems, face comparison algorithms are executed by front-end devices. Due to limitations in hardware resources, the face database cannot be set too large. Therefore, large face databases usually adopt a centralized approach. However, this inevitably places very high demands on the computing power of the center. Therefore, a method is needed to allow the front-end devices to perform face comparison filtering, filtering out faces that are not in the face database, thereby reducing the pressure on the center. Summary of the Invention

[0004] The purpose of this invention is to provide a novel high-speed scanning device for express delivery, which enables face comparison and filtering at the front-end device to filter out faces not in the face database, thereby reducing the pressure on the center and solving the existing technical defects and unmet technical requirements.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a 5G recorder face recognition pre-filtering system, comprising:

[0006] The image capture module (1) is used to capture video images, generate image frames, and send the image frames to the face feature extraction module (2);

[0007] The face feature extraction module (2) is used to receive the image frame sent by the image capture module (1), extract face features using the feature extractor, encode the feature data into a 128-dimensional face feature vector, and then send the 128-dimensional face feature vector to the face feature dimensionality reduction and normalization module (3).

[0008] Facial features include facial contours, eyes, nose, mouth, etc.

[0009] The face feature dimensionality reduction and normalization module (3) is used to receive the 128-dimensional face feature vectors from the face feature extraction module (2) and the face database module (4), and calculate and generate the normalization value of the face feature 1-dimensional space. The normalization value of the face feature 1-dimensional space obtained from the data in the face feature extraction module (2) and the 128-dimensional face feature vector in the face feature extraction module (2) are sent to the face pre-filtering module (6). The normalization value of the face feature 1-dimensional space obtained from the data in the face database module (4) and the 128-dimensional face feature vector in the face database module (4) are sent to the face feature block module (7). The normalization value of the face feature 1-dimensional space obtained from the data in the face database module (4) is sent separately to the face filtering tree module (5).

[0010] The face database module (4) is used to store the 128-dimensional face feature vectors of labeled personnel information and send the 128-dimensional face feature vectors to the face feature dimensionality reduction and normalization module (3);

[0011] The face filtering tree module (5) is used to receive the normalized value of the face feature 1-dimensional space in the vector dimensionality reduction and normalization module (3), establish the face filtering initial tree, and send the face filtering initial tree to the face pre-filtering module (6).

[0012] The face pre-filtering module (6) is used to receive the data of the face filtering initial tree from the face filtering tree module (5), reconstruct the face filtering tree, and at the same time receive the normalized value of the face feature 1D space sent by the face feature dimensionality reduction and normalization module (3) and the 128-dimensional face feature vector in the face feature extraction module (2). The normalized value of the face feature 1D space is converted into a face feature string, and then the face feature string is compared with the face filtering tree to generate a face feature matching string. The face feature matching string and the 128-dimensional face feature vector are sent to the face layer comparison module (8).

[0013] The face feature segmentation module (7) is used to receive the normalized value of the face feature 1D space and the 128-dimensional face feature vector sent by the face feature dimensionality reduction and normalization module (3), and then segment the face feature values ​​into blocks, and send the normalized value of the starting face feature 1D space and the normalized value of the ending face feature 1D space of each block to the face layer comparison module (8).

[0014] The face layer comparison module (8) receives the face feature matching string and 128-dimensional face feature vector sent by the face pre-filtering module (6), and at the same time receives the normalized value of the starting face feature 1D space and the normalized value of the ending face feature 1D space of each block sent by the face feature block module (7). The module matches the contents of the two modules to complete the confirmation of face matching.

[0015] A video inspection method based on a low-power 5G recorder includes:

[0016] 1) The image capture module (1) acquires video images, generates image frames, and sends the image frames to the face feature extraction module (2);

[0017] 2) The face feature extraction module (2) receives the image frame sent by the image capture module (1), and uses the feature extractor ResNet to extract face features. Then, the feature data is encoded into a 128-dimensional face feature vector through the SVM feature encoder, and the 128-dimensional face feature vector is sent to the face feature vector dimensionality reduction and normalization module (3).

[0018] 3) The face feature reduction and normalization module (3) receives the 128-dimensional face feature vector from the face feature extraction module (2), calculates the normalization value in 1-dimensional space, and sends the normalization value in 1-dimensional space and the 128-dimensional face feature vector to the face pre-filtering module (6).

[0019] 4) The face feature dimensionality reduction and normalization module (3) receives the 128-dimensional face feature vector from the face database module (4), calculates the normalization value of the face feature 1-dimensional space, sends the normalization value of the face feature 1-dimensional space to the face filter tree module (5), and the face feature dimensionality reduction and normalization module (3) combines the obtained normalization value of the face feature 1-dimensional space with the 128-dimensional face feature vector face feature block module (7).

[0020] 5) The face database module (4) stores the 128-dimensional face feature vectors of the labeled personnel information and sends the 128-dimensional face feature vectors to the face feature dimensionality reduction and normalization module (3). The face database module (4) sends all the 128-dimensional face feature vectors to the face feature dimensionality reduction and normalization module (3) for the first time.

[0021] Used in the face filtering tree module (5) to establish the initial face filtering tree;

[0022] After the face database module (4) changes (addition and deletion), only the changed 128-dimensional face feature vector is sent to the face feature dimensionality reduction and normalization module (3), which is used by the face filtering tree module (5) to generate the face filtering difference tree. Because the face database is a large database, only the changes and differences are modified to reduce the central computation.

[0023] 6) Face filtering tree module (5) Receive vector dimensionality reduction and normalization module (3) Normalization value of face feature 1-dimensional space;

[0024] 7) The face filtering tree module (5) uses the normalized value of the 1-dimensional space of face features to generate the initial face filtering tree and the face filtering difference array;

[0025] 8) Face filtering tree module (5) builds an initial face filtering tree from all data in the face database. The face filtering tree module (5) sends the initial face filtering tree to the face pre-filtering module (6).

[0026] 9) The face pre-filtering module (6) receives the face filtering initial tree data from the face filtering tree module (5) and reconstructs the face filtering tree;

[0027] 10) The face pre-filtering module (6) receives the normalized value of the face feature in 1-dimensional space and the 128-dimensional face feature vector from the face feature dimensionality reduction and normalization module (3);

[0028] 11) The face pre-filtering module (6) converts the normalized value of the face feature 1D space into a face feature string, and compares the strings of the face feature string one by one on the face filter tree. When no matching node is found, the query is terminated and a face feature matching string is obtained.

[0029] 12) Send the face feature matching string and the 128-dimensional face feature vector to the face layer comparison module (8).

[0030] Sending the face feature matching string to the face layer comparison module (8) is to reduce the number of face comparisons performed by the face layer comparison module (8);

[0031] 13) Compare the length of the face feature matching string with the previous matching threshold to perform pre-screening and filtering.

[0032] 14) The face feature segmentation module (7) receives the normalized value of the face feature 1-dimensional space and the 128-dimensional face feature vector from the face feature dimensionality reduction and normalization module (3), segments the face feature values, and provides a comparison database for the face layer comparison module (8).

[0033] 15) The face layer comparison module (8) receives the face feature matching string and 128-dimensional face feature vector from the face pre-filtering module (6);

[0034] 16) The face layer comparison module (8) compares the received face feature matching string with the 128-dimensional face feature vector to find the face feature blocks and complete the face matching.

[0035] Preferably, the face feature dimensionality reduction and normalization module (3) receives a 128-dimensional face feature vector and performs the step of calculating and generating a normalized value for a 1-dimensional face feature space:

[0036] 3.1) Face feature dimensionality reduction and normalization module (3) First, use random projection to project the 128-dimensional face feature vector onto the 2-dimensional plane space, and randomly generate a 128×2 projection matrix. Each element of the matrix is ​​randomly sampled from the standard normal distribution N(0,1).

[0037] The random projection transformation from 128-dimensional to 2-dimensional planar space is random, ensuring geometric features and thus preserving the distance relationships between 128-dimensional facial feature vectors.

[0038] 3.2) Face feature dimensionality reduction and normalization module (3) Then take the two values ​​of the 2-dimensional vector as the endpoint of the plane vector A, the coordinate center (0, 0) as the starting point of the plane vector A, and take the counterclockwise angle between the plane vector A and the positive x-axis as the normalization value of the 1-dimensional space.

[0039] Using the angle value as the final normalization value can avoid the distance error introduced by the difference in the size of the face image captured by the image capture module (1).

[0040] Preferably, the face filtering tree module (5) generates the initial face filtering tree in the following steps:

[0041] 4.1) The maximum tree depth of the face filtering tree module (5) is set to H;

[0042] 4.2) Convert the normalized values ​​of the 1D face feature space into a face feature string. Output the normalized values ​​of the 1D face feature space as a decimal string, take H-1 decimal places, remove the dot "." character from the string, and obtain a face feature string of length H.

[0043] 4.3) Store the facial feature string in a trie. A root node is created, which does not store any characters. Each character of the facial feature string is used as a node in the tree and stored sequentially in the tree. The sequence from the root node to each page node is a facial feature string.

[0044] 4.4) Use the initial face filtering tree as the face filtering tree.

[0045] After the face database module (4) changes (addition and deletion), only the changed 128-dimensional face feature vector is sent to the face feature dimensionality reduction and normalization module (3), which is used by the face filtering tree module (5) to generate the face filtering difference tree. Because the face database is a large database, only the changes and differences are modified to reduce the central computation.

[0046] Preferably, when the face filter tree module (5) changes, it generates a face filter differential array and sends it to the face pre-filter module (6) of the 5G recorder.

[0047] The new face filtering differential tree only includes the changed parts, reducing the amount of intermediate network data transmission and improving the overall system availability.

[0048] Preferably, the step of the face filtering tree module (5) generating the face filtering difference array is as follows:

[0049] 6.1) First, calculate the face feature string to be added or deleted. If a new face feature string is added, store each character of the face feature string as a node in the tree.

[0050] 6.2) If the character node A already exists in the face filter tree, use the character node of the face filter tree; otherwise, create a new character node, put the new node into the difference array, and add the new node to the face filter tree.

[0051] 6.3) If a face feature string deletion operation is performed, traverse the difference tree from the root node to find the character path of the face feature string, and check in reverse order whether the character node B of the character path has child nodes;

[0052] 6.4) If there are no child nodes, then make character node B the deletion node, put the deletion node into the difference array, and delete the node from the face filtering tree.

[0053] 6.5) Face filtering tree module (5) sends the difference array to the face pre-filtering module (6) of the 5G recorder.

[0054] Preferably, the pre-detection filtering method in the face pre-filtering module (6) is as follows:

[0055] 7.1) If the length of the face feature matching string is less than the previous matching threshold, the pre-detection filtering will fail, and the face will definitely not be in the face database.

[0056] 7.2) If the length of the face feature matching string is greater than or equal to the preceding matching threshold, the pre-detection filtering is successful, and the face may be in the face database.

[0057] Preferably, the method for obtaining the data comparison library in the face feature segmentation module (7) is as follows:

[0058] 8.1) Sort the normalized values ​​of the one-dimensional facial features;

[0059] 8.2) According to the square root of the face database size N The facial feature values ​​are divided into blocks to obtain the final result. Individual facial feature segmentation.

[0060] 8.3) The normalized values ​​of the starting and ending face feature 1D spaces of each block are used as a comparison database.

[0061] The normalized values ​​of the starting and ending face feature 1D spaces of each block are used by the face layer comparison module (8) to find face feature blocks. The face layer comparison module (8) performs face comparison in the face feature blocks, which greatly reduces the amount of computation.

[0062] Preferably, the face layer comparison module (8) uses a binary search method to find face feature blocks.

[0063] Preferably, the face layer comparison module (8) uses a layered method for face matching:

[0064] 10.1) The face layer comparison module (8) uses the normalized value of the face feature 1D space of the face feature matching string face feature block, extracts the 128-dimensional face feature vector corresponding to the normalized value of the 1D space of all matched face features, and performs cosine similarity calculation with the 128-dimensional face feature vector to be compared. If the cosine similarity calculation similarity is higher than the threshold, it is considered that a matching face has been found.

[0065] 10.2) If the 128-dimensional face feature vectors of this batch of face feature blocks are not higher than the threshold, the face layer comparison module (8) removes the end of the face feature matching string, then matches the normalized value of the face feature 1-dimensional space of the face feature blocks, expands the range of matching 128-dimensional face feature vectors, and then performs cosine similarity matching calculation.

[0066] 10.3) If the length of the face feature matching string is less than the previous matching threshold, and a 128-dimensional face feature vector is still not found, then the face matching is considered to have failed.

[0067] Compared with the prior art, the beneficial effects of the present invention are: by using the face feature filtering tree method, the face matching of devices such as 5G recorders can be pre-filtered, reducing the computational load of the server; in addition, by dividing the face database into blocks according to normalized face characteristics, and using a hierarchical method based on face feature matching strings for block matching, the matching face can be found as much as possible, further reducing the computational load of the server. Attached Figure Description

[0068] Figure 1 is a schematic diagram of the module relationship of the present invention;

[0069] In the figure: Image capture module 1, face feature extraction module 2, face feature dimensionality reduction and normalization module 3, face database module 4, face filtering tree module 5, face pre-filtering module 6, face feature segmentation module 7, face layer comparison module (8). Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include one or more of that feature.

[0072] Please refer to the figure for an embodiment of the present invention.

[0073] Example:

[0074] As shown in Figure 1: A 5G recorder face recognition pre-filtering system includes:

[0075] Image capture module 1 is used to capture video images, generate image frames, and send the image frames to face feature extraction module 2;

[0076] The face feature extraction module 2 is used to receive image frames sent by the image capture module 1, extract face features using a feature extractor, encode the feature data into a 128-dimensional face feature vector, and then send the 128-dimensional face feature vector to the face feature dimensionality reduction and normalization module 3.

[0077] Facial features include facial contours, eyes, nose, mouth, etc.

[0078] The face feature dimensionality reduction and normalization module 3 is used to receive 128-dimensional face feature vectors from the face feature extraction module 2 and the face database module 4, and calculate and generate normalization values ​​for the 1-dimensional face feature space. The normalization values ​​for the 1-dimensional face feature space obtained from the data in the face feature extraction module 2 and the 128-dimensional face feature vectors in the face feature extraction module 2 are sent to the face pre-filtering module 6. The normalization values ​​for the 1-dimensional face feature space obtained from the data in the face database module 4 and the 128-dimensional face feature vectors in the face database module 4 are sent to the face feature segmentation module 7. The normalization values ​​for the 1-dimensional face feature space obtained from the data in the face database module 4 are sent separately to the face filtering tree module 5.

[0079] The face database module 4 is used to store 128-dimensional face feature vectors with labeled personnel information and send the 128-dimensional face feature vectors to the face feature dimensionality reduction and normalization module 3;

[0080] The face filtering tree module 5 is used to receive the normalized values ​​of the face features in the 1-dimensional space of the vector dimensionality reduction and normalization module 3, establish the initial face filtering tree, and send the initial face filtering tree to the face pre-filtering module 6.

[0081] The face pre-filtering module 6 is used to receive the data of the initial face filtering tree from the face filtering tree module 5, reconstruct the face filtering tree, and simultaneously receive the normalized value of the 1D space of face features sent by the face feature dimensionality reduction and normalization module 3 and the 128-dimensional face feature vector from the face feature extraction module 2. It converts the normalized value of the 1D space of face features into a face feature string, compares the face feature string with the face filtering tree, generates a face feature matching string, and sends the face feature matching string and the 128-dimensional face feature vector to the face layer comparison module 8.

[0082] The face feature segmentation module 7 is used to receive the normalized value of the face feature 1D space and the 128-dimensional face feature vector sent by the face feature dimensionality reduction and normalization module 3, and then segment the face feature values ​​into blocks, and send the normalized value of the starting face feature 1D space and the normalized value of the ending face feature 1D space of each block to the face layer comparison module 8.

[0083] The face layer comparison module 8 receives the face feature matching string and 128-dimensional face feature vector sent by the face pre-filtering module 6, and at the same time receives the normalized value of the starting face feature 1D space and the normalized value of the ending face feature 1D space of each block sent by the face feature block module 7. It matches the contents of the two modules to complete the confirmation of face matching.

[0084] A video inspection method based on a low-power 5G recorder includes:

[0085] 1. Image capture module 1 acquires video images, generates image frames, and sends the image frames to face feature extraction module 2;

[0086] 2. The face feature extraction module 2 receives the image frame sent by the image capture module 1, and uses the feature extractor ResNet to extract face features. Then, the feature data is encoded into a 128-dimensional face feature vector through the SVM feature encoder, and the 128-dimensional face feature vector is sent to the face feature vector dimensionality reduction and normalization module 3.

[0087] 3. The face feature reduction and normalization module 3 receives the 128-dimensional face feature vector from the face feature extraction module 2, calculates the normalization value in 1-dimensional space, and sends the obtained normalization value in 1-dimensional space and the 128-dimensional face feature vector to the face pre-filtering module 6.

[0088] 4. The face feature dimensionality reduction and normalization module 3 receives the 128-dimensional face feature vector from the face database module 4, calculates the normalized value of the face feature 1-dimensional space, sends the normalized value of the face feature 1-dimensional space to the face filtering tree module 5, and the face feature dimensionality reduction and normalization module 3 combines the obtained normalized value of the face feature 1-dimensional space with the 128-dimensional face feature vector face feature segmentation module 7.

[0089] 5. The face database module 4 stores the 128-dimensional face feature vectors of the labeled personnel information and sends the 128-dimensional face feature vectors to the face feature dimensionality reduction and normalization module 3. For the first time, all 128-dimensional face feature vectors of the face database module 4 are sent to the face feature dimensionality reduction and normalization module 3.

[0090] Used to establish the initial face filtering tree in face filtering tree module 5;

[0091] After the face database module 4, when faces are added or deleted, only the changed 128-dimensional face feature vectors are sent to the face feature dimensionality reduction and normalization module 3, which is used by the face filtering tree module 5 to generate the face filtering difference tree. Because the face database is a large database, only the changes and differences are modified to reduce the central computation load.

[0092] 6. Face filtering tree module 5 receives vector dimensionality reduction and normalization module 3 normalization value of face features in 1-dimensional space;

[0093] 7. The face filtering tree module (5) uses the normalized value of the 1-dimensional space of face features to generate the initial face filtering tree and the face filtering difference array;

[0094] 8. Face Filtering Tree Module 5: Builds an initial face filtering tree from all data in the face database. The Face Filtering Tree Module 5 then distributes the initial face filtering tree to the Face Pre-Filtering Module 6.

[0095] 9. The face pre-filtering module 6 receives the data of the initial face filtering tree from the face filtering tree module 5 and reconstructs the face filtering tree;

[0096] 10. The face pre-filtering module 6 receives the normalized value of the face feature in 1-dimensional space and the 128-dimensional face feature vector from the face feature dimensionality reduction and normalization module 3.

[0097] 11. The face pre-filtering module 6 converts the normalized values ​​of the 1-dimensional space of face features into face feature strings, and compares them one by one on the face filtering tree according to the strings of the face feature strings. When no matching node is found, the query is terminated, and a face feature matching string is obtained.

[0098] 12. Send the face feature matching string and the 128-dimensional face feature vector to the face layer comparison module 8.

[0099] Sending the face feature matching string to the face layer comparison module 8 is to reduce the number of faces that the face layer comparison module 8 needs to compare.

[0100] 13. Compare the length of the face feature matching string with the previous matching threshold to perform pre-screening and filtering.

[0101] 14. The face feature segmentation module 7 receives the normalized value of the face feature in 1-dimensional space and the 128-dimensional face feature vector from the face feature dimensionality reduction and normalization module 3, and segments the face feature values ​​into blocks to provide a comparison database for the face layer comparison module 8.

[0102] 15. The face layer comparison module 8 receives the face feature matching string and 128-dimensional face feature vector from the face pre-filtering module 6;

[0103] 16. The face layer comparison module 8 compares the received face feature matching string with the 128-dimensional face feature vector to find the face feature blocks and complete the face matching.

[0104] The facial feature dimensionality reduction and normalization module 3 receives a 128-dimensional facial feature vector and performs the following steps to calculate and generate normalized values ​​in a 1-dimensional space of facial features:

[0105] 3.1 Face Feature Dimension Reduction and Normalization Module 3 First, random projection is used to project the 128-dimensional face feature vector onto a 2-dimensional plane space, and a 128×2 projection matrix is ​​randomly generated. Each element of the matrix is ​​randomly sampled from the standard normal distribution N(0,1).

[0106] The random projection transformation from 128-dimensional to 2-dimensional planar space is random, ensuring geometric features and thus preserving the distance relationships between 128-dimensional facial feature vectors.

[0107] 3.2 Face Feature Dimension Reduction and Normalization Module 3. Then, take the two values ​​of the 2D vector as the endpoint of the plane vector A, and the coordinate center 0, 0 as the starting point of the plane vector A. Take the counterclockwise angle between the plane vector A and the positive x-axis as the normalization value of the 1D space.

[0108] Using the angle value as the final normalized value can avoid the distance error caused by the difference in the size of the face image captured by the image capture module 1.

[0109] The steps of the face filtering tree module 5 in generating the initial face filtering tree are as follows:

[0110] 4.1 The maximum tree depth of the face filtering tree module 5 is set to H;

[0111] 4.2 Convert the normalized values ​​of the 1D face feature space into a face feature string. Output the normalized values ​​of the 1D face feature space as a decimal string, take H-1 decimal places, remove the dot "." character from the string, and obtain a face feature string of length H.

[0112] 4.3 Store the facial feature string in a trie. A root node is created, which does not store any characters. Each character of the facial feature string is used as a node in the tree and stored sequentially in the tree. The sequence from the root to each page node is a facial feature string.

[0113] 4.4 Use the initial face filtering tree as the face filtering tree.

[0114] After the face database module 4, when faces are added or deleted, only the changed 128-dimensional face feature vectors are sent to the face feature dimensionality reduction and normalization module 3, which is used by the face filtering tree module 5 to generate the face filtering difference tree. Because the face database is a large database, only the changes and differences are modified to reduce the central computation load.

[0115] When the face database changes, the face filtering tree module 5 generates a face filtering differential array and sends it to the face pre-filtering module 6 of the 5G recorder.

[0116] The new face filtering differential tree only includes the changed parts, reducing the amount of intermediate network data transmission and improving the overall system availability.

[0117] The steps of generating the face filtering difference array in the face filtering tree module 5 are as follows:

[0118] 6.1 First, calculate the face feature string to be added or deleted. If a new face feature string is added, store each character of the face feature string as a node in the tree.

[0119] 6.2 If character node A already exists in a face filtering tree, use the character node in the face filtering tree; otherwise, create a new character node, put the new node into the difference array, and add the new node to the face filtering tree.

[0120] 6.3 If a face feature string deletion operation is performed, traverse the difference tree from the root node to find the character path of the face feature string, and check in reverse order whether the character node B of the character path has child nodes;

[0121] 6.4 If there are no child nodes, then the character node B is the deletion node, the deletion node is placed in the difference array, and the node is deleted from the face filtering tree.

[0122] 6.5 Face Filtering Tree Module 5 sends the differential array to the 5G recorder's face pre-filtering module 6.

[0123] The pre-detection filtering method in the face pre-filtering module 6 is as follows:

[0124] 7.1 If the length of the face feature matching string is less than the previous matching threshold, the pre-detection filtering will fail, and the face will definitely not be in the face database.

[0125] 7.2 If the length of the face feature matching string is greater than or equal to the previous matching threshold, the pre-detection filtering is successful, and the face may be in the face database.

[0126] The method for obtaining the data comparison library in the face feature segmentation module 7:

[0127] 8.1 Sort the normalized values ​​of the one-dimensional facial features;

[0128] 8.2 According to the square root of the face database size N The facial feature values ​​are divided into blocks to obtain the final result.

[0129] Individual facial feature segmentation.

[0130] 8.3 The normalized values ​​of the starting and ending face feature 1D spaces of each block are used as a comparison database.

[0131] The normalized values ​​of the starting and ending face feature 1D spaces for each block are used by the face layer comparison module 8 to find face feature blocks. The face layer comparison module 8 then performs face comparison within the face feature blocks, greatly reducing the computational load.

[0132] The face layer comparison module 8 uses a binary search method to find face feature blocks.

[0133] The face layer comparison module 8 uses a layered method to perform face matching:

[0134] 10.1 The face layer comparison module 8 uses the normalized value of the face feature 1D space of the face feature matching string face feature block, extracts the 128-dimensional face feature vector corresponding to the normalized value of the 1D space of all matched face features, and performs cosine similarity calculation with the 128-dimensional face feature vector to be compared. If the cosine similarity calculation similarity is higher than the threshold, it is considered that a matching face has been found.

[0135] 10.2 If the 128-dimensional face feature vectors of this batch of face feature blocks are not higher than the threshold, then the face layer comparison module 8 removes the end of the face feature matching string, and then matches the normalized value of the face feature 1-dimensional space of the face feature blocks to expand the range of matching 128-dimensional face feature vectors, and then calculates the matching using the cosine similarity method.

[0136] 10.3 If the length of the face feature matching string is less than the previous matching threshold, and a 128-dimensional face feature vector is still not found, then the face matching is considered to have failed.

[0137] Working principle: This application sends the changed 128-dimensional face feature vector to the face feature dimensionality reduction and normalization module 3 after the face database module 4 changes (addition and deletion), and uses it to generate a face filtering difference tree in the face filtering tree module (5). Since the face database is a large database, only the changes and differences are modified, reducing the central computation load. The new face filtering difference tree only contains the changed part, reducing the amount of intermediate transmission network data and improving the availability of the whole system. The normalized value of the starting face feature 1D space and the normalized value of the ending face feature 1D space of each block are used by the face layer comparison module 8 to find the face feature block. The face layer comparison module 8 compares faces in the face feature block, greatly reducing the computation load. The face feature matching string is sent to the face layer comparison module 8 to reduce the number of face comparisons in the face layer comparison module 8. The application itself uses the cooperation between the above modules to perform face comparison filtering on the front-end device, filtering out faces that are not in the face database, thereby reducing the pressure on the center.

[0138] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0139] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A 5G recorder face recognition pre-filtering system, characterized in that, include: The image capture module (1) is used to capture video images, generate image frames, and send the image frames to the face feature extraction module (2); The face feature extraction module (2) is used to receive image frames sent by the image capture module (1), extract face features using a feature extractor, encode the feature data into a 128-dimensional face feature vector, and then send the 128-dimensional face feature vector to the face feature dimensionality reduction and normalization module (3); the face feature dimensionality reduction and normalization module (3) is used to receive the 128-dimensional face feature vectors from the face feature extraction module (2) and the face database module (4), and calculate and generate the normalization value of the face feature 1-dimensional space respectively. The normalization value of the face feature 1-dimensional space obtained based on the data in the face feature extraction module (2) and the 128-dimensional face feature vector in the face feature extraction module (2) are sent to the face pre-filtering module (6), and the normalization value of the face feature 1-dimensional space obtained based on the data in the face feature extraction module (2) and the 128-dimensional face feature vector in the face feature extraction module (2) are sent to the face pre-filtering module (6), and the face feature vector is sent to the face database module (6). The normalized value of the 1D space of facial features obtained from the data in module (4) and the 128-dimensional facial feature vector in the face library module (4) are sent to the face feature segmentation module (7), and the normalized value of the 1D space of facial features obtained from the data in the face library module (4) is sent separately to the face filtering tree module (5); the face library module (4) is used to store the 128-dimensional facial feature vector with labeled personnel information, and sends the 128-dimensional facial feature vector to the face feature dimensionality reduction and normalization module (3); the face filtering tree module (5) is used to receive the normalized value of the 1D space of facial features in the face feature dimensionality reduction and normalization module (3), establish the initial face filtering tree, and send the initial face filtering tree to the face pre-filtering module (6); The face pre-filtering module (6) is used to receive the face filtering initial tree data from the face filtering tree module (5), reconstruct the face filtering tree, and simultaneously receive the normalized value of the face feature 1D space sent by the face feature reduction and normalization module (3) and the 128-dimensional face feature vector in the face feature extraction module (2). It converts the normalized value of the face feature 1D space into a face feature string, compares the face feature string with the face filtering tree, generates a face feature matching string, and sends the face feature matching string and the 128-dimensional face feature vector to the face layer comparison module (8). The face feature segmentation module (7) is used to receive the normalized value of the face feature 1D space and the 128-dimensional face feature vector sent by the face feature reduction and normalization module (3), and then performs face feature segmentation on the face feature value. The face features are divided into blocks, and the normalized values ​​of the starting and ending face feature 1D spaces of each block are sent to the face layer comparison module (8) to provide a comparison database for the face layer comparison module (8). The face layer comparison module (8) receives the face feature matching string and 128-dimensional face feature vector sent by the face pre-filtering module (6), and at the same time receives the normalized values ​​of the starting and ending face feature 1D spaces of each block sent by the face feature block module (7). The received contents from the two modules are used to perform face matching using a layered method to complete the confirmation of face matching. The face feature block module (7) sorts the normalized values ​​of the face feature 1D space and then sorts them according to the square root of the face database size N. The facial feature values ​​are divided into blocks to obtain the final result. The face features are divided into blocks, and the normalized values ​​of the starting and ending face feature 1D spaces of each block are used as a comparison database. When the face layering comparison module (8) performs face matching using the layering method, it uses face feature matching strings to match the normalized values ​​of the face feature 1D spaces of the face feature blocks, extracts the 128-dimensional face feature vectors corresponding to the normalized values ​​of all matched face feature 1D spaces, and performs cosine similarity calculation with the 128-dimensional face feature vectors to be compared. If the cosine similarity calculation is similar... If the similarity is higher than the preset threshold, it is considered that a matching face has been found. If the 128-dimensional face feature vector of this batch of face feature blocks is not higher than the threshold, the face layer comparison module (8) removes the end of the face feature matching string, then matches the normalized value of the face feature 1-dimensional space of the face feature blocks, expands the range of matching 128-dimensional face feature vectors, and then calculates the cosine similarity method. If the length of the face feature matching string is less than the previous matching threshold, and a suitable 128-dimensional face feature vector is still not found, it is considered that the face matching has failed.

2. A video inspection method based on a low-power 5G recorder, characterized in that, include: 1) The image capture module (1) captures video images, generates image frames, and sends the image frames to the face feature extraction module (2); 2) The face feature extraction module (2) receives the image frames sent by the image capture module (1), extracts face features using the ResNet feature extractor, encodes the feature data into a 128-dimensional face feature vector through the SVM feature encoder, and then sends the 128-dimensional face feature vector to the face feature dimension reduction and normalization module (3); 3) The face feature dimension reduction and normalization module (3) receives the 128-dimensional face feature vector from the face feature extraction module (2), calculates the normalization value in 1-dimensional space, and then uses the obtained value to calculate the normalization value in 1-dimensional space. The normalized value of the 1D space and the 128-dimensional face feature vector are sent to the face pre-filtering module (6); 4) The face feature dimensionality reduction and normalization module (3) receives the 128-dimensional face feature vector from the face database module (4), calculates the normalized value of the 1D space of the face features, and sends the normalized value of the 1D space of the face features to the face filtering tree module (5). The face feature dimensionality reduction and normalization module (3) sends the normalized value of the 1D space of the face features and the 128-dimensional face feature vector to the face feature segmentation module (7); 5) The face database module (4) stores the 128-dimensional face feature vector with labeled personnel information and sends the 128-dimensional face feature vector to the face feature segmentation module (7). The dimensionality reduction and normalization module (3) and the face database module (4) generate all 128-dimensional face feature vectors for the face feature dimensionality reduction and normalization module (3); 6) The face filtering tree module (5) receives the normalization value of the face feature 1-dimensional space from the face feature dimensionality reduction and normalization module (3); 7) The face filtering tree module (5) generates the initial face filtering tree and the face filtering difference array using the normalization value of the face feature 1-dimensional space; 8) The face filtering tree module (5) establishes the initial face filtering tree with all data in the face database, and sends the initial face filtering tree to the face pre-filtering module (6); 9) The face pre-filtering module (6) receives the face filtering data from the face filtering tree module (5). Filter the data of the initial tree and reconstruct the face filtering tree; 10) The face pre-filtering module (6) receives the normalized value of the face feature 1D space and the 128-dimensional face feature vector from the face feature dimensionality reduction and normalization module (3); 11) The face pre-filtering module (6) converts the normalized value of the face feature 1D space into a face feature string, and compares it one by one on the face filtering tree according to the string of the face feature string. When no matching degree node is found, the query is terminated and a face feature matching string is obtained; 12) The face feature matching string and the 128-dimensional face feature vector are sent to the face layer comparison module (8); 13) The length of the face feature matching string is compared with the pre-matching threshold to perform pre-detection filtering;14) The face feature segmentation module (7) receives the normalized value of the face feature 1D space and the 128-dimensional face feature vector from the face feature dimensionality reduction and normalization module (3), segments the face feature values, and provides a comparison database for the face layer comparison module (8); 15) The face layer comparison module (8) receives the face feature matching string and the 128-dimensional face feature vector from the face pre-filtering module (6); 16) The face layer comparison module (8) compares the received face feature matching string with the 128-dimensional face feature vector, finds the face feature segment, and completes the face matching; The method for obtaining the data comparison database in the face feature segmentation module (7) is as follows: 8.1) Sort the normalized value of the face feature 1D space; 8.2) According to the square root of the face database size N The facial feature values ​​are divided into blocks to obtain the final result. 8.3) The normalized values ​​of the starting and ending face feature 1D spaces of each segment are used as a comparison database; The face layer comparison module (8) uses a layered method to perform face matching: 10.1) The face layer comparison module (8) uses face feature matching strings to match the normalized values ​​of the face feature 1D spaces of the face feature segments, extracts the 128-dimensional face feature vectors corresponding to the normalized values ​​of all matched face feature 1D spaces, and performs cosine similarity calculation with the 128-dimensional face feature vectors to be compared. If the cosine similarity is... If the similarity calculated by degree is higher than the preset threshold, it is considered that a matching face has been found; 10.2) If the 128-dimensional face feature vector of this batch of face feature blocks is not higher than the threshold, the face layer comparison module (8) removes the end of the face feature matching string, then matches the normalized value of the face feature 1-dimensional space of the face feature blocks, expands the range of matching 128-dimensional face feature vectors, and then performs cosine similarity matching calculation; 10.3) If the length of the face feature matching string is less than the previous matching threshold, and a suitable 128-dimensional face feature vector is still not found, it is considered that the face matching has failed.

3. The video inspection method based on a low-power 5G recorder according to claim 2, characterized in that, The face feature dimensionality reduction and normalization module (3) receives a 128-dimensional face feature vector and performs the following steps to calculate and generate a normalized value for the 1-dimensional space of face features: 3.1) The face feature dimensionality reduction and normalization module (3) first uses random projection to project the 128-dimensional face feature vector onto a 2-dimensional plane space, and randomly generates a 128×2 projection matrix. Each element of the matrix is ​​randomly sampled from the standard normal distribution N(0,1); 3.2) The face feature dimensionality reduction and normalization module (3) then takes the two values ​​of the 2-dimensional vector as the endpoint of the plane vector A, the coordinate center (0,0) as the starting point of the plane vector A, and takes the counterclockwise angle between the plane vector A and the positive x-axis as the normalized value of the 1-dimensional space.

4. The video inspection method based on a low-power 5G recorder according to claim 2, characterized in that, The steps of the face filtering tree module (5) to generate the initial face filtering tree are as follows: 4.1) The face filtering tree module (5) sets the maximum tree depth to H; 4.2) The normalized value of the face feature 1D space is converted into a face feature string, the output of the normalized value of the face feature 1D space is a decimal string, H-1 decimal places are taken, and the character dot "." in the string is removed to obtain a face feature string of length H; 4.3) Store the facial feature string in a trie. A root node is created, which does not store any characters. Each character of the facial feature string is used as a node in the tree and stored sequentially in the tree. The sequence from the root node to each page node is a facial feature string. 4.4) Use the initial face filtering tree as the face filtering tree.

5. A video inspection method based on a low-power 5G recorder according to claim 2, characterized in that, When the face database changes, the face filtering tree module (5) will generate a face filtering differential array and send it to the face pre-filtering module (6) of the 5G recorder.

6. A video inspection method based on a low-power 5G recorder according to claim 5, characterized in that, The face filtering tree module (5) generates the face filtering difference array in the following steps: 6.1) First, calculate the face feature string to be added or deleted. If the face feature string is added, each character of the face feature string is used as a node of the tree and stored in the tree in sequence. 6.2) If character node A already exists in a face filter tree, use the character node of the face filter tree; otherwise, create a new character node, put the new node into the difference array, and add the new node to the face filter tree. 6.3) If a face feature string deletion operation is performed, traverse the difference tree from the root node to find the character path of the face feature string, and check in reverse order whether the character node B of the character path has child nodes. 6.4) If there are no child nodes, then the character node B is the deletion node, the deletion node is placed in the difference array, and the node is deleted from the face filtering tree; 6.5) The face filtering tree module (5) sends the difference array to the face pre-filtering module (6) of the 5G recorder.

7. A video inspection method based on a low-power 5G recorder according to claim 2, characterized in that, The pre-detection filtering method in the face pre-filtering module (6) is as follows: 7.1) If the length of the face feature matching string is less than the preset pre-matching threshold, the pre-detection filtering fails and the face is definitely not in the face database; 7.2) If the length of the face feature matching string is greater than or equal to the pre-matching threshold, the pre-detection filtering succeeds and the face may be in the face database.

8. A video inspection method based on a low-power 5G recorder according to claim 2, characterized in that, The face layer comparison module (8) uses a binary search method to find face feature blocks.

Images