Dynamic image fusion method for big data and artificial intelligence virtual human panoramic live broadcast

Abstract

The invention relates to the technical field of man-machine interaction, and discloses a dynamic image fusion method for big data and artificial intelligence virtual human panoramic live broadcast, comprising the following steps: collecting live broadcast data; constructing graph structure data; inputting the graph structure data into an emotion recognition model, and outputting a result representing a fan emotion state by the emotion recognition model; inputting the graph structure data into a response model, and outputting a result representing a virtual idol interaction strategy by the response model; inputting second serial data obtained by interaction strategy coding into an auditing model, wherein the auditing model outputs a result representing the security state of the virtual idol interaction strategy; according to the method, the live broadcast information is collected regularly, the intention and the emotional state of the user are comprehensively considered, the interaction strategy of the virtual idol with a good emotional improvement effect and better safety is provided, and compared with the mode that the interaction strategy is provided only according to the intention of the user, the method can balance multiple interaction targets, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of human-computer interaction technology, and more specifically, to a dynamic image fusion method for panoramic live broadcast of virtual humans for big data and artificial intelligence. Background Art

[0002] With the rapid development of technology and the enthusiasm of young people for virtual culture, virtual idols are becoming a new form of entertainment; they not only have carefully designed appearances and personalities, but also have the ability to interact and communicate; however, the current interaction of virtual idols is still relatively simple, unable to give personalized and situational responses, and it is difficult to truly draw the audience into an immersive virtual world;

[0003] Use deep learning technology to analyze fans' voice commands, gestures, etc., and accurately identify their interaction intentions. Based on the identified intentions, combined with the idol's role settings and fans' historical interaction records, dynamically generate personalized voice and action responses that are highly relevant to the current scenario;

[0004] However, in an open interactive environment, malicious users may make inappropriate remarks or actions, leading to inappropriate responses from virtual idols. At the same time, virtual idols can serve as mental health support tools, guiding users to manage emotions and relieve stress through chat or actions, thereby promoting mental health. Summary of the invention

[0005] The purpose of the present invention is to provide a dynamic image fusion method for panoramic live broadcast of virtual humans for big data and artificial intelligence in order to solve the above problems.

[0006] The present invention provides a dynamic image fusion method for panoramic live broadcast of virtual human for big data and artificial intelligence, comprising the following steps:

[0007] Step 100, collecting live broadcast data, which is collected once at each time sampling point, including virtual idol information and fan information;

[0008] Step 200, constructing graph structure data, the graph structure data includes nodes and edges connecting the nodes, and a node represents a virtual idol or fan;

[0009] There are edges between virtual idol nodes in the same live broadcast, and there are edges between virtual idol nodes and all fan nodes in the same live broadcast;

[0010] The node characteristics of the node representing the virtual idol are equal to the virtual idol characteristics, and the node characteristics of the node representing the fan are equal to the fan characteristics;

[0011] Virtual idol characteristics and fan characteristics are obtained through feature engineering;

[0012] The live broadcast data of each node is respectively encoded into No. 1 sequence data, wherein the No. 1 sequence data includes n No. 1 sequence units, and the t-th No. 1 sequence unit represents the live broadcast data of a node at the t-th time sampling point;

[0013] Step 300, inputting the graph structure data into an emotion recognition model, the emotion recognition model comprising a first hidden layer, a second hidden layer and a first output layer, and the emotion recognition model outputs a result representing the emotional state of the fans;

[0014] Step 400, inputting the graph structure data into a response model, the response model comprising a third hidden layer, a fourth hidden layer, a feature fusion layer and a second output layer, and the response model outputs a result representing the virtual idol interaction strategy;

[0015] Step 500, inputting the second sequence data obtained by the interaction strategy encoding into the audit model, the audit model includes a sixth hidden layer and a third output layer, and the audit model outputs a result representing the security status of the virtual idol interaction strategy.

[0016] Furthermore, in the emotion recognition model, sequence data No. 1 is input into the first hidden layer, the first hidden layer outputs the first hidden state to the second hidden layer, the second hidden layer also inputs graph structure data, the second hidden layer outputs the second hidden state to the first output layer and the feature fusion layer, and the first output layer outputs the result representing the emotional state of the fans.

[0017] Furthermore, the loss function calculation method of the emotion recognition model is:

[0018] ;

[0019] in, represents the emotional loss value, is the number of sentiment samples, is the number of emotion categories. There are seven emotion categories, including happiness, sadness, anger, fear, surprise, disgust, and neutral. Emotion samples represent the actual emotion labels of all fans in each emotion category within a time period. The value is 0 or 1. 1 indicates that the fan belongs to the emotion category, and 0 indicates that the fan does not belong to the emotion category. Indicates The first emotion sample The actual sentiment label of the class sentiment, represents the first The first emotion sample The probability of class emotion, Represents the logarithmic function with the natural constant e as the base.

[0020] Furthermore, in the response model, the first sequence data is input into the third hidden layer, the third hidden layer outputs the third hidden state to the feature fusion layer, the feature fusion layer outputs the fifth fusion state to the fourth hidden layer, the fourth hidden layer also inputs the graph structure data, the fourth hidden layer outputs the fourth hidden state to the second output layer, and the second output layer outputs the result representing the virtual idol interaction strategy.

[0021] Furthermore, the steps of training the response model include:

[0022] Step 401, constructing a target network, the target network replicating a response model;

[0023] Step 402: randomly obtain an experience from the experience pool. The experience includes the state at time e. , the action performed at time e , the state at time e+1 , perform the action Rewards ;

[0024] The states at time e and e+1 represent the live data at time e and time e+1 of the response model input respectively;

[0025] award The calculation formula is as follows:

[0026] ;

[0027] in Represents an action to audit the output of a model The safety probability, and are the first and second hyperparameters, Indicates the degree of improvement of user emotions from time e to time e+1;

[0028] Then calculate the state The output of the target network corresponds to The probability value ;

[0029] ;

[0030] express The maximum probability value of the target network output;

[0031] Step 403, according to and Calculate the loss value loss and update the response model;

[0032] ;

[0033] express The output of the next response model corresponds to The probability value of

[0034] is the discount factor, a value between 0 and 1;

[0035] Step 404, updating the target network every fixed number of network updates so that its parameters are the same as those of the current response model;

[0036] Step 405, until the number of collected experiences reaches a set threshold or the response model converges, the step is terminated.

[0037] Further, The difference in emotion scores calculated from the emotion state vectors output by the emotion recognition model at time e and e+1. The emotion score is calculated as follows:

[0038] ;

[0039] in is the emotion score at time e, Is the emotional category The weight of is the emotion predicted by the emotion recognition model The probability of is the number of sentiment categories.

[0040] Furthermore, the second sequence data includes F second sequence units sorted by time, and the mth second sequence unit represents the interaction strategy collected at the mth moment.

[0041] Further, in the audit model, the second sequence data is input into the sixth hidden layer, the sixth hidden layer outputs the sixth hidden state to the third output layer, and the third output layer outputs the result representing the safety state of the virtual idol interaction strategy.

[0042] Furthermore, the loss function calculation method of the audit model is:

[0043] ;

[0044] in Represents the audit loss value, represents the total number of audit samples, It is The real security label of the audit sample has a value of 0 or 1, where 1 means that the interactive strategy of the audit sample is safe, and 0 means that the interactive strategy of the audit sample is unsafe. The audit model predicts Security labels for audit samples, Represents the logarithmic function with the natural constant e as the base.

[0045] The present invention provides a computer storage medium for storing computer-readable instructions, which, when read, can execute the aforementioned dynamic image fusion method for panoramic live broadcast of virtual humans for big data and artificial intelligence.

[0046] The beneficial effects of the present invention are as follows: the present invention collects live broadcast information at regular intervals, comprehensively considers the user's intention and emotional state, and provides an interactive strategy for a virtual idol with good emotion improvement effect and better safety. Compared with providing an interactive strategy based only on the user's intention, this method can balance multiple interactive goals and improve the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of a dynamic image fusion method for panoramic live broadcast of virtual humans for big data and artificial intelligence of the present invention. DETAILED DESCRIPTION

[0048] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0049] At least one embodiment of the present invention discloses a dynamic image fusion method for panoramic live broadcast of virtual human for big data and artificial intelligence, such as Figure 1 As shown, the following steps are included:

[0050] Step 100, collecting live broadcast data, which is collected once at each time sampling point, including virtual idol information and fan information;

[0051] Virtual idol information includes role settings and personality traits;

[0052] Fan information includes voice commands, gestures, text input, and interaction status;

[0053] The time sampling point refers to the moment when data is collected at a predetermined time interval;

[0054] In one embodiment of the present invention, the default value of the predetermined time interval is 1 minute.

[0055] Step 200, constructing graph structure data, the graph structure data includes nodes and edges connecting the nodes, and a node represents a virtual idol or fan;

[0056] There are edges between virtual idol nodes in the same live broadcast, and there are edges between virtual idol nodes and all fan nodes in the same live broadcast;

[0057] The node characteristics of the node representing the virtual idol are equal to the virtual idol characteristics, and the node characteristics of the node representing the fan are equal to the fan characteristics;

[0058] Virtual idol characteristics and fan characteristics are obtained through feature engineering;

[0059] The live broadcast data of each node is respectively encoded into No. 1 sequence data, wherein the No. 1 sequence data includes n No. 1 sequence units, and the t-th No. 1 sequence unit represents the live broadcast data of a node at the t-th time sampling point;

[0060] In one embodiment of the present invention, live broadcast data of a virtual idol is processed by feature engineering to obtain virtual idol features;

[0061] The live broadcast data of virtual idols is in text mode, and the Word Embedding algorithm is used as a feature engineering method;

[0062] In one embodiment of the present invention, the live broadcast data of fans is processed by feature engineering to obtain fan features;

[0063] Fans’ live streaming data includes text, audio, and image modes. For data in various modes, corresponding feature engineering methods are used to encode and generate features, and then splice them.

[0064] The information of text modality is used as a feature engineering method through the Word Embedding algorithm;

[0065] The information of audio modality is used as a feature engineering method through the MFCC (Mel-Frequency Cepstral Coefficient) algorithm;

[0066] The information of image modality is used as a feature engineering method through CNN algorithm;

[0067] Step 300, inputting the graph structure data into the emotion recognition model, the emotion recognition model includes a first hidden layer, a second hidden layer and a first output layer, wherein the first sequence data is input into the first hidden layer, the first hidden layer outputs the first hidden state to the second hidden layer, the second hidden layer also inputs the graph structure data, the second hidden layer outputs the second hidden state to the first output layer and the feature fusion layer, and the first output layer outputs the result representing the emotional state of the fans;

[0068] In one embodiment of the present invention, the formula of the first hidden layer is as follows:

[0069]

[0070] in Represents the tth first hidden state of the vth node of the graph structure data, Represents the t-1th first hidden state of the vth node of the graph structure data, Represents the tth first-order sequence unit of the first-order sequence data corresponding to the vth node of the graph structure data, and are the first and second sequence weight parameters, is the first sequence bias parameter, and tanh is the hyperbolic tangent function.

[0071] In one embodiment of the present invention, the formula of the second hidden layer is as follows:

[0072]

[0073] in represents the second hidden state of the vth node, represents the first graph weight parameter, , , and Represents the nth first hidden state of the vth and uth nodes of the graph structure data, is the set of nodes that have edges with the node v of the graph structure data. express The total number of nodes in ;

[0074]

[0075] in represents the first aggregation coefficient between the vth and uth nodes, tanh represents the hyperbolic tangent function, T represents the transpose, Indicates adjustable parameters; The default value is 1 / N, where N is the dimension of the node features.

[0076] In one embodiment of the present invention, the formula of the first output layer is as follows:

[0077]

[0078] in represents an emotional state vector, and the i-th component of the emotional state vector represents the probability that the fan belongs to the i-th emotion. represents the second hidden state of the vth fan node, is the first weight parameter, is the first bias parameter, Represents a sigmoid function.

[0079] In one embodiment of the present invention, the loss function calculation method of the emotion recognition model is:

[0080]

[0081] in, represents the emotional loss value, is the number of sentiment samples, is the number of emotion categories. There are seven emotion categories, including happiness, sadness, anger, fear, surprise, disgust, and neutral. Emotion samples represent the actual emotion labels of all fans in each emotion category within a time period. The value is 0 or 1. 1 indicates that the fan belongs to the emotion category, and 0 indicates that the fan does not belong to the emotion category. Indicates The first emotion sample The actual sentiment label of the class sentiment, represents the first The first emotion sample The probability of class emotion, Represents the logarithmic function with the natural constant e as the base.

[0082] Step 400, inputting graph structure data into a response model, the response model comprising a third hidden layer, a fourth hidden layer, a feature fusion layer, and a second output layer, wherein the first sequence data is input into the third hidden layer, the third hidden layer outputs a third hidden state to the feature fusion layer, the feature fusion layer outputs a fifth fusion state to the fourth hidden layer, the fourth hidden layer also inputs graph structure data, the fourth hidden layer outputs a fourth hidden state to the second output layer, and the second output layer outputs a result representing the virtual idol interaction strategy;

[0083] In one embodiment of the present invention, the formula of the third hidden layer is as follows:

[0084]

[0085]

[0086]

[0087]

[0088] in, , , represents the weight parameters of the third, fourth and fifth sequences, , , represents the bias parameters of the second, third and fourth sequences, represents the dot product, , and They represent the first, second, and third intermediate states of the vth node of the graph structure data, respectively. Represents the tth first-order sequence unit of the first-order sequence data corresponding to the vth node of the graph structure data, and They represent the t-th and t-1-th third hidden states of the v-th node of the graph structure data, n≥t≥1, n represents the total number of the first sequence units of the first sequence data, and t=1 , tanh is the hyperbolic tangent function, Represents a sigmoid function.

[0089] In one embodiment of the present invention, the formula of the feature fusion layer is as follows:

[0090]

[0091] in, Represents the fifth fusion state of the vth node of the graph structure data, Represents the second hidden state of the vth node of the graph structure data, Represents the nth third hidden state of the vth node of the graph structure data, represents the concatenation function, represents the sum weight matrix, Represents the summed bias parameter.

[0092] In one embodiment of the present invention, the formula of the fourth hidden layer is as follows:

[0093]

[0094] in Represents the fourth hidden state of the vth node of the graph structure data, Represents the second aggregation coefficient of the vth node of the graph structure data, represents the activation function, represents the weight parameter of the second graph;

[0095]

[0096]

[0097]

[0098] in , , and Respectively represent the fifth fusion state of the vth and uth nodes of the graph structure data, and Respectively represent the linear transformation features of the vth and uth nodes of the graph structure data, represents the aggregation weight coefficient, represents the splicing weight coefficient, T represents transposition, represents an exponential function with a natural constant as the base, LeakyRelu represents a modified linear unit function, Represents the set of nodes that have edges with the vth node.

[0099] In one embodiment of the present invention, the formula of the second output layer is as follows:

[0100]

[0101] in Represents an interactive response vector. The dth component value of the interactive response vector represents the probability value of the dth interactive strategy. The interactive strategy with the largest probability value is selected as the output. The interactive strategy group contains all executable interactive strategies. An interactive strategy represents the voice and action response of the virtual idol. Voice includes voice content and voice features, and actions include gestures, facial expressions, and body movements. represents the concatenation of the fourth hidden states of all nodes, M represents the set of all nodes, is the second weight parameter, is the second bias parameter, Represents a sigmoid function.

[0102] In one embodiment of the present invention, the step of training the response model includes:

[0103] Step 401, constructing a target network, the target network replicating a response model;

[0104] Step 402: randomly obtain an experience from the experience pool, the experience includes the state at time e , the action performed at time e , the state at time e+1 , perform the action Rewards ;

[0105] The states at time e and e+1 represent the live data at time e and time e+1 of the response model input respectively;

[0106] Then calculate the state The target network output corresponds to The probability value ;

[0107]

[0108] express The maximum probability value of the target network output;

[0109] Step 403, according to and Calculate the loss value loss and update the response model;

[0110]

[0111] express The output of the next response model corresponds to The probability value of

[0112] is the discount factor, a value between 0 and 1;

[0113] Step 404, updating the target network every fixed number of network updates so that its parameters are the same as those of the current response model;

[0114] Step 405, until the number of collected experiences reaches a set threshold or the response model converges, the step is terminated.

[0115] In one embodiment of the present invention, the experience is derived from the record of historical action execution.

[0116] In one embodiment of the present invention, the reward The calculation formula is as follows:

[0117]

[0118] in Represents an action to audit the output of a model The safety probability, and are the first and second hyperparameters, Indicates the degree of improvement of the user's mood from time e to time e+1.

[0119] In one embodiment of the present invention, The difference in emotion scores calculated from the emotion state vectors output by the emotion recognition model at time e and e+1. The emotion score is calculated as follows:

[0120]

[0121] in is the emotion score at time e, Is the emotional category The weight of is the emotion predicted by the emotion recognition model The probability of is the number of sentiment categories.

[0122] Step 500, inputting the second sequence data into the audit model, the audit model comprising a sixth hidden layer and a third output layer, wherein the second sequence data is input into the sixth hidden layer, the sixth hidden layer outputs a sixth hidden state to the third output layer, and the third output layer outputs a result indicating a safe state of the virtual idol interaction strategy;

[0123] In one embodiment of the present invention, the interaction strategy is encoded as second-order sequence data; the second-order sequence data includes F second-order sequence units sorted by time, and the mth second-order sequence unit represents the interaction strategy collected at the mth moment.

[0124] In one embodiment of the present invention, the formula of the sixth hidden layer is as follows:

[0125]

[0126] in represents the mth sixth hidden state, represents the m-1th sixth hidden state, represents the mth second sequence unit of the second sequence data, and is the sixth and seventh sequence weight parameter, is the fifth sequence bias parameter, and tanh is the hyperbolic tangent function.

[0127] In one embodiment of the present invention, the formula of the third output layer is as follows:

[0128]

[0129] in represents a safe state vector, which represents the safety probability of the interaction strategy. represents the Fth sixth hidden state, is the third weight parameter, is the third bias parameter, Represents a sigmoid function.

[0130] In one embodiment of the present invention, the loss function calculation method of the audit model is:

[0131]

[0132] in Represents the audit loss value, represents the total number of audit samples, It is The real security label of the audit sample has a value of 0 or 1, where 1 means that the interactive strategy of the audit sample is safe, and 0 means that the interactive strategy of the audit sample is unsafe. The audit model predicts Security labels for audit samples, Represents the logarithmic function with the natural constant e as the base.

[0133] In at least one embodiment of the present invention, a computer storage medium is provided for storing computer-readable instructions, which, when read, can execute the aforementioned dynamic image fusion method for panoramic live broadcast of virtual humans for big data and artificial intelligence.

[0134] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are within the protection of this embodiment.

Claims

1. A dynamic image fusion method for panoramic live broadcast of virtual humans for big data and artificial intelligence, characterized in that: include: Step 100, collecting live broadcast data, which is collected once at each time sampling point, including virtual idol information and fan information; Step 200, constructing graph structure data, the graph structure data includes nodes and edges connecting the nodes, and a node represents a virtual idol or fan; There are edges between virtual idol nodes in the same live broadcast, and there are edges between virtual idol nodes and all fan nodes in the same live broadcast; The node characteristics of the node representing the virtual idol are equal to the virtual idol characteristics, and the node characteristics of the node representing the fan are equal to the fan characteristics; Virtual idol characteristics and fan characteristics are obtained through feature engineering; The live broadcast data of each node is respectively encoded into No. 1 sequence data, wherein the No. 1 sequence data includes n No. 1 sequence units, and the t-th No. 1 sequence unit represents the live broadcast data of a node at the t-th time sampling point; Step 300, inputting the graph structure data into an emotion recognition model, the emotion recognition model comprising a first hidden layer, a second hidden layer and a first output layer, and the emotion recognition model outputs a result representing the emotional state of the fans; Step 400, inputting the graph structure data into a response model, the response model comprising a third hidden layer, a fourth hidden layer, a feature fusion layer and a second output layer, and the response model outputs a result representing the virtual idol interaction strategy; Step 500, inputting the second sequence data obtained by the interaction strategy encoding into the audit model, the audit model includes a sixth hidden layer and a third output layer, and the audit model outputs a result representing the security status of the virtual idol interaction strategy.

2. According to claim 1, a dynamic image fusion method for virtual human panoramic live broadcast for big data and artificial intelligence is characterized in that: In the emotion recognition model, sequence data No. 1 is input into the first hidden layer, the first hidden layer outputs the first hidden state to the second hidden layer, the second hidden layer also inputs graph structure data, the second hidden layer outputs the second hidden state to the first output layer and the feature fusion layer, and the first output layer outputs the result representing the emotional state of the fans.

3. The dynamic image fusion method for panoramic live broadcast of virtual human for big data and artificial intelligence according to claim 2 is characterized in that: The loss function calculation method of the emotion recognition model is: ； in, represents the emotional loss value, is the number of sentiment samples, is the number of emotion categories. There are seven emotion categories, including happiness, sadness, anger, fear, surprise, disgust, and neutral. Emotion samples represent the actual emotion labels of all fans in each emotion category within a time period. The value is 0 or 1. 1 indicates that the fan belongs to the emotion category, and 0 indicates that the fan does not belong to the emotion category. Indicates The first emotion sample The actual sentiment label of the class sentiment, represents the first The first emotion sample The probability of class emotion, Represents the logarithmic function with the natural constant e as the base.

4. The dynamic image fusion method for panoramic live broadcast of virtual human for big data and artificial intelligence according to claim 1 is characterized in that: In the response model, the first sequence data is input into the third hidden layer, the third hidden layer outputs the third hidden state to the feature fusion layer, the feature fusion layer outputs the fifth fusion state to the fourth hidden layer, the fourth hidden layer also inputs the graph structure data, the fourth hidden layer outputs the fourth hidden state to the second output layer, and the second output layer outputs the result representing the virtual idol interaction strategy.

5. The method for dynamic image fusion of virtual human panoramic live broadcast for big data and artificial intelligence according to claim 4 is characterized in that: The steps of training the response model include: Step 401, constructing a target network, the target network replicating a response model; Step 402: randomly obtain an experience from the experience pool, the experience includes the state at time e , the action performed at time e , the state at time e+1 , perform the action Rewards ; The states at time e and e+1 represent the live data at time e and time e+1 of the response model input respectively; award The calculation formula is as follows: ； in Represents an action to audit the output of a model The safety probability, and are the first and second hyperparameters, Indicates the degree of improvement of user emotions from time e to time e+1; Then calculate the state The target network output corresponds to The probability value ; ； express The maximum probability value of the target network output; Step 403, according to and Calculate the loss value loss and update the response model; ； express The output of the next response model corresponds to The probability value of is the discount factor, a value between 0 and 1; Step 404, updating the target network every fixed number of network updates so that its parameters are the same as those of the current response model; Step 405, until the number of collected experiences reaches a set threshold or the response model converges, the step is terminated.

6. The method for dynamic image fusion of virtual human panoramic live broadcast for big data and artificial intelligence according to claim 5 is characterized in that: The difference in emotion scores calculated from the emotion state vectors output by the emotion recognition model at time e and e+1. The emotion score is calculated as follows: ； in is the emotion score at time e, Is the emotional category The weight of is the emotion predicted by the emotion recognition model The probability of is the number of sentiment categories.

7. The method for dynamic image fusion of virtual human panoramic live broadcast for big data and artificial intelligence according to claim 1 is characterized in that: The second sequence data includes F second sequence units sorted by time, and the mth second sequence unit represents the interaction strategy collected at the mth moment.

8. The method for dynamic image fusion of virtual human panoramic live broadcast for big data and artificial intelligence according to claim 1 is characterized in that: In the audit model, the second sequence data is input into the sixth hidden layer, the sixth hidden layer outputs the sixth hidden state to the third output layer, and the third output layer outputs the result representing the security state of the virtual idol interaction strategy.

9. The method for dynamic image fusion of virtual human panoramic live broadcast for big data and artificial intelligence according to claim 8, characterized in that: The loss function calculation method of the audit model is: ； in Represents the audit loss value, represents the total number of audit samples, It is The real security label of the audit sample has a value of 0 or 1, where 1 means that the interactive strategy of the audit sample is safe, and 0 means that the interactive strategy of the audit sample is unsafe. The audit model predicts Security labels for audit samples, Represents the logarithmic function with the natural constant e as the base.

10. A computer storage medium, characterized in that: It is used to store computer-readable instructions, which, when read, can execute a dynamic image fusion method for panoramic live broadcast of virtual humans for big data and artificial intelligence as described in any one of claims 1-9.

Images