Face recognition system and method based on self-learning prediction model

Abstract

The present invention provides a face recognition system and method based on a self-learning prediction model, comprising: a self-learning strategy generator module, a portrait photo generator module, an automatic calculation module, a pass rate analyzer module, an algorithm acquisition strategy module, and an algorithm-specific portrait database module; the self-learning strategy generator module is combined with the portrait photo generator module to generate multiple sets of multi-type portrait photo databases, which are passed through the automatic calculation module and then through the pass rate analyzer module to determine the portrait quality photo parameter set calculation model of the most matching third-party algorithm platform, and a portrait photo acquisition channel is established according to the most matching portrait quality photo parameter set calculation model, and finally the collected portrait photos are converted into portrait feature values ​​for encrypted storage to form a portrait database exclusive to the third-party algorithm platform. The present invention improves the recognition rate of portrait data used by each algorithm platform, and fundamentally solves the problems of low portrait acquisition pass rate and low recognition rate.

Description

Technical Field

[0001] The present invention relates to the technical field of face recognition, and in particular to a face recognition system and method based on a self-learning prediction model. Background Art

[0002] In the context of digital construction, the application of face recognition technology is becoming more and more extensive, including: campus security, access control, venue management, dormitory management, conference room sign-in and other scenarios, and has now penetrated into all aspects of people's work, study and life. The portrait feature data that supports face recognition technology has become particularly important in management as personal sensitive information. The security, identification and standardization of portrait data have become important issues faced in digital construction.

[0003] The current algorithms for portrait biometric recognition on the market are mainly based on the algorithm manufacturer's comprehensive calculation of the facial structure points of the face in the photo, and the portrait data model is obtained for comparison and recognition. Each algorithm manufacturer has a portrait recognition algorithm mechanism based on its own algorithm characteristics, and it is impossible to achieve algorithm mutual recognition and identification. In this way, when the equipment and systems of multiple portrait algorithm manufacturers are applied in many business scenarios, it will cause the system to collect face data and the collected portrait data photos will be given to each algorithm manufacturer, and it will not be possible to perfectly match them for use.

[0004] In response to the above problems, the present invention proposes a face recognition system and method based on a self-learning prediction model. By customizing and generating a variety of portrait algorithm photos, the face recognition algorithm of a third-party manufacturer is actively and repeatedly calculated. Finally, the calculation model that best matches the face recognition algorithm of a third-party manufacturer is obtained by analyzing the calculation model data (based on face real person detection, whether wearing glasses detection, left and right eye opening and closing detection, mask wearing detection, face up and down tilt angle detection, face photo proportion and position detection, face light detection, face occlusion detection, face overflow photo detection analysis). Then, based on this model, portrait collection can achieve high recognition and precise application of this portrait under this algorithm platform. Summary of the invention

[0005] The present invention provides a face recognition system and method based on a self-learning prediction model, which aims to solve the problem in the prior art that when the equipment and systems of multiple portrait algorithm manufacturers are applied in many business scenarios, the system will give the collected portrait data photos to each algorithm manufacturer when collecting face data, and the problem cannot be perfectly matched and used.

[0006] In order to solve the above problems, the present invention discloses a face recognition system based on a self-learning prediction model, comprising: a self-learning strategy generator module, a portrait photo generator module, an automatic measurement module, a pass rate analyzer module, an algorithm acquisition strategy module, and an algorithm-specific portrait database module;

[0007] The self-learning strategy generator module is used to generate detection items under different facial states based on portrait photos by establishing a self-learning strategy model, and independently measure each detection item one by one. The measurement method is to measure the standardization of the detection items through a large number of portrait photos, and record the parameter interval with the highest pass rate, where all parameter intervals are collectively referred to as an adaptive parameter set;

[0008] The portrait photo generator module is used to divide the adaptive parameter set of each detection item into segmented intervals, cross-combine the value intervals of each detection item to generate a plurality of power value parameter sets, establish a portrait photo collection channel that meets the parameter set requirements, collect portrait photos that meet the channel parameter requirements through the collection channel and record and store them;

[0009] The automatic measurement module is used to push the collected portrait photos to the interface of the third-party manufacturer's algorithm for automatic measurement to obtain the pass rate of the portrait photos of each parameter set;

[0010] The pass rate analyzer module is used to import the measured data and perform parameter average value analysis after the automatic measurement is completed, so as to obtain the most matching portrait quality photo parameter set measurement model identified by the third-party manufacturer's algorithm;

[0011] The algorithm acquisition strategy module is used to reversely establish an acquisition model through the best matching portrait quality photo parameter set measurement model. This model is the optimal portrait acquisition algorithm model;

[0012] The algorithm-specific portrait database module is used to establish a portrait photo collection channel exclusive to the third-party manufacturer's algorithm platform through the optimal portrait collection algorithm model and collect portrait photo data. Finally, the collected portrait photos are converted into portrait feature values ​​for encrypted storage to form an exclusive portrait database for the third-party algorithm platform.

[0013] Preferably, the detection items for different facial states in portrait photos include: real face detection, glasses wearing detection, left and right eye open and closed detection, mask wearing detection, up and down and left and right tilt angle detection of the face, proportion and position detection in the face photo, facial light detection, facial occlusion detection, and facial overflow photo detection.

[0014] As a preference, the ranges and testing standards for different parameter intervals are as follows:

[0015] Face detection: By capturing the facial structure points of the face, the detection capture range is within the range of 80-150 structure points, which is the effective capture range, that is, it is determined to be a real person photo;

[0016] Glasses detection: By judging the occlusion of facial structure points, ensuring that the recognizable range of the eyes and eye periorbital structure points is between 80% and 100%, it is determined that the person is not wearing glasses, otherwise it is determined that the person is wearing glasses;

[0017] Left and right eye opening and closing detection: The eye is judged by detecting the distance between the upper and lower eyelid structure points. If the distance between the upper and lower eyelid structure points is less than 1-3mm, and the pupil and iris structure points are not detected, the eye is judged as closed, otherwise it is judged as open.

[0018] Mask wearing detection: By detecting the occlusion of the chin and the surrounding structural points of the orbicularis oris muscle, if the occlusion rate reaches between 80% and 100%, it is determined that the person is wearing a mask.

[0019] As a preference, the ranges and testing standards for different parameter intervals are as follows:

[0020] Detection of the up-down and left-right tilt angle of the face: The facial structure points of the portrait are connected through the three-dimensional structure interconnection method to establish a three-dimensional portrait model. The horizontal coordinate deviation method is used on the model to measure the relative tilt angle of the facial structure points of the portrait, and then to calculate whether the real portrait is deviated in the photo. If the deviation tilt angle is greater than 10-25 degrees, the portrait is judged to be in a tilted state, otherwise it is judged to be in a non-tilted state;

[0021] Detection of face proportion and position in photos: Use image recognition technology to identify the position range of the face in the photo and analyze the proportion of the face in the photo. When the face area is less than 40% of the photo area, it is determined that the portrait is in an abnormal proportion state.

[0022] As a preference, the ranges and testing standards for different parameter intervals are as follows:

[0023] Face light detection: polarization, color level, color difference, exposure;

[0024] Facial occlusion detection: If the ratio of facial structure points reaches between 60% and 80%, it is considered facial occlusion;

[0025] Face overflow photo detection: If the completeness ratio of facial structure points is less than 90%, the portrait is judged to be overflowed or blocked.

[0026] Preferably, the 80-150 structural points of the face real person detection can be segmented into 3, 5 or 7 segments, wherein the 7 segments are: 80-90, 90-100, 100-110, 110-120, 120-130, 130-140, 140-150.

[0027] Preferably, the portrait photo generator module collects portrait photos by setting up a fixed collection device camera collector or a mobile collection device mobile phone camera to collect portrait photos of the experimental population.

[0028] Preferably, the interface of the third-party manufacturer's algorithm in the automatic measurement module is SDK, and the collected portrait photos are sequentially passed through the SDK of the third-party manufacturer's algorithm to record the recognition pass rate of the portrait photos, and the pass rate is recorded in the interval of 60%-100%; the highest pass rate and the lowest pass rate of the portrait photos in the interval of 60%-100% are deleted, and then the average value is calculated to obtain the average pass rate, and finally the portrait photo parameters closest to the average pass rate are selected as the portrait photo parameter set measurement model.

[0029] As a preference, when the algorithm SDK of a third-party manufacturer is upgraded or changed, this system also generates a collection channel and database exclusive to the new algorithm.

[0030] To solve the above problems, the present invention also discloses a face recognition method based on a self-learning prediction model, comprising the following steps:

[0031] A. Generate detection items under different facial states based on portrait photos by establishing a self-learning strategy model, and independently measure each detection item one by one. The measurement method is to measure the standardization of the detection items through a large number of portrait photos, and record the parameter interval with the highest pass rate. Among them, all parameter intervals are collectively referred to as adaptive parameter sets;

[0032] B. Taking values ​​of the adaptive parameter sets of each test item in segmented intervals, cross-combining the value intervals of each test item to generate a plurality of power value parameter sets, establishing a portrait photo collection channel that meets the parameter set requirements, and collecting portrait photos that meet the channel parameter requirements through the collection channel and recording and storing them;

[0033] C. Push the collected portrait photos to the interface of the third-party manufacturer's algorithm for automatic measurement to obtain the pass rate of the portrait photos of each parameter set;

[0034] D. After the automatic measurement is completed, the measured data is imported and analyzed through the rate analyzer module for parameter average value, so as to obtain the most matching portrait photo parameter set measurement model recognized by the third-party manufacturer's algorithm;

[0035] E. Reversely establish the acquisition model through the best matching portrait photo parameter set measurement model, which is the optimal portrait acquisition algorithm model;

[0036] F. Establish a portrait photo collection channel exclusive to the third-party manufacturer's algorithm platform through the optimal portrait collection algorithm model and collect portrait photo data. Finally, convert the collected portrait photos into portrait feature values ​​for encrypted storage to form an exclusive portrait database for the third-party algorithm platform.

[0037] The beneficial effects of the present invention are as follows:

[0038] The present invention repeatedly measures third-party algorithms through a portrait photo generator, establishes the highest recognition and precision portrait recognition model for each algorithm platform, and establishes a standardized portrait acquisition strategy through the model; this method is applied to various business scenarios for self-learning measurement, and can automatically analyze and establish individual algorithm acquisition strategies; through strategy acquisition, a high-recognition and precision exclusive portrait database for each algorithm platform is established, including: photo data, portrait feature value data; a standardized interface is established, and the automatic measurement of the portrait recognition model of the newly added algorithm and the establishment and generation of acquisition strategies can be realized through the interface. The present invention improves the recognition rate of portrait data used by each algorithm platform, and fundamentally solves the problems of low portrait acquisition pass rate and low recognition rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the system flow of the present invention;

[0040] Figure 2 It is a schematic diagram of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0043] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0044] like Figure 1 As shown, the present invention provides a face recognition system based on a self-learning prediction model, including: a self-learning strategy generator module, a portrait photo generator module, an automatic measurement module, a pass rate analyzer module, an algorithm acquisition strategy module, and an algorithm-specific portrait database module;

[0045] The self-learning strategy generator module is used to generate detection items under different facial states based on portrait photos by establishing a self-learning strategy model, and independently measure each detection item one by one. The measurement method is to measure the standardization of the detection items through a large number of portrait photos, and record the parameter interval with the highest pass rate, where all parameter intervals are collectively referred to as an adaptive parameter set;

[0046] The portrait photo generator module is used to divide the adaptive parameter set of each detection item into segmented intervals, cross-combine the value intervals of each detection item to generate a plurality of power value parameter sets, establish a portrait photo collection channel that meets the parameter set requirements, collect portrait photos that meet the channel parameter requirements through the collection channel and record and store them;

[0047] The automatic measurement module is used to push the collected portrait photos to the interface of the third-party manufacturer's algorithm for automatic measurement to obtain the pass rate of the portrait photos of each parameter set;

[0048] The pass rate analyzer module is used to import the measured data and perform parameter average value analysis after the automatic measurement is completed, so as to obtain the most matching portrait quality photo parameter set measurement model identified by the third-party manufacturer's algorithm;

[0049] The algorithm acquisition strategy module is used to reversely establish an acquisition model through the best matching portrait quality photo parameter set measurement model. This model is the optimal portrait acquisition algorithm model;

[0050] The algorithm-specific portrait database module is used to establish a portrait photo collection channel exclusive to the third-party manufacturer's algorithm platform through the optimal portrait collection algorithm model and collect portrait photo data. Finally, the collected portrait photos are converted into portrait feature values ​​for encrypted storage to form an exclusive portrait database for the third-party algorithm platform.

[0051] Furthermore, the detection items for different facial states in portrait photos include: face real person detection, whether wearing glasses detection, left and right eye open and closed detection, mask wearing detection, face up and down and left and right tilt angle detection, face proportion and position detection in the photo, face light detection, face occlusion detection, and face overflow photo detection.

[0052] Furthermore, the ranges and detection criteria of different parameter intervals are as follows:

[0053] Face detection: By capturing the facial structure points of the face, the detection capture range is within the range of 80-150 structure points, which is the effective capture range, that is, it is determined to be a real person photo;

[0054] Glasses detection: By judging the occlusion of facial structure points, ensuring that the recognizable range of the eyes and eye periorbital structure points is between 80% and 100%, it is determined that the person is not wearing glasses, otherwise it is determined that the person is wearing glasses;

[0055] Left and right eye opening and closing detection: The eye is judged by detecting the distance between the upper and lower eyelid structure points. If the distance between the upper and lower eyelid structure points is less than 1-3mm, and the pupil and iris structure points are not detected, the eye is judged as closed, otherwise it is judged as open.

[0056] Mask wearing detection: By detecting the occlusion of the chin and the surrounding structural points of the orbicularis oris muscle, if the occlusion rate reaches between 80% and 100%, it is determined that the person is wearing a mask.

[0057] Furthermore, the ranges and detection criteria of different parameter intervals are as follows:

[0058] Detection of the up-down and left-right tilt angle of the face: The facial structure points of the portrait are connected through the three-dimensional structure interconnection method to establish a three-dimensional portrait model. The horizontal coordinate deviation method is used on the model to calculate the relative tilt angle of the facial structure points of the portrait, and then calculate whether the real portrait is deviated in the photo. If the deviation tilt angle is greater than 10-25 degrees, the portrait is judged to be in a tilted state, otherwise it is judged to be in a non-tilted state.

[0059] Detection of face proportion and position in photos: Use image recognition technology to identify the position range of the face in the photo and analyze the proportion of the face in the photo. When the face area is less than 40% of the photo area, it is determined that the portrait is in an abnormal proportion state.

[0060] Furthermore, the ranges and detection criteria of different parameter intervals are as follows:

[0061] Face light detection: polarization, color level, color difference, exposure;

[0062] Facial occlusion detection: If the ratio of facial structure points reaches between 60% and 80%, it is considered facial occlusion;

[0063] Face overflow photo detection: If the completeness ratio of facial structure points is less than 90%, the portrait is judged to be overflowed or blocked.

[0064] Furthermore, the 80-150 structural points of the face real person detection can be segmented into 3, 5 or 7 segments, wherein the 7 segments are: 80-90, 90-100, 100-110, 110-120, 120-130, 130-140, 140-150.

[0065] Furthermore, the portrait photo generator module collects portrait photos by setting up a fixed collection device, a camera collector, or a mobile collection device, a mobile phone camera, to collect portrait photos of the experimental group.

[0066] Furthermore, the interface of the third-party manufacturer's algorithm in the automatic measurement module is SDK, and the collected portrait photos are passed through the SDK of the third-party manufacturer's algorithm in turn to record the recognition pass rate of the portrait photos, and the pass rate is recorded in the interval of 60%-100%; the highest pass rate and the lowest pass rate of the portrait photos in the interval of 60%-100% are deleted, and then the average value is calculated to obtain the average pass rate, and finally the portrait photo parameters closest to the average pass rate are selected as the portrait photo parameter set measurement model.

[0067] Furthermore, when the algorithm SDK of a third-party manufacturer is upgraded or changed, this system also generates a collection channel and database exclusive to the new algorithm.

[0068] The present invention also proposes a face recognition method based on a self-learning prediction model, comprising the following steps:

[0069] A. Generate detection items under different facial states based on portrait photos by establishing a self-learning strategy model, and independently measure each detection item one by one. The measurement method is to measure the standardization of the detection items through a large number of portrait photos, and record the parameter interval with the highest pass rate. Among them, all parameter intervals are collectively referred to as adaptive parameter sets;

[0070] B. Taking values ​​of the adaptive parameter sets of each test item in segmented intervals, cross-combining the value intervals of each test item to generate a plurality of power value parameter sets, establishing a portrait photo collection channel that meets the parameter set requirements, and collecting portrait photos that meet the channel parameter requirements through the collection channel and recording and storing them;

[0071] C. Push the collected portrait photos to the interface of the third-party manufacturer's algorithm for automatic measurement to obtain the pass rate of the portrait photos of each parameter set;

[0072] D. After the automatic measurement is completed, the measured data is imported and analyzed through the rate analyzer module for parameter average value, so as to obtain the most matching portrait photo parameter set measurement model recognized by the third-party manufacturer's algorithm;

[0073] E. Reversely establish the acquisition model through the best matching portrait photo parameter set measurement model, which is the optimal portrait acquisition algorithm model;

[0074] F. Establish a portrait photo collection channel exclusive to the third-party manufacturer's algorithm platform through the optimal portrait collection algorithm model and collect portrait photo data. Finally, convert the collected portrait photos into portrait feature values ​​for encrypted storage to form an exclusive portrait database for the third-party algorithm platform.

[0075] like Figure 2 As shown, the specific embodiment:

[0076] S1. Generate a calculation strategy based on photo face real person detection, glasses detection, left and right eye open and closed detection, mask detection, face up and down tilt angle detection, face proportion and position detection, face light detection, face occlusion detection, face overflow photo detection and other parameters by establishing a self-learning strategy model;

[0077] S2. Take values ​​of the adaptive parameter sets of each detection item in segmented intervals (take values ​​according to the average ratio interval, for example, the 80-150 structural points of face real person detection can be segmented by 3, 5, and 7, for example, 7 segments are: 80-90, 90-100, 100-110, 110-120, 120-130, 130-140, 140-150), cross-combine the value intervals of each detection item (random combination of parameters of multiple detection items) to generate many power value parameter sets, such as the Hikvision algorithm test parameter set: face real person detection (facial structure points Range 100-140), whether to wear glasses detection (periorbital structure point range 20-80), left and right eye opening and closing detection (upper and lower eyelid distance 1-5), mask detection (lower half of the face structure point range 50-100), face up and down left and right tilt angle detection (deviation angle 1-180), face proportion and position detection in the photo (proportion position 1:2-1:1.5), face light detection (brightness detection 1-100), face occlusion detection (facial structure points 100-150), face overflow photo detection (facial structure points 100-150);

[0078] S3. Establish a portrait photo collection channel that meets the parameter set requirements (mainly to collect a large amount of sample data to support the subsequent verification and measurement pass rate stability calculation. Construction process: develop an H5 user photo entrance -> user authorization -> user photo -> system storage data), such as: hang the collection H5 page in WeChat for Enterprise, open the H5 collection entrance through WeChat for Enterprise, click the collection button to take a photo to collect portrait data. Collect portrait photos that meet the channel parameter requirements through the collection channel and record and store them;

[0079] S4. Collect portrait photos. Collect portrait photos of the experimental population by setting up fixed collection equipment (camera collector) or mobile (mobile phone camera self-service collection) collection equipment.

[0080] S5. Calculate the pass rate of the third-party algorithm SDK. Pass the collected portrait photos through the third-party algorithm SDK in turn and record the recognition pass rate of the photos. The pass rate is recorded in intervals such as 60%-100%.

[0081] S6. Analysis of the stability of the pass rate: To ensure the "central tendency" and data stability of the calculation, the photos in the 60%-100% interval are deleted with the highest and lowest pass rates, and then the average pass rate is calculated. The parameters of the photos closest to the average pass rate are selected as the model parameters.

[0082] S7. Based on the standard project parameters of the photos with average passing rate, a portrait quality photo parameter set measurement model is established.

[0083] Each standard item parameter is verified through the Hikvision portrait algorithm interface to obtain the pass rate of each standard item parameter. The set of each standard item parameter is obtained through verification of a large number of sample data. The pass rate stability of each set of standard item parameters is then measured to obtain the average value of each item. The average values ​​of each item are then combined to form the following quality photo parameter set measurement model: For example, the average pass rate of each test item is A1...AN, and the average pass rate of the Hikvision algorithm is M. The calculation formula for M is as follows:

[0084]

[0085] Based on the above formula, it can be calculated that the average pass rate of Hikvision portrait algorithm is 89%. The measurement model of the quality photo parameter set with a pass rate of 89% is: face real person detection (facial structure point interval 110-120), whether wearing glasses detection (eye structure point interval 50-70), left and right eye open and closed detection (upper and lower eyelid distance 2-3), mask detection (lower half of the face structure point interval 70-90), face up and down tilt angle detection (deviation angle 1-90), face proportion and position detection in the photo (proportion position is 1:1.7-1:1.5), face light detection (brightness detection 30-60), face occlusion detection (facial structure points 110-120), face overflow photo detection (facial structure points 110-120); based on this model, the portrait acquisition channel of S8 is established.

[0086] S8. Based on the portrait quality photo parameter set measurement model, establish a portrait collection channel, such as: hang the collection H5 page in the enterprise WeChat, open the H5 collection entrance through the enterprise WeChat, click the collection button to take a photo to collect portrait data.

[0087] S9. Collect portrait photos of the target population by setting up fixed collection equipment (such as camera collectors) or mobile (self-service collection by mobile phone cameras) collection equipment.

[0088] S10. Record the collection channel number (i.e., H5 collection entry number) and data and encrypt and store the portrait photo data. Establish an algorithm-specific collection channel and database, such as: hang the collection H5 page in WeChat for Business, open the H5 collection entry through WeChat for Business, click the collection button to take a photo to collect portrait data, and encrypt and store the data through SM4+SM2 after data collection.

[0089] When the third-party algorithm SDK is upgraded or changed, the system executes S11-S16 to generate a new algorithm-specific collection channel and database.

[0090] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A face recognition system based on a self-learning prediction model, characterized in that: include: Self-learning strategy generator module, portrait photo generator module, automatic measurement module, pass rate analyzer module, algorithm acquisition strategy module, algorithm-specific portrait database module; The self-learning strategy generator module is used to generate detection items under different facial states based on portrait photos by establishing a self-learning strategy model, and independently measure each detection item one by one. The measurement method is to measure the standardization of the detection items through a large number of portrait photos, and record the parameter interval with the highest pass rate, where all parameter intervals are collectively referred to as an adaptive parameter set; The portrait photo generator module is used to divide the adaptive parameter set of each detection item into segmented intervals, cross-combine the value intervals of each detection item to generate a plurality of power value parameter sets, establish a portrait photo collection channel that meets the parameter set requirements, collect portrait photos that meet the channel parameter requirements through the collection channel and record and store them; The automatic measurement module is used to push the collected portrait photos to the interface of the third-party manufacturer's algorithm for automatic measurement to obtain the pass rate of the portrait photos of each parameter set; The pass rate analyzer module is used to import the measured data and perform parameter average value analysis after the automatic measurement is completed, so as to obtain the most matching portrait quality photo parameter set measurement model identified by the third-party manufacturer's algorithm; The algorithm acquisition strategy module is used to reversely establish an acquisition model through the best matching portrait quality photo parameter set measurement model. This model is the optimal portrait acquisition algorithm model; The algorithm-specific portrait database module is used to establish a portrait photo collection channel exclusive to the third-party manufacturer's algorithm platform through the optimal portrait collection algorithm model and collect portrait photo data. Finally, the collected portrait photos are converted into portrait feature values ​​for encrypted storage to form an exclusive portrait database for the third-party algorithm platform.

2. A face recognition system based on a self-learning prediction model as claimed in claim 1, characterized in that: The detection items for different facial states in portrait photos include: real face detection, glasses wearing detection, left and right eye open or closed detection, mask wearing detection, up and down, left and right tilt angle detection of the face, proportion and position detection in the face photo, facial light detection, facial occlusion detection, and facial overflow photo detection.

3. A face recognition system based on a self-learning prediction model as claimed in claim 2, characterized in that: The ranges and testing standards for different parameter intervals are as follows: Face detection: By capturing the facial structure points of the face, the detection capture range is within the range of 80-150 structure points, which is the effective capture range, that is, it is determined to be a real person photo; Glasses detection: By judging the occlusion of facial structure points, ensuring that the recognizable range of the eyes and eye periorbital structure points is between 80% and 100%, it is determined that the person is not wearing glasses, otherwise it is determined that the person is wearing glasses; Left and right eye opening and closing detection: The eye is judged by detecting the distance between the upper and lower eyelid structure points. If the distance between the upper and lower eyelid structure points is less than 1-3mm, and the pupil and iris structure points are not detected, the eye is judged as closed, otherwise it is judged as open. Mask wearing detection: By detecting the occlusion of the chin and the surrounding structural points of the orbicularis oris muscle, if the occlusion rate reaches between 80% and 100%, it is determined that the person is wearing a mask.

4. A face recognition system based on a self-learning prediction model as claimed in claim 2 or 3, characterized in that: The ranges and testing standards for different parameter intervals are as follows: Detection of the up-down and left-right tilt angle of the face: The facial structure points of the portrait are connected through the three-dimensional structure interconnection method to establish a three-dimensional portrait model. The horizontal coordinate deviation method is used on the model to measure the relative tilt angle of the facial structure points of the portrait, and then to calculate whether the real portrait is deviated in the photo. If the deviation tilt angle is greater than 10-25 degrees, the portrait is judged to be in a tilted state, otherwise it is judged to be in a non-tilted state; Detection of face proportion and position in photos: Use image recognition technology to identify the position range of the face in the photo and analyze the proportion of the face in the photo. When the face area is less than 40% of the photo area, it is determined that the portrait is in an abnormal proportion state.

5. A face recognition system based on a self-learning prediction model as claimed in claim 4, characterized in that: The ranges and testing standards for different parameter intervals are as follows: Face light detection: polarization, color level, color difference, exposure; Facial occlusion detection: If the ratio of facial structure points reaches between 60% and 80%, it is considered facial occlusion; Face overflow photo detection: If the completeness ratio of facial structure points is less than 90%, the portrait is judged to be overflowed or blocked.

6. A face recognition system based on a self-learning prediction model as claimed in claim 3, characterized in that: The 80-150 structural points of the face real person detection can be segmented into 3, 5 or 7 segments, wherein the 7 segments are: 80-90, 90-100, 100-110, 110-120, 120-130, 130-140, 140-150.

7. A face recognition system based on a self-learning prediction model as claimed in claim 1, characterized in that: The portrait photo generator module collects portrait photos by setting up a fixed collection device camera collector or a mobile collection device mobile phone camera to collect portrait photos of the experimental group.

8. A face recognition system based on a self-learning prediction model as claimed in claim 1, characterized in that: The interface of the third-party manufacturer's algorithm in the automatic measurement module is SDK. The collected portrait photos are passed through the SDK of the third-party manufacturer's algorithm in turn to record the recognition pass rate of the portrait photos, and the pass rate is recorded in the range of 60%-100%; the highest pass rate and the lowest pass rate of the portrait photos in the 60%-100% interval are deleted, and then the average value is calculated to obtain the average pass rate, and finally the portrait photo parameters closest to the average pass rate are selected as the portrait photo parameter set measurement model.

9. A face recognition system based on a self-learning prediction model as claimed in claim 8, characterized in that: When the algorithm SDK of a third-party manufacturer is upgraded or changed, this system also generates a collection channel and database exclusive to the new algorithm.

10. A face recognition method based on a self-learning prediction model, characterized in that: Using a face recognition system based on a self-learning prediction model as described in any one of claims 1 to 9, comprising the following steps: A. Generate detection items under different facial states based on portrait photos by establishing a self-learning strategy model, and independently measure each detection item one by one. The measurement method is to measure the standardization of the detection items through a large number of portrait photos, and record the parameter interval with the highest pass rate. Among them, all parameter intervals are collectively referred to as adaptive parameter sets; B. Taking values ​​of the adaptive parameter sets of each test item in segmented intervals, cross-combining the value intervals of each test item to generate a plurality of power value parameter sets, establishing a portrait photo collection channel that meets the parameter set requirements, and collecting portrait photos that meet the channel parameter requirements through the collection channel and recording and storing them; C. Push the collected portrait photos to the interface of the third-party manufacturer's algorithm for automatic measurement to obtain the pass rate of the portrait photos of each parameter set; D. After the automatic measurement is completed, the measurement data is imported into the pass rate analyzer module and the parameter average value analysis is performed to obtain the most matching portrait photo parameter set measurement model recognized by the third-party manufacturer's algorithm; E. Reversely establish the acquisition model through the best matching portrait photo parameter set measurement model, which is the optimal portrait acquisition algorithm model; F. Establish a portrait photo collection channel exclusive to the third-party manufacturer's algorithm platform through the optimal portrait collection algorithm model and collect portrait photo data. Finally, convert the collected portrait photos into portrait feature values ​​for encrypted storage to form an exclusive portrait database for the third-party algorithm platform.

Images