A satisfaction evaluation system integrating AI facial expression recognition and subjective scoring

Abstract

The present invention relates to a satisfaction evaluation system integrating AI facial expression recognition and subjective scoring, wherein a video monitoring unit obtains a facial video stream of a subject, a recognition frame extraction unit extracts one or more frames in the facial video stream for recognition; a face image processing unit obtains information of the subject based on the recognition frame, extracts facial expression features and performs facial expression recognition to obtain primary satisfaction, a satisfaction evaluation unit obtains subjective scoring information of the subject, obtains secondary satisfaction, integrates primary satisfaction and secondary satisfaction, and a display unit outputs and displays the integrated satisfaction evaluation result. The present invention integrates the evaluation results of AI facial expression recognition and subjective scoring, and the two can verify each other and complement each other, thereby improving the accuracy and objectivity of satisfaction evaluation, and overcoming the disadvantages of electroencephalogram analysis method and eye tracking method in sensory engineering that are easily affected by noise, the resolution of the signal is not high, and the result accuracy of the subjective evaluation method is poor.

Description

Technical Field

[0001] The present invention relates to the technical field of data recognition, data representation, recording medium and recording medium processing, and in particular to a satisfaction evaluation system integrating AI facial expression recognition and subjective scoring in the technical field of computer control. Background Art

[0002] Satisfaction evaluation uses whether the needs of service recipients are met as an objective evaluation measure. It is a method of final quality evaluation. It is applicable to satisfaction surveys at the national and industry levels and can significantly and effectively manage and improve service levels.

[0003] In the existing technology, satisfaction evaluation adopts a variety of collection methods, including service implementation survey, perceived quality survey, satisfaction index model survey, satisfaction and dissatisfaction survey, satisfaction and shortcoming improvement, satisfaction and KANO analysis, satisfaction and U&A research, satisfaction and excellent service research, satisfaction and user experience research, satisfaction and service management research; through satisfaction surveys, managers can establish VOC systems, service improvement systems, and service performance evaluation systems to easily and effectively listen to customer voices, understand service status, discover service shortcomings, evaluate service performance, and promote service improvement.

[0004] However, most of the satisfaction evaluations in the existing technologies are implemented in a subjective scoring manner, which is highly subjective. There is a situation where users deliberately conceal their true evaluations, and thus the actual value of the product or service cannot be truly reflected due to overly high or low scores.

[0005] With the advancement of computer technology, the concept of Kansei Engineering has begun to be applied in the context of satisfaction evaluation. It translates sensibility into engineering, and can transform the results of human sensory analysis into physical design elements of products, and then manufacture products according to people's preferences. In actual applications, the types of measurement data involved in Kansei Engineering include brain waves, gaze tracking, electromyography, facial expressions, behaviors and many other physiological manifestations of people. Of course, questionnaire surveys as a classic form are also one of the objects, but because brain waves, gaze tracking, electromyography are easily affected by noise, the resolution of signals is not high, and the results of subjective evaluation methods such as facial expressions and behaviors are inaccurate, therefore, although there is no fixed method for the research on Kansei Engineering, it still needs to be promoted in a scientific way to gradually obtain more accurate sensory analysis results of the subjects. Summary of the invention

[0006] The present invention solves the problems existing in the prior art and provides an optimized satisfaction evaluation system integrating AI facial expression recognition and subjective scoring.

[0007] The technical solution adopted by the present invention is a satisfaction evaluation system integrating AI facial expression recognition and subjective scoring, the system comprising:

[0008] A video monitoring unit, used to obtain a facial video stream of the subject;

[0009] an identification frame extraction unit, used to extract one or more frames from the facial video stream for identification;

[0010] A face image processing unit, used to obtain the subject's information based on the recognition frame, extract facial expression features and perform facial expression recognition to obtain a primary satisfaction level M;

[0011] A satisfaction evaluation unit is used to obtain the subjective scoring information of the subjects and obtain the secondary satisfaction M', integrating the primary satisfaction and the secondary satisfaction;

[0012] A display unit is used to output and display the satisfaction evaluation result S after the primary satisfaction and the secondary satisfaction are integrated.

[0013] Preferably, a facial expression recognition model is provided in conjunction with the face image processing unit;

[0014] The facial expression recognition model includes:

[0015] The first recognition model includes an Alexnet network and a VGG19 network arranged in parallel, and an attention mechanism network and a classification decision function are sequentially arranged at the output ends of the Alexnet network and the VGG19 network;

[0016] The second recognition model is an Alexnet network, wherein the Alexnet network is connected with a SVM classifier;

[0017] The third recognition model is a VGG19 network, and the VGG19 network is connected with a SVM classifier.

[0018] Preferably, the attention mechanism network fuses the outputs of the Alexnet network and the VGG19 network, satisfying qL(K)=q(K)+J(K);

[0019] q(K) and J(K) are global features and local features respectively, q(K) = sig(Ba(D(K))), J(K) = sig(Ba(K));

[0020]

[0021]

[0022]

[0023] K=wA V Alexnet +w g V vgg ,

[0024] Among them, V Alexnet and V vgg They are the fc7 layer output of the Alexnet network and the penultimate layer fc-4096 output of the VGG19 network, and both are 4096*1 vectors. A and w g is the corresponding weight, L A and L g are the lengths of the vectors output by the fc7 layer of the Alexnet network and the penultimate fc-4096 layer of the VGG19 network, i is the sequence number of each element of the vectors output by the fc7 layer of the Alexnet network and the penultimate fc-4096 layer of the VGG19 network, and w r To find D(K r ) is the weight of the rth element of the corresponding vector in the process; μ A and σ A are the mean and variance of the elements of the D(K) vector, μ g and σ g are the mean and variance of the elements of the K vector respectively;

[0025] The classification decision function Among them, w i is the weight corresponding to the output of qL(K), m is an integer from 1 to p, and p is the number of expression categories.

[0026] Preferably, for each recognition frame, a copy is made and then respectively input into the Alexnet network and the VGG19 network of the first recognition model, the output of the fc7 layer of the Alexnet network and the output of the penultimate layer fc-4096 of the VGG19 network are input into the attention mechanism network, and the expression features output by the attention mechanism network are classified by the classification decision function, and the result with the highest probability after classification is taken as the first facial expression recognition result A;

[0027] For each recognition frame, the fc7 feature of the face corresponding to the recognition frame is obtained by using the Alexnet network and classified by using the SVM classifier to obtain the second facial expression recognition result B;

[0028] The penultimate layer fc-4096 features of the face corresponding to the recognition frame are obtained by using the VGG19 network and classified by the SVM classifier to obtain the third facial expression recognition result C;

[0029] A, B, and C are determined at the decision layer. When B is not equal to C, A is used as the facial expression recognition result; when B is equal to C, B is used as the facial expression recognition result.

[0030] Preferably, in the face image processing unit, satisfaction weights are assigned to the satisfied emotions, neutral emotions and dissatisfied emotions of the facial expression recognition results respectively; L recognition frames of the subject within a certain time T are obtained, and the primary satisfaction M is calculated,

[0031]

[0032]

[0033] Among them, λ 1 and λ 2 are the satisfaction weights corresponding to satisfied emotions and neutral emotions respectively, θ(t) and φ(t) are the satisfaction values ​​assigned when in satisfied emotions and neutral emotions respectively.

[0034] Preferably, when the facial expression recognition result corresponding to any recognition frame is a satisfied emotion, θ(t) is 1;

[0035] When the facial expression recognition result corresponding to any recognition frame is dissatisfaction, θ(t) is -1;

[0036] In other cases, θ(t) is 0;

[0037] When the facial expression recognition result corresponding to any recognition frame is neutral emotion, φ(t) is 1, otherwise φ(t) is 0.

[0038] Preferably, λ 1 is 1, λ 2 <0.03.

[0039] Preferably, S=αM+βM′, wherein −1≤M′≤1.

[0040] Preferably, if F≥1 or F≤-1, α=1, β=0; otherwise, α=0, β=0.1.

[0041] Preferably, the video surveillance unit includes several cameras, and a comparison unit is provided in conjunction with the video surveillance unit. The comparison unit uses VLAN to compare the similarity between the avatar corresponding to the identification frame of the person being collected extracted at the same time and the built-in standard avatar, and takes the avatar corresponding to the identification frame with the highest similarity as the face image; and pre-processes the identification frame.

[0042] The present invention provides an optimized satisfaction evaluation system integrating AI facial expression recognition and subjective scoring. A video monitoring unit is used to obtain a facial video stream of a subject, and a recognition frame extraction unit is used to extract one or more frames in the facial video stream for recognition. A face image processing unit is used to obtain subject information based on recognition frames, extract facial expression features, and perform facial expression recognition to obtain primary satisfaction M. A satisfaction evaluation unit is used to obtain the subject's subjective scoring information and obtain secondary satisfaction M', and the primary satisfaction and secondary satisfaction are integrated. A display unit is used to output a satisfaction evaluation result S after the primary satisfaction and secondary satisfaction are integrated.

[0043] The beneficial effect of the present invention is that the evaluation results of AI facial expression recognition and subjective scoring are integrated, and the two can verify each other and complement each other, thereby improving the accuracy and objectivity of satisfaction evaluation, overcoming the disadvantages of electroencephalogram analysis method and eye tracking method in sensory engineering that are easily affected by noise, the signal resolution is not high, and the result accuracy of subjective evaluation method is poor. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0045] Figure 2 It is a structural diagram of the facial expression recognition model of the present invention;

[0046] Figure 3 This is an application structure diagram of the first recognition model in the present invention. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto.

[0048] The present invention relates to a satisfaction evaluation system integrating AI facial expression recognition and subjective scoring, the system comprising:

[0049] A video monitoring unit, used to obtain a facial video stream of the subject;

[0050] The video surveillance unit includes several cameras, and a comparison unit is provided in conjunction with the video surveillance unit. The comparison unit uses VLAN to compare the similarity between the avatar corresponding to the recognition frame of the person being collected and extracted at the same time and the built-in standard avatar, and takes the avatar corresponding to the recognition frame with the highest similarity as the face image; and pre-processes the recognition frame.

[0051] In the present invention, since it is impossible to ensure that the subject is facing the display with his front face at all times, in order to obtain the subject's front face image as much as possible, there are at least three cameras, which simultaneously capture the subject's facial expression once every 2 seconds, and use the camera in the middle in front of the subject to calibrate the subject's front face, ensuring that the camera in the middle can capture the subject's front face to the maximum extent.

[0052] an identification frame extraction unit, used to extract one or more frames from the facial video stream for identification;

[0053] In the present invention, one or more recognition frames are automatically extracted from the video stream collected by the video monitoring unit, emotion recognition is performed on the recognition frames, and the final result based on machine recognition is obtained.

[0054] In the present invention, the facial expressions of the subjects are summarized as satisfaction, neutrality, and disgust; when the subject is satisfied with the evaluation object, his eyebrows are symmetrically spread, and the corners of his mouth and the muscles on both sides are bent downward, indicating that his expression is satisfied; when the subject is dissatisfied with the evaluation object, his eyebrows are tightened and the corners of his mouth are asymmetrical, indicating that his expression is disgusted (dissatisfied); when the subject has no feelings for the evaluation object, he will show an expressionless state, indicating that he has a neutral expression; based on the above summary of facial expression types, the facial expressions of face images can be recognized through a deep learning model.

[0055] In the present invention, the frontal face portraits of all the subjects are recorded in advance as standard portraits. Generally, only one subject is allowed to be evaluated at the same time. Three cameras extract key frame face portraits of the subjects at the same time. It is planned to use VLAN algorithm to compare the similarity of the key frame portraits and standard portraits of the subjects extracted at the same time, and take the one with the highest similarity as the face image of the subject for extraction; the face image is denoised and normalized, the face area features are extracted, and facial emotion recognition is performed.

[0056] A face image processing unit, used to obtain the subject's information based on the recognition frame, extract facial expression features and perform facial expression recognition to obtain a primary satisfaction level M;

[0057] A facial expression recognition model is provided in conjunction with the face image processing unit;

[0058] The facial expression recognition model includes:

[0059] The first recognition model includes an Alexnet network and a VGG19 network arranged in parallel, and an attention mechanism network and a classification decision function are sequentially arranged at the output ends of the Alexnet network and the VGG19 network;

[0060] The attention mechanism network fuses the outputs of the Alexnet network and the VGG19 network, satisfying qL(K)=q(K)+J(K);

[0061] q(K) and J(K) are global features and local features respectively, q(K) = sig(Ba(D(K))), J(K) = sig(Ba(K));

[0062]

[0063]

[0064]

[0065] K=w A V Alexnet +w g V vgg ,

[0066] Among them, V Alexnet and V vgg They are the fc7 layer output of the Alexnet network and the penultimate layer fc-4096 output of the VGG19 network, and both are 4096*1 vectors. A and w g is the corresponding weight, L A and L g are the lengths of the vectors output by the fc7 layer of the Alexnet network and the penultimate fc-4096 layer of the VGG19 network, i is the sequence number of each element of the vectors output by the fc7 layer of the Alexnet network and the penultimate fc-4096 layer of the VGG19 network, and w r To find D(K r ) is the weight of the rth element of the corresponding vector in the process; μ A and σ A are the mean and variance of the elements of the vector output by the fc7 layer of the Alexnet network, μ g and σ g The mean and variance of the elements of the vector output by the penultimate layer fc-4096 of the VGG19 network; r is 1 to L A A positive integer between 1 and L g A positive integer.

[0067] The classification decision function Among them, w i is the weight corresponding to the output of qL(K), m is an integer from 1 to p, and p is the number of expression categories.

[0068] The second recognition model is an Alexnet network, wherein the Alexnet network is connected with a SVM classifier;

[0069] The third recognition model is a VGG19 network, and the VGG19 network is connected with a SVM classifier.

[0070] For each recognition frame, a copy is made and then input into the Alexnet network and VGG19 network of the first recognition model respectively. The output of the fc7 layer of the Alexnet network and the output of the penultimate fc-4096 layer of the VGG19 network are input into the attention mechanism network together. The expression features output by the attention mechanism network are classified by the classification decision function, and the result with the highest probability after classification is taken as the first facial expression recognition result A;

[0071] For each recognition frame, the fc7 feature of the face corresponding to the recognition frame is obtained by using the Alexnet network and classified by using the SVM classifier to obtain the second facial expression recognition result B;

[0072] The penultimate layer fc-4096 features of the face corresponding to the recognition frame are obtained by using the VGG19 network and classified by the SVM classifier to obtain the third facial expression recognition result C;

[0073] A, B, and C are determined at the decision layer. When B is not equal to C, A is used as the facial expression recognition result; when B is equal to C, B is used as the facial expression recognition result.

[0074] In the present invention, the output of the fc7 layer of the Alexnet network and the output of the penultimate fc-4096 layer of the VGG19 network are fused with an attention mechanism network and used as the input of the classification decision function, and the output of the classification decision function is used as the first facial expression recognition result A;

[0075] The attention mechanism network combines the global features and local features corresponding to each element in the vector output by the Alexnet network and the VGG19 network to obtain a more accurate feature fusion result. The calculation of the mean and variance is a well-known technique in the art, namely, where x i The elements in the sequence for which the mean and variance are to be calculated; D(K) is global feature fusion, K is local feature fusion, sig(a) represents Sigmoid function, and Ba(x) represents z-score standardization;

[0076] The classification decision function makes a category judgment on each input fusion feature;

[0077] Finally, the category corresponding to the maximum category probability is taken as the first facial expression recognition result A. Generally speaking, the expression categories include satisfied expression, neutral expression and disgusted (unsatisfied) expression;

[0078] The first recognition model makes full use of the fusion of the features of the vectors output by the Alexnet network and the VGG19 network to improve the accuracy of classification.

[0079] In the present invention, before using the second recognition model and the third recognition model, the last layer recognition categories of Alexnet and VGG19 are changed into three categories, and then they are respectively trained in the training set for the recognition of satisfied expression face images, neutral expression face images, and disgust (dissatisfied) expression face images, so as to obtain their final parameters; and then during recognition, the Alexnet network of the second recognition model is used to obtain the fc7 features of the face corresponding to the recognition frame and is classified by an SVM classifier to obtain the second facial expression recognition result B, and the VGG19 network of the third recognition model is used to obtain the penultimate layer fc-4096 features of the face corresponding to the recognition frame and is classified by an SVM classifier to obtain the third facial expression recognition result C.

[0080] In the present invention, the facial expression recognition result Q corresponding to the decision layer satisfies:

[0081]

[0082] In the face image processing unit, satisfaction weights are assigned to the satisfied emotions, neutral emotions and dissatisfied emotions of the facial expression recognition results respectively; L recognition frames of the subject within a certain time T are obtained, and the primary satisfaction M is calculated.

[0083]

[0084]

[0085] Among them, λ 1 and λ 2 are the satisfaction weights corresponding to satisfied emotions and neutral emotions respectively, θ(t) and φ(t) are the satisfaction values ​​assigned when in satisfied emotions and neutral emotions respectively.

[0086] When the facial expression recognition result corresponding to any recognition frame is a satisfied emotion, θ(t) is 1;

[0087] When the facial expression recognition result corresponding to any recognition frame is dissatisfaction, θ(t) is -1;

[0088] In other cases, θ(t) is 0;

[0089] When the facial expression recognition result corresponding to any recognition frame is neutral emotion, φ(t) is 1, otherwise φ(t) is 0.

[0090] λ 1 is 1, λ 2 <0.03.

[0091] In the present invention, when satisfaction or disgust is identified, it means that the person does not have the problem of not expressing emotions clearly, and it also means that this system is suitable for automatically identifying his satisfaction directly, so the corresponding weight is relatively high, that is, λ 1 is 1, and the value of θ(t) is more explicit (1 or -1);

[0092] When a neutral emotion is recognized, it means that the person may be introverted and not good at expressing emotions. In this case, facial expression recognition alone will result in errors or inaccuracies, and subjective scoring is required for verification. Based on this, λ 2 A setting less than 0.03 is more appropriate, in which case the subjective score will be more valuable.

[0093] In the present invention, since the system generally acquires multiple recognition frames, the facial expression of the observer is acquired once, and the satisfaction measurement of a person can be completed within 1 minute (generally, the acquisition is performed once every 2 seconds). Therefore, for the convenience of calculation, λ 2 Just set it to 0.01.

[0094] In the present invention, generally speaking, -1≤M≤1.

[0095] A satisfaction evaluation unit is used to obtain the subjective scoring information of the subjects and obtain the secondary satisfaction M', integrating the primary satisfaction and the secondary satisfaction;

[0096] In the present invention, the final satisfaction is obtained by integrating facial expression recognition technology and subjective scoring evaluation method.

[0097] A display unit is used to output and display the satisfaction evaluation result S after the primary satisfaction and the secondary satisfaction are integrated.

[0098] S=αM+βM', where -1≤M'≤1.

[0099] If F≥1 or F≤-1, then α=1, β=0; otherwise, α=0, β=0.1.

[0100] In the present invention, when F≥1 or F≤-1, it can be explained that the expression of the subject is obvious, and it is suitable to automatically obtain his satisfaction by facial expression recognition technology. Therefore, α=1, M must be included, and the subjective score M' may not play a role;

[0101] When F is greater than -1 and less than 1, it means that the expression of the subject is not obvious, and it is not suitable to automatically obtain his satisfaction by facial expression recognition technology alone. Subjective scoring M' is needed to assist in obtaining his satisfaction index. Therefore, α is 0 and β is included in this case; but at the same time, subjective scoring is obviously not objective, and there is a high possibility of data falsification. Therefore, its weight is smaller than that of direct panel expression recognition and is only set to 0.1. This error or inaccuracy of the result will be offset in the end through the accumulation of a large amount of data, making the evaluation results of the system tend to be accurate and stable.

[0102] In the present invention, the user can view the final result of the satisfaction evaluation through a display unit, such as a large screen.

Claims

1. A satisfaction evaluation system integrating AI facial expression recognition and subjective scoring, characterized by: The system comprises: A video monitoring unit, used to obtain a facial video stream of the subject; an identification frame extraction unit, used to extract one or more frames from the facial video stream for identification; A face image processing unit is used to obtain the subject information based on the recognition frame, extract the facial expression features and perform facial expression recognition; a facial expression recognition model is provided in conjunction with the face image processing unit; the facial expression recognition model includes: The first recognition model includes an Alexnet network and a VGG19 network arranged in parallel, and an attention mechanism network and a classification decision function are sequentially arranged at the output ends of the Alexnet network and the VGG19 network; for each recognition frame, a copy is made and input into the Alexnet network and the VGG19 network of the first recognition model respectively, the output of the fc7 layer of the Alexnet network and the output of the penultimate layer fc-4096 of the VGG19 network are input into the attention mechanism network, the expression features output by the attention mechanism network are classified by the classification decision function, and the result with the largest probability after classification is taken as the first facial expression recognition result A; The second recognition model is an Alexnet network, and the Alexnet network is connected with an SVM classifier; for each recognition frame, the fc7 feature of the face corresponding to the recognition frame is obtained by the Alexnet network and classified by the SVM classifier to obtain a second facial expression recognition result B; The third recognition model is a VGG19 network, wherein the VGG19 network is connected with an SVM classifier; the penultimate layer fc-4096 features of the face corresponding to the recognition frame are obtained by the VGG19 network and classified by the SVM classifier to obtain a third facial expression recognition result C; A, B, and C are determined in the decision layer. When B is not equal to C, A is used as the facial expression recognition result; when B is equal to C, B is used as the facial expression recognition result; Get primary satisfaction M; A satisfaction evaluation unit is used to obtain the subjective scoring information of the subjects and obtain the secondary satisfaction M', integrating the primary satisfaction and the secondary satisfaction; A display unit is used to output and display the satisfaction evaluation result S after the primary satisfaction and the secondary satisfaction are integrated.

2. A satisfaction evaluation system integrating AI facial expression and subjective scoring according to claim 1, characterized in that: The attention mechanism network fuses the outputs of the Alexnet network and the VGG19 network, satisfying qL(K)=q(K)+J(K); q(K) and J(K) are global features and local features respectively, q(K)=sig(Ba(D(K))), J(K)= sig(Ba(K)); ， ； ， ； ， K=w A V Alexnet +w g V vgg , Among them, V Alexnet and V vgg They are the fc7 layer output of the Alexnet network and the penultimate layer fc-4096 output of the VGG19 network, and both are 4096*1 vectors. A and w g is the corresponding weight, and are the lengths of the vectors output by the fc7 layer of the Alexnet network and the penultimate fc-4096 layer of the VGG19 network, respectively; i is the sequence number of each element of the vectors output by the fc7 layer of the Alexnet network and the penultimate fc-4096 layer of the VGG19 network, For The weight of the rth element of the corresponding vector in the process; and are the mean and variance of the elements of the vector output by the fc7 layer of the Alexnet network, and They are the mean and variance of the elements of the vector output by the penultimate fc-4096 layer of the VGG19 network; The classification decision function ,in, is the weight corresponding to the output of qL(K), m is an integer from 1 to p, and p is the number of expression categories.

3. The satisfaction evaluation system integrating AI facial expression and subjective scoring according to claim 1, characterized in that: In the face image processing unit, satisfaction weights are assigned to the satisfied emotions, neutral emotions and dissatisfied emotions of the facial expression recognition results respectively; L recognition frames of the subject within a certain time T are obtained, and the primary satisfaction M is calculated. ， ， in, and are the satisfaction weights corresponding to satisfied emotions and neutral emotions, and Assign values ​​to satisfaction when in satisfied mood and neutral mood respectively.

4. The satisfaction evaluation system integrating AI facial expression and subjective scoring according to claim 3, characterized in that: When the facial expression recognition result corresponding to any recognition frame is a satisfied emotion, is 1; When the facial expression recognition result corresponding to any recognition frame is dissatisfied, is -1; In other cases, is 0; When the facial expression recognition result corresponding to any recognition frame is neutral emotion, is 1, otherwise is 0.

5. The satisfaction evaluation system integrating AI facial expression and subjective scoring according to claim 3 is characterized in that: is 1, <0.

03.

6. A satisfaction evaluation system integrating AI facial expression and subjective scoring according to claim 3, characterized in that: S=αM+βM', where -1≤M'≤1.

7. A satisfaction evaluation system integrating AI facial expression and subjective scoring according to claim 6, characterized in that: If F≥1 or F≤-1, then α=1, β=0; otherwise, α=0, β=0.

1.

8. The satisfaction evaluation system integrating AI facial expression and subjective scoring according to claim 1, characterized in that: The video surveillance unit includes several cameras, and a comparison unit is provided in conjunction with the video surveillance unit. The comparison unit uses VLAN to compare the similarity between the avatar corresponding to the recognition frame of the person being collected and extracted at the same time and the built-in standard avatar, and takes the avatar corresponding to the recognition frame with the highest similarity as the face image; and pre-processes the recognition frame.

Images