A risk control identity authentication system and method based on facial expression recognition

Abstract

The application discloses a risk control identity authentication system and method based on facial expression recognition, and relates to the technical field of image recognition.The system comprises an identity information acquisition module, a facial feature extraction module, a dynamic feature extraction accuracy analysis module, a multi-dimensional comparison module and a multi-dimensional comparison rationality analysis module.The application obtains the dynamic feature extraction accuracy evaluation by obtaining the feature extraction parameter quantization, judges whether to perform the dynamic feature extraction accuracy optimization, if yes, then performs the facial expression feature comparison link after optimization, otherwise directly performs the facial expression feature comparison link and obtains the multi-dimensional comparison rationality evaluation by obtaining the multi-dimensional comparison parameter quantization, and judges whether to perform the multi-dimensional comparison rationality optimization, thereby improving the risk control identity authentication reliability, and solving the problem of low risk control identity authentication reliability caused by inaccurate expression dynamic feature extraction in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, in particular to a risk control identity authentication system and method based on facial expression recognition. BACKGROUND

[0002] The user makes the specified expression according to the prompt (such as the screen display please smile and blink), the collection device captures the expression dynamic video stream or sequence image in real time, simultaneously judges whether the user is a real person through infrared imaging and 3D depth detection, prevents photo, video, mask and other forgery attacks, pre-processes the collected dynamic data (such as denoising, face alignment, key frame extraction), extracts the dynamic features of the expression such as facial muscle movement vector, expression duration, feature point time sequence change rule and the like and the static features such as facial geometric structure through computer vision algorithms such as CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory) models based on deep learning.

[0003] The real-time extracted expression features are compared with the facial expression-identity benchmark model stored during the user registration in multiple dimensions, the matching degree (such as dynamic feature time sequence consistency, key expression action similarity, feature point motion trajectory coincidence degree and the like) is calculated, if all the verification links are passed, the system feeds back to the user that the identity authentication is successful, and allows the user to continue to operate; if not, the failure reason such as insufficient expression matching degree and non-living body is explicitly prompted, and the user is guided to complete the identity verification through other security methods (such as SMS verification code, fingerprint recognition).

[0004] For example, the Chinese invention patent with the publication number CN105809111B discloses an identity authentication method, device, equipment and system, which comprises: acquiring the certificate information of a user; shooting the face photo of the user; judging whether the person facing the camera is a real person according to the head image of the user acquired by the camera in real time; and sending the certificate information and the face photo of the user to the authentication server in association under the condition that the person facing the camera is a real person.

[0005] For example, the Chinese invention patent with publication number CN108875646A discloses a real face image and identity card registration double comparison authentication method and system, which comprises: acquiring a current face image and a registration face image of a target human body; performing feature extraction on the current face image and the registration face image respectively to generate corresponding texture feature vectors; using a pre-set correction model to correct the texture feature vector of the registration face image to generate a corrected feature vector; calculating the difference degree of the texture feature vector of the current face image and the corrected feature vector, and judging whether the value of the difference degree is less than a preset threshold; when the value of the difference degree is less than the preset threshold, it is determined that the current face image and the registration face image are matched.

[0006] The above-mentioned technology at least has the following technical problems:

[0007] In the risk control identity authentication process, the expression dynamic features are extracted by CNN and LSTM. For the LSTM model, the core is to capture the time sequence change rule of the expression, which depends on the integrity of the continuous key frames. Fuzzy frames and lost frames caused by sudden changes in environmental light or network packet loss will destroy the continuity of the expression action, so that the LSTM model cannot accurately extract the time sequence change rule of the expression, and further causes deviation in dynamic feature extraction. The CNN model is easily affected by the appearance change of makeup. The difference between the bare face and the makeup, the beard and the glasses during user registration and identification can cause the face feature matching score to decrease. At the same time, the CNN model is sensitive to light and posture. Side light will cause distortion of key feature calculation, affecting the accurate extraction of static features by the CNN model, and further exacerbating the inaccuracy of dynamic feature extraction.

[0008] Multi-dimensional comparison has significant accuracy defects. On the one hand, the weight distribution mechanism is logically unbalanced. Some systems excessively rely on a single indicator (such as setting a high weight for “expression switching speed”), but do not consider the natural motion differences caused by user tension, fatigue or emotional fluctuations. On the other hand, the deformation of the chin-neck trajectory caused by the forward leaning of the head during sitting, and the physiological differences in pupil reflection intensity under different light, are all predictable scene variables. However, existing systems do not design a threshold dynamic adjustment mechanism for these variables, resulting in systematic bias in feature comparison results of the same user in different scenes. There is a problem of low reliability of risk control identity authentication caused by inaccurate extraction of expression dynamic features. SUMMARY

[0009] To solve the problem of low reliability of risk control identity authentication caused by inaccurate extraction of expression dynamic features in the prior art, the present application provides a risk control identity authentication system and method based on facial expression recognition. The technical solution is as follows:

[0010] In one aspect, a face expression recognition-based risk control identity authentication system is provided, comprising the following steps: an identity information acquisition module, a face feature extraction module, a dynamic feature extraction accuracy analysis module, a multi-dimensional comparison module, and a multi-dimensional comparison rationality analysis module; wherein the identity information acquisition module is configured to acquire a face image sequence of a user to be authenticated; the face feature extraction module is configured to preprocess the face image sequence to obtain feature extraction parameters to quantify the accuracy of CNN and LSTM models in the face expression feature extraction process, and obtain a dynamic feature extraction accuracy evaluation; the dynamic feature extraction accuracy analysis module is configured to determine whether to perform dynamic feature extraction accuracy optimization according to the dynamic feature extraction accuracy evaluation, and if so, to perform a face expression feature comparison link after optimization, or to directly perform the face expression feature comparison link, wherein the dynamic feature extraction accuracy optimization comprises dynamic adjustment of regional feature weights and dynamic adjustment of face region mask weights; the multi-dimensional comparison module is configured to obtain multi-dimensional comparison parameters in the face expression feature comparison link to quantify the rationality in the face expression feature comparison process, and obtain a multi-dimensional comparison rationality evaluation; the multi-dimensional comparison rationality analysis module is configured to determine whether to perform multi-dimensional comparison rationality optimization according to the multi-dimensional comparison rationality evaluation, and if so, to generate a risk control identity authentication decision after optimization, or to directly generate the risk control identity authentication decision, wherein the multi-dimensional comparison rationality optimization comprises dynamic adjustment of transition frame weights and dynamic adjustment of static-dynamic feature weights.

[0011] In another aspect, a face expression recognition-based risk control identity authentication method is provided, comprising: acquiring a face image sequence of a user to be authenticated; preprocessing the face image sequence to obtain feature extraction parameters to quantify the accuracy of CNN and LSTM models in the face expression feature extraction process, and obtain a dynamic feature extraction accuracy evaluation; determining whether to perform dynamic feature extraction accuracy optimization according to the dynamic feature extraction accuracy evaluation, and if so, performing a face expression feature comparison link after optimization, or directly performing the face expression feature comparison link, wherein the dynamic feature extraction accuracy optimization comprises dynamic adjustment of regional feature weights to improve the discrimination in the face expression feature extraction, and dynamic adjustment of face region mask weights to improve the accuracy of feature representation; obtaining multi-dimensional comparison parameters in the face expression feature comparison link to quantify the rationality in the face expression feature comparison process, and obtain a multi-dimensional comparison rationality evaluation; determining whether to perform multi-dimensional comparison rationality optimization according to the multi-dimensional comparison rationality evaluation, and if so, generating a risk control identity authentication decision after optimization, or directly generating the risk control identity authentication decision, wherein the multi-dimensional comparison rationality optimization comprises dynamic adjustment of transition frame weights to improve the matching accuracy of dynamic feature sequences in the expression switching process, and dynamic adjustment of static-dynamic feature weights to improve the consistency of multi-dimensional comparison with the real identity.

[0012] The one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0013] 1. By collecting the face image sequence of the user to be authenticated and pre-processing to obtain the feature extraction parameters, clear and standardized face image data can be obtained, providing high-quality input basis for subsequent face feature extraction and identity authentication, effectively reducing the influence of interference factors on authentication accuracy, quantizing the dynamic feature extraction accuracy evaluation according to the feature extraction parameters to represent the accuracy of the CNN and LSTM model in the face expression feature extraction process, judging whether to perform dynamic feature extraction accuracy optimization based on the dynamic feature extraction accuracy evaluation to dynamically control and improve the accuracy in the face expression feature extraction process, thereby improving the face feature extraction reliability, obtaining multi-dimensional comparison parameter quantization of the face expression feature comparison link to obtain multi-dimensional comparison rationality evaluation to represent the rationality in the face expression feature comparison process, and judging whether to perform multi-dimensional comparison rationality optimization based on the multi-dimensional comparison rationality evaluation to realize dynamic control and targeted optimization of the whole process of multi-dimensional comparison, thereby improving the risk control identity authentication reliability.

[0014] 2. According to the dynamic feature extraction accuracy evaluation, it is judged whether to perform dynamic feature extraction accuracy optimization, if yes, according to the brightness deviation, it is judged whether to perform regional feature weight dynamic adjustment, specifically, the brightness deviation reflects the uniformity of illumination in different regions of the face image, when some regions are too bright or too dark, the expression features of these regions will be covered or misread, thereby affecting the accuracy of feature extraction, by combining the brightness deviation to determine whether to adjust the regional feature weight, the model can give higher weight to the regions with appropriate illumination and clear features in the feature extraction process, so that they occupy a more important position in feature calculation, and appropriately reduce the weight of the regions with blurred features caused by brightness deviation, reduce the interference of these regions on the overall feature extraction result, thereby significantly improving the accuracy and discrimination of face expression feature extraction, and according to the mask entropy, it is judged whether to perform face region mask weight dynamic adjustment, which can optimize the focusing ability of the mask to the effective feature region, by improving the mask weight of the effective feature region and reducing the weight of the irrelevant region, the model can more accurately capture the key expression features and reduce the interference of redundant information, to improve the accuracy of feature representation, thereby improving the risk control identity authentication reliability.

[0015] 3. According to the multi-dimension comparison rationality evaluation, it is judged whether to perform multi-dimension comparison rationality optimization, if yes, according to the expression switching interval, it is judged whether to perform transition frame weight dynamic adjustment, which avoids the interference of invalid transition information and fully utilizes effective transition features, thereby significantly improving the rationality, accuracy and robustness of multi-dimension comparison, and providing more reliable support for accurate matching of facial expression features, and according to the single feature point positioning error, it is judged whether to perform static-dynamic feature weight dynamic adjustment, thereby avoiding invalid transition information interference, fully utilizing effective transition features, improving the rationality, accuracy of comparison, and further judging whether to perform dynamic adjustment of static-dynamic feature weight according to the single feature point positioning error, thereby comprehensively improving the reliability of risk control identity authentication. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0017] Figure 1 A structural schematic diagram of a risk control identity authentication system based on facial expression recognition provided by the embodiment of the present application;

[0018] Figure 2 A facial region mask weight dynamic adjustment flowchart of a risk control identity authentication system based on facial expression recognition provided by the embodiment of the present application;

[0019] Figure 3 A transition frame weight dynamic adjustment whole flowchart of a risk control identity authentication system based on facial expression recognition provided by the embodiment of the present application;

[0020] Figure 4 A risk control identity authentication method flowchart based on facial expression recognition provided by the embodiment of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the present application will be described below with reference to the drawings.

[0022] In the embodiments of the present application, the words such as "example", "for example" are used to represent an example, illustration or description. Any embodiment or design scheme described as "example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be one of the two.

[0023] The embodiment of the application provides a risk control identity authentication system and method based on facial expression recognition, and solves the problem of low risk control identity authentication reliability caused by inaccurate dynamic feature extraction in the prior art. The facial image sequence of a user to be authenticated is collected and preprocessed to obtain feature extraction parameter quantization, dynamic feature extraction accuracy evaluation, and dynamic feature extraction accuracy optimization. If the dynamic feature extraction accuracy optimization is performed, the facial expression feature comparison link is performed after optimization. Otherwise, the facial expression feature comparison link is directly performed, multi-dimensional comparison parameter quantization is obtained, multi-dimensional comparison rationality evaluation is obtained, and multi-dimensional comparison rationality optimization is determined. The risk control identity authentication reliability is improved.

[0024] The technical solution in the embodiment of the application solves the problem of low risk control identity authentication reliability caused by inaccurate dynamic feature extraction, and the general idea is as follows:

[0025] The facial image sequence of a user to be authenticated is collected and preprocessed to obtain feature extraction parameter quantization, dynamic feature extraction accuracy evaluation, and dynamic feature extraction accuracy optimization. If the dynamic feature extraction accuracy optimization is performed, the facial expression feature comparison link is performed after optimization. Otherwise, the facial expression feature comparison link is directly performed, multi-dimensional comparison parameter quantization is obtained, multi-dimensional comparison rationality evaluation is obtained, and multi-dimensional comparison rationality optimization is determined. The risk control identity authentication reliability is improved.

[0026] To make the technical problems, technical solutions and advantages of the application clearer, specific embodiments will be described in detail below with reference to the drawings.

[0027] As Figure 1As shown, it is a structural schematic diagram of a risk control identity authentication system based on facial expression recognition provided by an embodiment of the present application. The processing flow of the system can include the following steps: an identity information acquisition module, a facial feature extraction module, a dynamic feature extraction accuracy analysis module, a multi-dimensional comparison module, and a multi-dimensional comparison rationality analysis module. The output end of the identity information acquisition module is connected to the input end of the facial feature extraction module. The first output end of the facial feature extraction module is connected to the first input end of the multi-dimensional comparison module, and the second output end is connected to the input end of the dynamic feature extraction accuracy analysis module. The output end of the dynamic feature extraction accuracy analysis module is connected to the second input end of the multi-dimensional comparison module. The third input end of the multi-dimensional comparison module is connected to an identity authentication database, and the output end is connected to the input end of the multi-dimensional comparison rationality analysis module. The data end of the multi-dimensional comparison rationality analysis module is connected to the identity authentication database.

[0028] The identity information acquisition module is configured to acquire a facial image sequence of a user to be authenticated. The facial feature extraction module is configured to pre-process the facial image sequence to obtain a feature extraction parameter to quantify the accuracy of the CNN and LSTM models in the facial expression feature extraction process and obtain a dynamic feature extraction accuracy evaluation.

[0029] It should be noted that the data of the present application will be normalized, only considering the numerical results. The feature extraction parameter includes: key frame missing rate, single feature point displacement vector field, and frame interval time. The feature extraction parameter threshold and the feature extraction parameter correction factor are obtained from the constructed identity verification database. The feature extraction parameter threshold includes the key frame missing rate threshold, the single feature point displacement vector field threshold, and the frame interval time standard value. The feature extraction parameter correction factor includes the key frame missing rate correction factor, the single feature point displacement vector field correction factor, and the frame interval time correction factor. Specifically, the number of key frames is extracted by a video analysis tool such as FFmpeg. The absolute value of the difference between the number of key frames and the preset number of key frames is divided by the preset number of key frames, which is denoted as the key frame missing rate. The key frame refers to a special frame containing one frame of expression image information. The key feature points are located in the continuous video frames by a face detection algorithm such as MTCNN. The coordinates of the feature points in each frame are recorded. The direction of the feature points is calculated by the inverse tangent. Then the coordinate difference of the same feature point in adjacent two frames is calculated by the Euclidean distance formula to obtain the displacement vector of the single feature point, which is denoted as the single feature point displacement vector field. The time interval between adjacent two frames of images is obtained by synchronizing the hardware clock of the acquisition device, which is denoted as the frame interval time.

[0030] Regarding the above, it should be noted that to support the efficient operation of a facial expression recognition-based risk control and identity authentication system, a pre-built identity authentication database is constructed specifically for storing various key setting parameters, covering essential preset data such as keyframe missing rate, single feature point displacement vector field, and frame interval time. The initial values ​​of these parameters are not subjectively determined but are based on actual data. For example, the initial value of the keyframe missing rate threshold is calculated by summing and averaging the historical keyframe missing rates accumulated in the database, ensuring that the initial settings have objective basis and representativeness. Furthermore, considering the complexity of actual application scenarios and the dynamic changes in requirements, all parameter values ​​in the database remain adjustable. Technicians can manually set, adjust, and fine-tune these values ​​based on the system's performance during actual debugging, ultimately achieving optimal configuration of system parameters.

[0031] It needs to be explained that, firstly, the keyframe missing rate threshold and the proportion analysis results of the keyframe missing rate are corrected by a keyframe missing rate correction factor, denoted as the first component of accuracy. Here, the proportion analysis represents a division operation, and the specific constraint expression is as follows:

[0032] ;

[0033] In the formula, Q represents the first component of accuracy, P1 represents the keyframe missing rate correction factor obtained from the authentication database, A0 represents the keyframe missing rate threshold obtained from the authentication database, A represents the keyframe missing rate, and k1 represents a constant term to avoid meaningless numbers. The keyframe missing rate is negatively correlated with the accuracy evaluation of dynamic feature extraction; a higher keyframe missing rate means more keyframes are lost during video transmission, storage, or processing, resulting in a lower accuracy evaluation of dynamic feature extraction.

[0034] Next, the ratio analysis results of the single feature point displacement vector field and the single feature point displacement vector field threshold are corrected by the single feature point displacement vector field correction factor, which is denoted as the second component of accuracy. The specific constraint expression is as follows:

[0035] ;

[0036] In the formula, L represents the second component of accuracy, P2 represents the correction factor for the single feature point displacement vector field obtained from the authentication database, B represents the single feature point displacement vector field, and B0 represents the threshold of the single feature point displacement vector field obtained from the authentication database. The single feature point displacement vector field is positively correlated with the accuracy evaluation of dynamic feature extraction. A higher single feature point displacement vector field means that it can accurately describe the movement trajectory of various parts of the face, providing rich dynamic features such as velocity, acceleration, and direction changes, resulting in a higher accuracy evaluation for dynamic feature extraction.

[0037] Then, the proportion analysis result of the frame interval time standard value and the interval time deviation is corrected by a frame interval time correction factor, denoted as an accuracy third component, wherein the interval time deviation represents the absolute value of the difference between the frame interval time and the frame interval time standard value, and the specific limit expression is:

[0038] ;

[0039] In the formula, D represents the accuracy third component, P3 represents the frame interval time correction factor obtained from the identity verification database, R0 represents the frame interval time standard value obtained from the identity verification database, and R represents the frame interval time. Among them, the frame interval time and the dynamic feature extraction accuracy evaluation exist negative correlation, the longer the frame interval time, the lower the sampling rate, the weaker the ability to capture dynamic changes of expression, resulting in the loss of intermediate gradual frames, which cannot accurately describe the expression change trajectory, and the lower the dynamic feature extraction accuracy evaluation.

[0040] Finally, the accuracy first component, the accuracy second component and the accuracy third component are coupled to obtain the dynamic feature extraction accuracy evaluation, and the specific limit expression is:

[0041] ;

[0042] In the formula, U represents the dynamic feature extraction accuracy evaluation.

[0043] It needs to be explained that the identity verification database stores influence factors related to feature extraction parameters, including key frame missing rate correction factors, single feature point displacement vector field correction factors and frame interval time correction factors. The values of these influence factors are mostly in the range of 0 to 1, and the sum of the three values is 1. At the same time, there is a pre-set mapping relationship between these influence factors and feature extraction parameters. This relationship may be a one-to-one direct correspondence, or a many-to-one aggregate correspondence.

[0044] Specifically, there is a correlation between feature extraction parameters, as follows: the larger the frame interval time, the smaller the speed under the same displacement distance, so the motion speed will decrease, and the smaller the single feature point displacement vector field; the higher the key frame missing rate, the more likely the feature point matching fails, and there is no reference frame, so the single feature point displacement vector field is smaller.

[0045] It needs to be further explained that by understanding the correlation between the feature extraction parameters, it is helpful to grasp the influence that different parameter changes may have on the feature point displacement vector field in the actual extraction process: when the interframe interval time is detected to increase, the sensitivity of the feature tracking algorithm can be adjusted accordingly to avoid misjudgment of the displacement vector field caused by speed calculation deviation, and it can be adjusted adaptively according to its correlation with other parameters, thereby ensuring the consistency and reliability of the feature extraction results, and providing higher quality basic data for subsequent facial expression feature comparison and identity authentication and other links. By understanding the positive and negative correlation between the feature extraction parameters and the dynamic feature extraction accuracy evaluation, it is helpful to realize the fine regulation and optimization of the feature extraction process in actual operation, which can accurately identify the key parameters affecting the accuracy and their action direction, avoid resource waste or effect fluctuation caused by blind adjustment, and thus realize the continuous and stable improvement of the dynamic feature extraction accuracy, output more reliable feature data for subsequent facial expression feature comparison and identity authentication and other links, and fundamentally enhance the performance of the entire authentication system.

[0046] The dynamic feature extraction accuracy analysis module is configured to determine whether to perform dynamic feature extraction accuracy optimization according to the dynamic feature extraction accuracy evaluation, and if so, to perform the facial expression feature comparison link after optimization, otherwise to directly perform the facial expression feature comparison link. The dynamic feature extraction accuracy optimization includes dynamic adjustment of regional feature weight to improve the distinguishability in facial expression feature extraction, and dynamic adjustment of face region mask weight to improve the accuracy of feature representation.

[0047] The multi-dimensional comparison module is configured to obtain multi-dimensional comparison parameters in the facial expression feature comparison link to quantify the rationality in the facial expression feature comparison process, and obtain a multi-dimensional comparison rationality evaluation.

[0048] It should be noted that the multi-dimensional comparison parameters include: to-be-compared dynamic feature extraction accuracy evaluation, single feature point positioning error and expression switching interval; Specifically, if dynamic feature extraction accuracy optimization is performed, the optimized dynamic feature extraction accuracy evaluation is recorded as the to-be-compared dynamic feature extraction accuracy evaluation, otherwise the current dynamic feature extraction accuracy evaluation is directly obtained and recorded as the to-be-compared dynamic feature extraction accuracy evaluation; The image is processed using a feature point detection algorithm such as SIFT, the coordinates of the positioned feature points are extracted and output, and the absolute value of the difference between the positioned coordinates of the feature points and the preset coordinates is calculated by the Euclidean distance. The absolute value is recorded as the single feature point positioning error; The time interval required for the user to switch from one expression to another is recorded by a high-speed camera and is recorded as the expression switching interval. The multi-dimensional comparison parameter threshold and the multi-dimensional comparison parameter correction factor are obtained from the constructed identity verification database. The multi-dimensional comparison parameter threshold includes the dynamic feature extraction accuracy evaluation threshold, the single feature point positioning error standard value and the expression switching interval standard value, and the multi-dimensional comparison parameter correction factor includes the dynamic feature extraction accuracy evaluation correction factor, the single feature point positioning error correction factor and the expression switching interval correction factor.

[0049] It should be explained that first, the to-be-compared dynamic feature extraction accuracy evaluation and the dynamic feature extraction accuracy evaluation threshold are modified by the dynamic feature extraction accuracy evaluation correction factor, and the result of the proportion analysis is recorded as the first component of rationality. The specific limit expression is:

[0050] ;

[0051] In the formula, V represents the first component of rationality, C1 represents the dynamic feature extraction accuracy evaluation correction factor obtained from the identity verification database, F0 represents the dynamic feature extraction accuracy evaluation threshold obtained from the identity verification database, and F represents the to-be-compared dynamic feature extraction accuracy evaluation. Among them, the to-be-compared dynamic feature extraction accuracy evaluation and the multi-dimensional comparison rationality evaluation have a positive correlation. The greater the to-be-compared dynamic feature extraction accuracy evaluation, the closer the time sequence features such as AU intensity change and expression conversion trajectory to the real situation, and the higher the multi-dimensional comparison rationality evaluation.

[0052] Next, the single feature point positioning error standard value and the positioning error deviation degree are modified by the single feature point positioning error correction factor, wherein the positioning error deviation degree represents the absolute value of the difference between the single feature point positioning error and the single feature point positioning error standard value, and is recorded as the second component of rationality. The specific limit expression is:

[0053] ;

[0054] In the formula, S represents the second component of rationality, C2 represents the single feature point positioning error correction factor obtained from the identity verification database, W0 represents the single feature point positioning error standard value obtained from the identity verification database, and W represents the single feature point positioning error. Among them, the single feature point positioning error has a negative correlation with the multi-dimensional comparison rationality evaluation, and the higher the single feature point positioning error, the more likely it is to cause deviation in spatial dimension comparison, and the lower the multi-dimensional comparison rationality evaluation.

[0055] Then, the proportion analysis result of the expression switching interval standard value and the switching interval deviation degree is corrected by the expression switching interval correction factor, and is recorded as the third component of rationality, wherein the switching interval deviation degree represents the absolute value of the difference between the expression switching interval and the expression switching interval standard value, and the specific limit expression is:

[0056] ;

[0057] In the formula, N represents the third component of rationality, C3 represents the expression switching interval correction factor obtained from the identity verification database, H0 represents the expression switching interval standard value obtained from the identity verification database, and H represents the expression switching interval. Among them, the expression switching interval has a negative correlation with the multi-dimensional comparison rationality evaluation, and the longer the expression switching interval, the more dynamic information is lost in the long interval, resulting in feature loss and lower multi-dimensional comparison rationality evaluation.

[0058] It needs to be explained that the identity verification database stores influence factors related to multi-dimensional comparison parameters, including dynamic feature extraction accuracy evaluation correction factor, single feature point positioning error correction factor, and expression switching interval correction factor. The values of these influence factors are mostly in the range of 0 to 1, and the sum of the three values is 1. At the same time, there is a pre-set mapping relationship between these influence factors and multi-dimensional comparison parameters, which may be one-to-one direct correspondence or many-to-one aggregate correspondence.

[0059] Specifically, there is a correlation between multi-dimensional comparison parameters, as follows: the higher the dynamic feature extraction accuracy evaluation, the higher the accuracy in the process of facial expression feature extraction, and the smaller the single feature point positioning error; the shorter the expression switching interval, the higher the sampling density and real-time requirement of dynamic features, and the higher the dynamic feature extraction accuracy evaluation; the larger the single feature point positioning error, the longer the expression switching interval threshold value of the system is forced to extend to avoid misjudgment, and the longer the expression switching interval.

[0060] Further, by understanding the correlation between multi-dimensional comparison parameters, it helps to reduce the contradictions and deviations caused by isolated parameter adjustment, and also improves the adaptability of the system to different expression change scenarios, provides more reliable mechanism support for accurate comparison of facial expression features, and further enhances the stability and effectiveness of the entire identity authentication process; by understanding the positive and negative correlation between multi-dimensional comparison parameters and multi-dimensional comparison rationality evaluation, it helps to quickly locate the parameters that may cause fluctuations and make targeted adjustments, ensuring that multi-dimensional comparison always remains at a high level of rationality, providing a more solid guarantee for accurate matching of facial expression features, and further promoting the entire identity authentication process to a more efficient and reliable direction.

[0061] The multi-dimensional comparison rationality analysis module is configured to determine whether to perform multi-dimensional comparison rationality optimization according to the multi-dimensional comparison rationality evaluation. If yes, a risk identity authentication decision is generated after optimization, otherwise, a risk identity authentication decision is directly generated. The risk identity authentication decision includes identity verification pass and identity verification fail. The multi-dimensional comparison rationality optimization includes transition frame weight dynamic adjustment to improve the matching accuracy of dynamic feature sequences in the expression switching process, and static-dynamic feature weight dynamic adjustment to improve the consistency of multi-dimensional comparison with the real identity.

[0062] As an embodiment of the first aspect, through the multi-module collaborative work and dynamic optimization mechanism, the reliability of the risk identity authentication is significantly improved. The closed-loop design of dynamic optimization and multi-dimensional evaluation effectively solves the problems of insufficient feature extraction discrimination, large dynamic sequence matching error, and unreasonable static and dynamic feature fusion in traditional facial expression authentication, making the generated risk identity authentication decision (including identity verification pass and fail) more accurate and reliable, and comprehensively improving the performance of the entire risk identity authentication system.

[0063] Further, the specific determination steps of determining whether to perform dynamic feature extraction accuracy optimization are as follows: obtaining an accuracy evaluation threshold from the constructed identity authentication database; if the dynamic feature extraction accuracy evaluation is not less than the accuracy evaluation threshold, no dynamic feature extraction accuracy optimization is performed, and a feature extraction accuracy qualified instruction is sent to the multi-dimensional comparison module; if the dynamic feature extraction accuracy evaluation is less than the accuracy evaluation threshold, it is determined whether to perform regional feature weight dynamic adjustment according to the brightness deviation; if yes, it is determined whether to perform face region mask weight dynamic adjustment after dynamic adjustment, otherwise, it is directly determined whether to perform face region mask weight dynamic adjustment.

[0064] As a further specific description, according to the luminance deviation degree, whether to perform the dynamic adjustment of the regional feature weight is judged. The specific process is as follows: obtaining the preset shadow luminance upper limit, the preset shadow luminance lower limit, the preset highlight luminance lower limit and the preset highlight luminance upper limit from the constructed identity authentication database, the preset shadow luminance upper limit is less than the preset shadow luminance lower limit, the preset shadow luminance lower limit is less than the preset highlight luminance lower limit, and the preset highlight luminance lower limit is less than the preset highlight luminance upper limit. The image is converted from the RGB (Red, Green, Blue) color space to the HSV (Hue, Saturation, Value) color space, the V channel (Value) is extracted, the image is divided into regions according to the preset key regions (such as the corners of the mouth, the alae nasi and the lower jaw line), the V values of all pixels in each region are averaged to obtain the actual luminance value of the region, and the difference between the actual luminance value and the preset reference luminance value is recorded as the luminance deviation degree. It needs to be explained that the highlight of the non-key region such as the forehead and the cheek can be directly adjusted according to the threshold value, and the eye highlight (such as the pupil reflection) needs to be processed separately in combination with the pupil detection (to avoid misjudgment as a highlight region). The adjusted regional weight is converted into a “pixel-level weight matrix” (consistent with the input image size, and the weight of each pixel is equal to the adjusted weight of the region to which it belongs); at the feature extraction layer of the CNN model (such as after the first convolutional layer), the matrix is multiplied point by point with the feature map to realize the dynamic suppression of the feature of the abnormal light region.

[0065] If the luminance deviation degree is within the preset deviation degree standard interval, it means that the luminance deviation within the interval belongs to the normal fluctuation category and is not caused by the feature distortion of the abnormal light (such as extreme shadow and excessive highlight). In order to reduce invalid calculation and resource consumption and improve processing efficiency, the dynamic adjustment of the regional feature weight is not performed. The preset deviation degree standard interval represents the closed interval formed by the preset shadow luminance lower limit and the preset highlight luminance lower limit.

[0066] If the luminance deviation is within the preset shadow luminance interval, it is recorded as a first-level shadow region. The weighted average of the accuracy evaluation correction amount and the luminance deviation is taken to the negative infinity integer, and the result is taken as the shadow region weight reduction amount. This can effectively balance the region's own characteristics and system evaluation errors. By taking the negative infinity integer, the weight reduction intensity can be strengthened, ensuring that the weight of the first-level shadow region can accurately match its actual impact in the image - neither will it interfere with normal feature extraction due to too high a weight, nor will it ignore its due role due to too low a weight. Thus, a more reliable feature basis is provided for subsequent image analysis (such as target detection, scene understanding, etc.), and the result of multiplying the preset shadow region weight and the shadow region weight reduction amount is taken as the current shadow region weight to strengthen the feature differentiation between the first-level shadow region and other regions, optimize the overall image feature distribution, and the accuracy evaluation correction amount represents the difference between the accuracy evaluation threshold and the dynamic feature extraction accuracy evaluation. The preset shadow luminance interval represents the open interval formed by the preset shadow luminance lower limit and the preset shadow luminance upper limit.

[0067] If the luminance deviation is greater than or equal to the preset shadow luminance upper limit, it is recorded as a second-level shadow region. The weighted average of the accuracy evaluation correction amount and the deviation compensation amount is taken to the positive infinity integer, and the result is taken as the shadow region weight increase amount. The result of multiplying the preset shadow region weight and the shadow region weight increase amount is taken as the current shadow region weight. By taking the integer, the increase intensity is strengthened, and the weight of the second-level shadow region is accurately increased to match its significant feature influence. The dynamic correlation between the system state and the region feature enhances the adaptive adjustment capability. The explicit calculation logic improves the explainability and controllability, and at the same time, the feature differentiation with other regions is strengthened, and the overall feature distribution is optimized, thereby providing a more reliable feature basis for subsequent image analysis, improving the processing accuracy, and the deviation compensation amount represents the difference between the luminance deviation and the preset shadow luminance upper limit.

[0068] In this embodiment, through the coordinated operation of multiple links, accurate processing of image light abnormal regions is achieved, bringing multiple-dimensional significant technical effects. The way of directly adjusting the non-key region highlight according to the threshold and separately processing the eye highlight combined with pupil detection reflects the flexibility and pertinence of the processing - it simplifies the processing process of the non-key region to improve efficiency, and avoids misjudgment of the eye highlight (such as pupil reflection) through special processing, ensuring the processing accuracy of the key feature region. Overall, through standardized parameter definition, refined region processing, differentiated weight adjustment, and deep fusion with model feature extraction, the accuracy and pertinence of light abnormal region processing are ensured, and the running efficiency, adaptive ability, and explainability of the system are improved, providing a more reliable feature basis for subsequent tasks such as image target detection and scene understanding, effectively reducing misjudgment and omission caused by light problems, and significantly enhancing the robustness of the image processing system in complex lighting environments.

[0069] As a further specific description, the determining whether to perform the region feature weight dynamic adjustment further comprises:

[0070] If the luminance deviation is within the preset high-light luminance interval, it is recorded as a first-level high-light region, the result of taking the arithmetic mean of the accuracy evaluation correction amount and the luminance deviation and then rounding down is taken as the high-light region weight reduction amount, the result of multiplying the preset high-light region weight and the high-light region weight reduction amount is taken as the current high-light region weight, the down-regulation intensity is strengthened by rounding down to accurately match the actual influence of the first-level high-light region, to avoid excessive interference with normal feature extraction while retaining necessary weights, to optimize the overall feature distribution, to improve the accuracy and pertinence of the system in processing high-light regions, to provide a reliable feature basis for subsequent image analysis, and the preset high-light luminance interval represents an open interval formed by the preset high-light luminance lower limit and the preset high-light luminance upper limit.

[0071] If the luminance deviation is greater than or equal to the preset high-light luminance upper limit, it is recorded as a second-level high-light region, the result of taking the arithmetic mean of the accuracy evaluation correction amount and the deviation adjustment amount and then rounding up is taken as the high-light region weight increase amount, the result of multiplying the preset high-light region weight and the high-light region weight increase amount is taken as the current high-light region weight, the dynamic adjustment is realized through the increase amount, which not only retains the empirical basis setting embodied by the preset weight, but also can be flexibly corrected according to the actual lighting conditions and the system evaluation state, can reflect the importance of this type of region under normal circumstances, and can be real-time adapted to the high-light intensity of specific scenes and the system processing precision, to ensure that the weight value can accurately reflect the actual feature importance of the second-level high-light region in the current image, through accurate region definition, multi-parameter fusion adjustment logic and intensive rounding operation, the weight can be dynamically optimized according to the actual situation of the second-level high-light region, to effectively offset the interference of strong light on feature extraction, to enable the subsequent image analysis (such as target detection, feature matching, etc.) to more clearly capture the effective information of the region, to optimize the feature distribution of the overall image, to provide a more reliable feature basis for various image processing tasks, to reduce the feature loss or misjudgment caused by strong light, to improve the robustness of the system in complex lighting environments, and the deviation adjustment amount represents the difference between the luminance deviation and the preset high-light luminance.

[0072] In this embodiment, by finely dividing the first and second high light regions and implementing a differentiated weight adjustment strategy, precise suppression and feature optimization of high light interference are achieved. Through precise region definition, multi-parameter fusion adjustment logic and directional rounding operation (first downward reinforcement suppression, second upward dynamic adaptation), the interference of high light of different intensities on feature extraction is effectively offset, avoiding excessive influence of the first high light region and ensuring effective information retention of the second high light region under strong light, thereby optimizing the overall image feature distribution, providing a reliable feature basis for subsequent target detection, feature matching and other analysis tasks, reducing feature loss or misjudgment caused by light, and significantly improving the processing accuracy of the system in complex lighting environments.

[0073] As shown in Figure 2 The specific logic is: determining whether to perform dynamic adjustment of the face region mask weight according to the mask entropy, if the mask entropy is within a preset mask entropy standard interval, then not performing dynamic adjustment of the face region mask weight, if the mask entropy is less than a preset lower mask entropy limit, then taking the result of the harmonic average of the accuracy evaluation correction amount and the mask entropy compensation amount and then rounding to zero as the face region mask weight up adjustment amount, and if the mask entropy is greater than a preset upper mask entropy limit, then taking the result of the harmonic average of the accuracy evaluation correction amount and the mask entropy correction amount and then rounding to zero as the face region mask weight down adjustment amount.

[0074] As a further specific description, the specific steps of determining whether to perform dynamic adjustment of the face region mask weight are: the face region mask weight is essentially a weight matrix or mask graph, which has the same spatial size (such as height x width) as the input face image, and the value of each pixel or region represents the importance degree of the region in the model decision process. For example: a weight of 1 means that the region fully participates in the calculation; a weight of 0 means that the region is ignored; an intermediate value (such as 0.5) means partial participation or reduced influence; the mask entropy is a quantitative index for measuring the dispersion degree of the face region mask weight distribution information.

[0075] If the mask entropy is within a preset mask entropy standard interval, it means that the uncertainty of the current face region mask is moderate, the information amount distribution is reasonable, and dynamic weight adjustment is not needed, then not performing dynamic adjustment of the face region mask weight, the preset mask entropy standard interval represents a closed interval formed by the preset lower mask entropy limit and the preset upper mask entropy limit.

[0076] If the mask entropy is less than the preset mask entropy lower bound, it means that the uncertainty of the current face region mask is too low, the information is too concentrated, and there is a risk of insufficient feature expression ability or excessive sensitivity to local details. Then, the result of harmonically averaging the accuracy evaluation correction amount and the mask entropy compensation amount and rounding to zero is taken as the face region mask weight up-regulation amount, the result of multiplying the face region mask weight up-regulation amount and the preset face region mask weight is taken as the current face region mask weight, and rounding to zero strengthens the smoothness and conservatism of weight up-regulation, avoiding feature distribution imbalance caused by excessive adjustment. After the weight up-regulation, the importance of the face region in feature extraction is effectively enhanced, the feature sparsity or local overfitting problem caused by too low mask entropy is compensated, the feature expression ability of the face details is improved, the fine compensation optimization of feature weight is realized by dynamically responding to the abnormally low state of mask entropy, ensuring that the feature expression is still rich in low uncertainty scenarios, providing a more reliable feature basis for subsequent face recognition, expression analysis and other tasks, and the mask entropy compensation amount represents the difference between the preset mask entropy lower bound and the mask entropy.

[0077] If the mask entropy is greater than the preset mask entropy upper bound, it means that the uncertainty of the current face region mask is too high, the information distribution is too dispersed, and noise or interference features may be introduced, affecting the effective expression of the face region. Then, the result of harmonically averaging the accuracy evaluation correction amount and the mask entropy correction amount and rounding to zero is taken as the face region mask weight down-regulation amount, the result of multiplying the face region mask weight down-regulation amount and the preset face region mask weight is taken as the current face region mask weight, the harmonically averaged fusion integrates the system evaluation accuracy and the correction demand of mask entropy over-limit, and rounding to zero ensures the smoothness and conservatism of weight down-regulation, avoiding the loss of key features caused by excessive inhibition. After the weight down-regulation, the interference of high uncertainty mask on feature extraction is effectively reduced, and the excessive influence of noise or redundant information is suppressed, thereby improving the purity and stability of the face region features. Through dynamic response to the abnormally high state of mask entropy, fine inhibition optimization of feature weight is realized, ensuring that key features can still be reliably extracted in high uncertainty scenarios, providing a more robust feature basis for subsequent face recognition, expression analysis and other tasks, while optimizing the adaptability and anti-interference ability of the system under complex or low-quality mask conditions. The mask entropy correction amount represents the difference between the mask entropy and the preset mask entropy upper bound.

[0078] In this embodiment, a dynamic weight adjustment mechanism based on mask entropy is used to achieve intelligent response and refined optimization to the uncertainty of facial region masks, thereby significantly improving the robustness and adaptability of feature extraction. The overall mechanism dynamically responds to abnormal states of mask entropy, achieving refined compensation or suppression optimization of feature weights. This avoids feature distortion caused by over-adjustment and ensures rich feature representation and anti-interference capabilities in both low and high uncertainty scenarios. This provides a more stable and reliable feature foundation for subsequent tasks such as face recognition and expression analysis, significantly improving the system's processing accuracy under complex or low-quality mask conditions.

[0079] Furthermore, the specific steps for determining whether to perform multi-dimensional comparison rationality optimization based on the multi-dimensional comparison rationality assessment are as follows: If the multi-dimensional comparison rationality assessment is greater than or equal to the preset upper bound of the rationality assessment, then no multi-dimensional comparison rationality optimization is performed, and a comparison qualified instruction is sent; if the multi-dimensional comparison rationality assessment is within the preset rationality range, then it is determined whether to dynamically adjust the transition frame weight based on the expression switching interval. If so, then it is determined whether to dynamically adjust the static-dynamic feature weight after the dynamic adjustment; otherwise, it is directly determined whether to dynamically adjust the static-dynamic feature weight. The preset rationality range represents the open interval formed by the preset lower bound of the rationality assessment and the preset upper bound of the rationality assessment; if the multi-dimensional comparison rationality assessment is less than or equal to the preset lower bound of the rationality assessment, then a comparison unqualified instruction is sent to the preset reviewer.

[0080] like Figure 3 The diagram shows a flowchart of the dynamic adjustment of transition frame weights in a risk control identity authentication system based on facial expression recognition, provided in an embodiment of this application. The specific logic is as follows: Whether to dynamically adjust the transition frame weights is determined based on the expression switching interval. If the expression switching interval is greater than the preset upper limit of the expression switching interval, the multi-dimensional comparison rationality assessment and the switching interval compensation amount are input into the identity authentication database to map and obtain the transition frame weight increase amount. If the expression switching interval is within the preset range, no dynamic adjustment of the transition frame weights is performed. If the expression switching interval is less than the preset lower limit of the expression switching interval, the multi-dimensional comparison rationality assessment and the switching interval correction amount are input into the identity authentication database to map and obtain the transition frame weight decrease amount.

[0081] As a further explanation, the specific process for determining whether to dynamically adjust the weight of transition frames based on the expression switching interval is as follows: Transition frames refer to all intermediate frames between the starting frame (initial expression) and the target frame (final expression) during the process of an expression switching from one state to another.

[0082] If the expression switching interval is greater than the preset expression switching interval upper limit, the multi-dimensional comparison rationality evaluation and switching interval compensation quantity are input to the identity authentication database mapping to obtain a transition frame weight up-regulation quantity, and a result of multiplying the transition frame weight up-regulation quantity and the preset transition frame weight is taken as a current transition frame weight. The up-regulated transition frame weight effectively enhances the importance of the transition frame in a long-time expression state, makes up for the feature sparseness or identity information weakening problem caused by the excessively long switching interval, and thus improves the contribution of the transition frame in identity authentication or expression analysis. Through dynamic response to the abnormally prolonged state of the expression switching interval, fine compensation optimization of the transition frame weight is realized, ensuring that the identity feature continuity and reliability can still be maintained in the slow or stagnant expression switching scene, providing more stable feature support for subsequent identity authentication, expression recognition and the like, while enhancing the adaptability and accuracy of the system under complex or non-ideal expression switching conditions. The switching interval compensation quantity represents the difference between the expression switching interval and the preset expression switching interval upper limit.

[0083] If the expression switching interval is within the preset expression switching interval range, it means that the current expression state switching frequency is determined to be reasonable, and the information distribution of the transition frame is continuous and sufficient, so that dynamic optimization of the transition frame weight is not needed. The preset expression switching interval represents a closed interval formed by the preset expression switching interval lower limit and the preset expression switching interval upper limit.

[0084] If the expression switching interval is less than the preset expression switching interval lower limit, it means that the current expression state switching is determined to be too frequent, and the transition frame may have information redundancy or noise interference, thereby affecting the accuracy and stability of the identity feature expression. Then, the multi-dimensional comparison rationality evaluation and switching interval correction quantity are input to the identity authentication database mapping to obtain a transition frame weight down-regulation quantity, and a result of multiplying the transition frame weight down-regulation quantity and the preset transition frame weight is taken as a current transition frame weight. By using the identity verification database mapping relationship, the abnormality degree of the expression switching frequency and the multi-dimensional identity feature comparison result are comprehensively considered, ensuring that the weight down-regulation quantity not only reflects the deviation degree of the switching interval, but also takes into account the rationality of identity authentication, thereby realizing accurate suppression. By down-regulating the transition frame weight, the system effectively reduces the interference of the transition frame in the frequent switching state, suppresses the influence of redundant information and noise on identity feature extraction, thereby improving the dominant role of the key frame in identity authentication or expression analysis. Through dynamic response to the abnormally shortened state of the expression switching interval, fine suppression optimization of the transition frame weight is realized, ensuring that the identity feature purity and reliability can still be maintained in the scene where the expression switching is too frequent, providing more stable feature support for subsequent identity authentication, expression recognition and the like, while enhancing the adaptability and anti-interference ability of the system under fast switching or non-ideal expression change conditions. The switching interval correction quantity represents the difference between the preset expression switching interval lower limit and the expression switching interval.

[0085] In the embodiment, through the dynamic weight adjustment mechanism based on the expression switching interval, intelligent response and fine optimization of the importance of transition frames are realized, and the accuracy of identity authentication and expression analysis under complex expression change conditions is significantly improved. The overall mechanism dynamically responds to different abnormal states of the expression switching interval, realizes fine compensation or inhibition optimization of the transition frame weight, avoids feature distortion caused by excessive adjustment, and ensures that stable and reliable features are provided in various expression switching scenarios, laying an efficient technical foundation for subsequent identity authentication, expression recognition and other tasks.

[0086] As a further specific description, the specific steps of judging whether to perform static-dynamic feature weight dynamic adjustment are as follows: the static feature weight is used to quantify the contribution of static facial features in multi-dimensional comparison rationality evaluation. The static facial features refer to the inherent features of the relatively stable face, such as the geometric proportion of the face contour (the proportion of eye distance, nose width and face width), skin texture (distribution pattern of freckles and wrinkles), coordinate position of key feature points (such as fixed relative position of eye corner, nose tip and mouth corner) and the like, which are not easy to change with expression and are the core basis for identity recognition. The dynamic feature weight is used to measure the importance of dynamic facial features in multi-dimensional comparison rationality evaluation. The dynamic facial features refer to the motion features generated with expression changes, such as the speed of expression switching, the trajectory of muscle movement in transition frames (the amplitude change of mouth corner lifting, the rate of brow contraction and the like), the time interval of different expression state switching and the like, which reflect the naturalness and continuity of expression changes and can be used to distinguish real expressions from fake expressions (such as photo, video synthesized expressions).

[0087] If the single feature point positioning error is less than the preset positioning error asymptotic lower bound, it means that the current feature point positioning accuracy is extremely high, the static feature information is extremely reliable, but there may be a risk of relatively insufficient or redundant dynamic feature expression, then the multi-dimensional comparison rationality evaluation and positioning error comparison amount are input to the identity authentication database mapping to obtain a static feature weight adjustment instruction, the static feature weight adjustment instruction represents the result obtained by arithmetically averaging the rationality evaluation and positioning error comparison amount, the static feature weight adjustment instruction is input to the identity authentication database mapping to obtain a dynamic feature weight adjustment instruction, which comprehensively considers the accuracy of feature point positioning and the multi-dimensional identity feature comparison result, ensuring that the static feature weight adjustment amount reflects both the excellent performance of the positioning error and the rationality of identity authentication, thereby achieving precise enhancement. By adjusting the static feature weight and simultaneously adjusting the dynamic feature weight, the system effectively enhances the dominant role of high-precision static features in identity authentication and suppresses the possible redundancy or noise interference in dynamic features, thereby improving the purity and stability of the overall feature expression. By dynamically responding to the abnormally excellent state of the feature point positioning error, fine balance optimization of the static and dynamic feature weights is achieved, ensuring that the reliable expression of identity features can still be maintained in the case of extremely accurate feature point positioning, providing more robust feature support for subsequent identity authentication, feature matching and other tasks, while enhancing the adaptability and anti-interference ability of the system under the condition of high-quality feature point positioning. The positioning error comparison amount represents the difference between the preset positioning error asymptotic lower bound and the single feature point positioning error.

[0088] If the single feature point positioning error is within the preset positioning error interval, it is determined that the current feature point positioning accuracy is within a reasonable range, and the information distribution between the static feature and the dynamic feature is balanced and complementary, and there is no need for static-dynamic feature weight dynamic adjustment. The preset positioning error interval represents a closed interval formed by the preset positioning error asymptotic lower bound and the preset positioning error asymptotic upper bound.

[0089] If the single feature point positioning error is greater than the preset positioning error asymptotic upper bound, it means that the current feature point positioning accuracy is insufficient, the static feature information has unreliable risk, and it may affect the accuracy of identity authentication or feature matching. Therefore, the multi-dimensional comparison rationality evaluation and the positioning error correction amount are input to the identity authentication database mapping to obtain a static feature weight down-regulation instruction. The static feature weight down-regulation instruction represents the result obtained by arithmetically averaging the rationality evaluation and the positioning error correction amount. The static feature weight down-regulation instruction is input to the identity authentication database mapping to obtain a dynamic feature up-regulation instruction. By down-regulating the static feature weight and synchronously up-regulating the dynamic feature weight, the misjudgment risk introduced by the excessive static feature positioning error is effectively reduced, and the compensation effect of the dynamic feature in identity authentication is strengthened, thereby improving the adaptability and robustness of the overall feature expression. Through the dynamic response to the abnormally high state of the feature point positioning error, fine balance optimization of the static and dynamic feature weights is realized, ensuring that the reliable expression and stability of the identity feature can be maintained in the scene with poor feature point positioning quality, providing more robust feature support for subsequent identity authentication, feature matching, and other tasks, while enhancing the adaptability and anti-interference ability of the system under the condition of low-quality feature point positioning. The positioning error correction amount represents the difference between the single feature point positioning error and the preset positioning error asymptotic upper bound.

[0090] In the embodiment, through the dynamic weight adjustment mechanism based on the single feature point positioning error, intelligent response and fine balance optimization of the static and dynamic feature weights are realized, thereby significantly improving the accuracy of identity authentication and feature matching. Through the database mapping relationship, the positioning error and the multi-dimensional comparison rationality are comprehensively considered to realize fine regulation and control of the feature weight under different positioning accuracy scenarios, which not only avoids feature distortion caused by excessive adjustment, but also ensures that the reliable expression and robustness of the identity feature can be maintained under the conditions of extremely high, moderate, or poor feature point positioning, thereby providing more robust feature support for subsequent identity authentication, feature matching, and other tasks, while significantly enhancing the adaptability and anti-interference ability of the system under the condition of complex or low-quality feature point positioning.

[0091] As Figure 4As shown, a flow chart of a risk control identity authentication method based on facial expression recognition is provided in the embodiment of the application, the risk control identity authentication method based on facial expression recognition provided in the embodiment of the application comprises: collecting a facial image sequence of a user to be authenticated; pre-processing the facial image sequence to obtain a feature extraction parameter to quantify the accuracy of a CNN and an LSTM model in a facial expression feature extraction process, and to obtain a dynamic feature extraction accuracy evaluation; judging whether to perform dynamic feature extraction accuracy optimization according to the dynamic feature extraction accuracy evaluation, if yes, then performing a facial expression feature comparison link after optimization, otherwise directly performing the facial expression feature comparison link, the dynamic feature extraction accuracy optimization comprising regional feature weight dynamic adjustment and facial region mask weight dynamic adjustment; obtaining a multi-dimension comparison parameter in the facial expression feature comparison link to quantify the rationality in the facial expression feature comparison process, and to obtain a multi-dimension comparison rationality evaluation; judging whether to perform multi-dimension comparison rationality optimization according to the multi-dimension comparison rationality evaluation, if yes, then generating a risk control identity authentication decision after optimization, otherwise directly generating the risk control identity authentication decision, the multi-dimension comparison rationality optimization comprising transition frame weight dynamic adjustment and static-dynamic feature weight dynamic adjustment.

[0092] As an embodiment of the second aspect, a user logs in through a mobile banking App, adopts a dynamic identity authentication mechanism based on facial expression sequence, is required to complete a sequence of expression actions of smiling -> natural -> blinking in the login process, and a facial image sequence thereof is collected, dynamic feature extraction and comparison are performed by combining a CNN and an LSTM model, and finally a risk control identity authentication decision is generated. Subsequently, the image sequence is pre-processed through face detection, alignment, normalization and illumination compensation, a standardized facial region input is generated, and the CNN-LSTM hybrid model is input for feature extraction, wherein the CNN extracts spatial features (such as feature contours and textures), and the LSTM models time sequence dynamic features (such as expression change trajectories), and meanwhile, a dynamic feature extraction accuracy evaluation is generated by calculating the cross-entropy loss of the model output and the labeled expression label and the cosine similarity of the feature vector; the extracted dynamic feature sequence is compared with a user registration template, the system performs multi-dimension comparison rationality optimization, and outputs “identity verification passed” to allow login based on the final comparison result, if an abnormality is detected, triggers secondary verification or restricts high-risk operations, so that through the dynamic weight adjustment and multi-dimension optimization mechanism, the accuracy, robustness and security of identity authentication can still be maintained in complex lighting, expression blur or switching abnormality and other scenes.

[0093] The above is only a specific embodiment of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the application, which should be covered within the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A face expression recognition based anti-fraud identity authentication system, characterized in that, The system comprises an identity information collection module, a facial feature extraction module, a dynamic feature extraction accuracy analysis module, a multi-dimensional comparison module and a multi-dimensional comparison rationality analysis module. The identity information collection module is configured to collect a face image sequence of a user to be authenticated. The facial feature extraction module is configured to pre-process the face image sequence to obtain feature extraction parameters for quantifying the accuracy of the CNN and LSTM models in the facial expression feature extraction process, and to obtain a dynamic feature extraction accuracy evaluation. The dynamic feature extraction accuracy analysis module is configured to determine whether to perform dynamic feature extraction accuracy optimization based on the dynamic feature extraction accuracy evaluation. If so, the facial expression feature comparison step is performed after optimization. Otherwise, the facial expression feature comparison step is directly performed. The dynamic feature extraction accuracy optimization includes regional feature weight dynamic adjustment and facial region mask weight dynamic adjustment. The multi-dimensional comparison module is configured to obtain multi-dimensional comparison parameters in the facial expression feature comparison step to quantify the rationality of the facial expression feature comparison process and obtain a multi-dimensional comparison rationality evaluation. The multi-dimensional comparison rationality analysis module is configured to determine whether to perform multi-dimensional comparison rationality optimization based on the multi-dimensional comparison rationality evaluation. If so, the risk control identity authentication decision is generated after optimization. Otherwise, the risk control identity authentication decision is directly generated. The multi-dimensional comparison rationality optimization includes transition frame weight dynamic adjustment and static-dynamic feature weight dynamic adjustment.

2. The facial expression recognition based anti-fraud identity authentication system according to claim 1, characterized in that, The multi-dimensional comparison parameters include: dynamic feature extraction accuracy evaluation, single feature point positioning error and expression switching interval. The multi-dimensional comparison parameter threshold and the multi-dimensional comparison parameter correction factor are obtained from the constructed identity verification database. The multi-dimensional comparison parameter threshold includes dynamic feature extraction accuracy evaluation threshold, single feature point positioning error standard value and expression switching interval standard value. The multi-dimensional comparison parameter correction factor includes dynamic feature extraction accuracy evaluation correction factor, single feature point positioning error correction factor and expression switching interval correction factor. The specific determination steps of determining whether to perform multi-dimensional comparison rationality optimization based on the multi-dimensional comparison rationality evaluation are as follows: If the multi-dimensional comparison rationality evaluation is greater than or equal to the preset rationality evaluation upper bound, no multi-dimensional comparison rationality optimization is performed, and a comparison qualified instruction is sent. If the multi-dimensional comparison rationality evaluation is within the preset rationality interval, it is determined whether to perform transition frame weight dynamic adjustment based on the expression switching interval. If so, it is determined whether to perform static-dynamic feature weight dynamic adjustment after dynamic adjustment. Otherwise, it is directly determined whether to perform static-dynamic feature weight dynamic adjustment. The preset rationality interval represents an open interval formed by the preset rationality evaluation lower bound and the preset rationality evaluation upper bound. If the multi-dimensional comparison rationality evaluation is less than or equal to the preset rationality evaluation lower bound, a comparison unqualified instruction is sent to the preset audit personnel. The feature extraction parameters include: key frame missing rate, single feature point displacement vector field and frame interval time. Obtain the feature extraction parameter threshold and the feature extraction parameter correction factor from the constructed identity authentication database, wherein the feature extraction parameter threshold comprises a key frame missing rate threshold, a single feature point displacement vector field threshold and a frame interval time standard value, and the feature extraction parameter correction factor comprises a key frame missing rate correction factor, a single feature point displacement vector field correction factor and a frame interval time correction factor.

3. The facial expression recognition based anti-fraud identity authentication system according to claim 1, wherein, The specific judging step of judging whether to perform dynamic feature extraction accuracy optimization is: Obtain the accuracy evaluation threshold from the constructed identity authentication database; If the dynamic feature extraction accuracy evaluation is not less than the accuracy evaluation threshold, do not perform dynamic feature extraction accuracy optimization, and send a feature extraction accuracy qualified instruction to the multi-dimension comparison module; If the dynamic feature extraction accuracy evaluation is less than the accuracy evaluation threshold, judge whether to perform regional feature weight dynamic adjustment according to the brightness deviation, if yes, judge whether to perform face region mask weight dynamic adjustment after dynamic adjustment, otherwise, directly judge whether to perform face region mask weight dynamic adjustment.

4. The facial expression recognition based anti-fraud identity authentication system according to claim 3, characterized in that, The specific process of judging whether to perform regional feature weight dynamic adjustment according to the brightness deviation is: Obtain the preset shadow brightness upper limit, the preset shadow brightness lower limit, the preset highlight brightness lower limit and the preset highlight brightness upper limit from the constructed identity authentication database, wherein the preset shadow brightness upper limit is less than the preset shadow brightness lower limit, the preset shadow brightness lower limit is less than the preset highlight brightness lower limit, and the preset highlight brightness lower limit is less than the preset highlight brightness upper limit; If the brightness deviation is within the preset deviation standard interval, do not perform regional feature weight dynamic adjustment, wherein the preset deviation standard interval represents a closed interval formed by the preset shadow brightness lower limit and the preset highlight brightness lower limit; If the brightness deviation is within the preset shadow brightness interval, record it as a first-level shadow region, take the weighted average of the accuracy evaluation correction amount and the brightness deviation, and then take the integer part of the result towards negative infinity as the shadow region weight down adjustment amount, multiply the preset shadow region weight and the shadow region weight down adjustment amount to obtain the current shadow region weight, wherein the accuracy evaluation correction amount is used to represent the deviation amount of the accuracy evaluation threshold and the dynamic feature extraction accuracy evaluation, and the preset shadow brightness interval represents an open interval formed by the preset shadow brightness lower limit and the preset shadow brightness upper limit; If the brightness deviation is greater than or equal to the preset shadow brightness upper limit, record it as a second-level shadow region, take the weighted average of the accuracy evaluation correction amount and the deviation compensation amount, and then take the integer part of the result towards positive infinity as the shadow region weight up adjustment amount, multiply the preset shadow region weight and the shadow region weight up adjustment amount to obtain the current shadow region weight, wherein the deviation compensation amount is used to represent the deviation amount of the brightness deviation and the preset shadow brightness upper limit.

5. The facial expression recognition based anti-fraud identity authentication system according to claim 4, characterized in that, The judgment of whether to perform regional feature weight dynamic adjustment further comprises: If the luminance deviation degree is in the preset high-light luminance interval, it is recorded as a first high-light region, the result of arithmetically averaging the accuracy evaluation correction amount and the luminance deviation degree and then rounding down is taken as a high-light region weight reduction amount, the result of multiplying the preset high-light region weight and the high-light region weight reduction amount is taken as the current high-light region weight, and the preset high-light luminance interval represents an open interval formed by the preset high-light luminance lower limit and the preset high-light luminance upper limit; If the luminance deviation degree is greater than or equal to the preset high-light luminance upper limit, it is recorded as a second high-light region, the result of arithmetically averaging the accuracy evaluation correction amount and the deviation degree adjustment amount and then rounding up is taken as a high-light region weight increase amount, the result of multiplying the preset high-light region weight and the high-light region weight increase amount is taken as the current high-light region weight, and the deviation degree adjustment amount is used to represent the deviation amount of the luminance deviation degree from the preset high-light luminance.

6. The facial expression recognition based anti-fraud identity authentication system according to claim 3, characterized in that, The specific steps of judging whether to perform dynamic adjustment of the face region mask weight are: If the mask entropy is in the preset mask entropy standard interval, no dynamic adjustment of the face region mask weight is performed, and the preset mask entropy standard interval represents a closed interval formed by the preset mask entropy lower limit and the preset mask entropy upper limit; If the mask entropy is less than the preset mask entropy lower limit, the result of harmonic averaging the accuracy evaluation correction amount and the mask entropy compensation amount and then rounding to zero is taken as a face region mask weight increase amount, the result of multiplying the face region mask weight increase amount and the preset face region mask weight is taken as the current face region mask weight, and the mask entropy compensation amount is used to represent the deviation amount of the preset mask entropy lower limit from the mask entropy; If the mask entropy is greater than the preset mask entropy upper limit, the result of harmonic averaging the accuracy evaluation correction amount and the mask entropy correction amount and then rounding to zero is taken as a face region mask weight decrease amount, the result of multiplying the face region mask weight decrease amount and the preset face region mask weight is taken as the current face region mask weight, and the mask entropy correction amount is used to represent the deviation amount of the mask entropy from the preset mask entropy upper limit.

7. The facial expression recognition based anti-fraud identity authentication system according to claim 1, wherein, The specific process of judging whether to perform dynamic adjustment of the transition frame weight according to the expression switching interval is: If the expression switching interval is greater than the preset expression switching interval upper limit, the result of inputting the multi-dimensional comparison rationality evaluation and the switching interval compensation amount into the identity authentication database mapping is taken as a transition frame weight increase amount, and the result of multiplying the transition frame weight increase amount and the preset transition frame weight is taken as the current transition frame weight, and the switching interval compensation amount is used to represent the deviation amount of the expression switching interval from the preset expression switching interval upper limit; If the expression switching interval is in the preset expression switching interval, no transition frame weight optimization is performed, and the preset expression switching interval represents a closed interval formed by the preset expression switching interval lower limit and the preset expression switching interval upper limit. If the expression switching interval is less than the preset expression switching interval lower limit, the multi-dimension comparison rationality evaluation and the switching interval correction amount are input into the identity authentication database mapping to obtain a transition frame weight down-regulation amount, and a result of multiplying the transition frame weight down-regulation amount and the preset transition frame weight is taken as the current transition frame weight, and the switching interval correction amount is used to represent a deviation amount of the preset expression switching interval lower limit and the expression switching interval.

8. The facial expression recognition based anti-fraud identity authentication system according to claim 1, wherein, The specific steps of judging whether to perform the static-dynamic feature weight dynamic adjustment are: If the single feature point positioning error is less than the preset positioning error asymptotic lower limit, the multi-dimension comparison rationality evaluation and the positioning error comparison amount are input into the identity authentication database mapping to obtain a static feature weight up-regulation instruction, the static feature weight up-regulation instruction represents a result of performing an arithmetic average on the rationality evaluation and the positioning error comparison amount, the static feature weight up-regulation instruction is input into the identity authentication database mapping to obtain a dynamic feature down-regulation instruction, and the positioning error comparison amount is used to represent a deviation amount of the preset positioning error asymptotic lower limit and the single feature point positioning error. If the single feature point positioning error is within the preset positioning error interval, the static-dynamic feature weight dynamic adjustment is not performed, and the preset positioning error interval represents a closed interval formed by the preset positioning error asymptotic lower limit and the preset positioning error asymptotic upper limit. If the single feature point positioning error is greater than the preset positioning error asymptotic upper limit, the multi-dimension comparison rationality evaluation and the positioning error correction amount are input into the identity authentication database mapping to obtain a static feature weight down-regulation instruction, the static feature weight down-regulation instruction represents a result of performing an arithmetic average on the rationality evaluation and the positioning error correction amount, the static feature weight down-regulation instruction is input into the identity authentication database mapping to obtain a dynamic feature up-regulation instruction, and the positioning error correction amount is used to represent a deviation amount of the single feature point positioning error and the preset positioning error asymptotic upper limit.

9. A face expression recognition based anti-fraud identity authentication method, the face expression recognition based anti-fraud identity authentication method applies the face expression recognition based anti-fraud identity authentication system according to any one of claims 1-8, characterized in that, Comprise: Collecting a face image sequence of a user to be authenticated; Pretreating the face image sequence to obtain feature extraction parameters to quantify the accuracy of the CNN and the LSTM model in the face expression feature extraction process, and obtaining a dynamic feature extraction accuracy evaluation; According to the dynamic feature extraction accuracy evaluation, it is judged whether to perform dynamic feature extraction accuracy optimization, if yes, then the face expression feature comparison link is performed after optimization, otherwise the face expression feature comparison link is directly performed, and the dynamic feature extraction accuracy optimization comprises regional feature weight dynamic adjustment and face region mask weight dynamic adjustment; Obtaining multi-dimension comparison parameters in the face expression feature comparison link to quantify the rationality in the face expression feature comparison process, and obtaining a multi-dimension comparison rationality evaluation; According to the multi-dimension comparison rationality evaluation, it is judged whether to perform multi-dimension comparison rationality optimization, if yes, then the risk control identity authentication decision is generated after optimization, otherwise the risk control identity authentication decision is directly generated, and the multi-dimension comparison rationality optimization comprises transition frame weight dynamic adjustment and static-dynamic feature weight dynamic adjustment.

Images