An AI-driven video content automatic review method and system

Abstract

The application discloses an AI driving-based video content automatic auditing method and system, relates to the technical field of video content analysis, and comprises the following steps: acquiring video information to be audited, and performing multi-level segmentation on the video information; in the slice, node simulation is performed according to the content independence of the video information; the analysis of associated branches and historical characteristics is performed on each node to generate attribute characteristics of the node; content auditing is performed through the auditing strategy of multiple virtual intelligent agents, and the analysis of a violation space is performed according to unstable parts in an auditing result; and a final video content auditing report is generated according to the auditing result and the analysis result of the violation space. The application improves the structured identification and intelligent auditing efficiency of multi-modal video content, realizes accurate positioning, label attribution and automatic early warning of complex violation behaviors, and enhances the content security management capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of video content analysis, and in particular to an AI-driven video content automatic auditing method and system. BACKGROUND

[0002] With the rapid popularization of mobile Internet and smart terminals, the quantity and influence of video content on social, entertainment, education and other platforms are increasing. The richness and diversity of network video content have promoted the efficiency of information dissemination, but have also brought a series of social governance problems such as content compliance, information security, and youth protection. Especially in the scenes of short video, live broadcast, and video sharing, video content often contains multi-modal information such as pictures, audio, subtitles, and dubbing. Traditional rule-based automatic auditing or manual review methods are difficult to cope with the rapid screening and determination of large-scale content.

[0003] Most existing video auditing technologies focus on image recognition or text analysis, lacking multi-modal deep collaborative understanding and cross-modal consistency judgment capabilities. With the increasing complexity of video content editing methods, editing, splicing, and special effects are widely used, and the hiding, variation, and camouflage of illegal information show stronger concealment and diversity. Single modal or single level auditing methods are difficult to capture the complex semantic relationship between segments and cannot achieve a balance between global structure and local details.

[0004] In addition, how to improve the real-time, accuracy and scalability of content auditing, reduce the cost of manual intervention, and improve the self-adaptive ability of the model to new illegal behaviors has become an important technical direction in the industry. In the face of increasingly complex application scenarios and compliance requirements, building a more intelligent, structured, and multi-level collaborative video content auditing system has become a major technical challenge for the industry. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] Therefore, the present application provides an AI-driven video content automatic auditing method to solve the problems of multi-modal content difficult to be collaboratively analyzed, node segmentation and structured modeling insufficient, complex illegal behavior recognition not comprehensive, and content auditing efficiency and accuracy to be improved.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] In a first aspect, the present application provides an AI-driven video content automatic auditing method, which comprises: acquiring video information to be audited, and performing multi-level segmentation on the video information;

[0009] In the slice, the node simulation is performed according to the content independence of the video information; and the analysis of the associated branch and the historical characteristics of each node is performed to generate the attribute characteristics of the node;

[0010] The content is audited through the auditing strategy of the multiple virtual agents, and the analysis of the violation space is performed according to the unstable part in the auditing result;

[0011] According to the analysis result of the auditing result and the violation space, a final video content auditing report is generated;

[0012] The auditing strategy of the multiple virtual agents includes matching different preferences for all agents, voting according to the auditing result of each agent, and generating an auditing result; wherein the preference matching process of the agent is: in the preference library, the attention weight of each agent is randomly matched.

[0013] As a preferred scheme of the AI-driven video content automatic auditing method, wherein: the video information includes video frame sequence data and audio stream data in the video to be audited;

[0014] After aligning the video frame sequence data and the audio stream data according to the timestamp, the insertion of the split point is performed on the time axis of the video to be audited, and the video information is divided into multiple levels according to the inserted split point:

[0015] Step 1: according to the out point of each program in the video content, the insertion of the first-level split point is performed;

[0016] Step 2: in the video frame sequence data, the time of the shot cut is identified, and the insertion of the second-level split point is performed;

[0017] Step 3: using the AI-driven large language model, each sentence in the audio stream data is identified, and according to the identification result of the sentence semantics, the continuous sentences in the semantic series are identified; the second-level split point is inserted at the time after the continuous sentences; the third-level split point is inserted at the time after each sentence;

[0018] Step 4: according to the first-level split point, the time axis is divided to obtain the first-level slice of the video;

[0019] The second-level split points obtained in steps 2 and 3 are divided on the time axis of the first-level slice to obtain the second-level slice under each first-level slice; according to the third-level split point, the time axis of the second-level slice is divided to obtain the third-level slice under each second-level slice.

[0020] As a preferred scheme of the AI-driven video content automatic auditing method, the node simulation includes taking each slice of the video information on the time axis as a node; meanwhile, the hierarchical relationship of each node is synchronized according to the hierarchical relationship between the slices, and three levels of nodes are obtained. ; wherein, represents a first-level node corresponding to the i-th first-level slice; represents a second-level node corresponding to the j-th second-level slice under the i-th first-level slice; represents a third-level node corresponding to the third-level slice contained in the j-th second-level slice under the i-th first-level slice.

[0021] As a preferred scheme of the AI-driven video content automatic auditing method, the attribute features include constructing the association branch by using a knowledge graph to analyze the node features in the video information environment; and the habitual features of the current node are analyzed by using the features of similar nodes in the historical records.

[0022] The audio part in each node is encoded by using the AI-driven large language model to obtain a feature vector of the audio part; the video part in each node is encoded by using a time sequence Transformer network to obtain a feature vector of the video part; and the feature of each node is represented as: ; wherein, represents the feature vector of the audio part of the node, represents the feature vector of the video part of the node; the edges between the nodes are generated by analyzing the correlation of the feature vectors between the nodes, and the knowledge graph is constructed.

[0023] The correlation analysis includes generating edges between every two nodes by using principal component analysis strategies and time sequence correlation analysis, respectively; wherein, the edge relationship includes two similarity correlation sequences and time sequence correlation of four feature vector pairing results.

[0024] The principal component analysis strategy includes measuring the similarity of two feature vectors of the audio part and the video part in the node to obtain a similarity correlation sequence of the audio part and a similarity correlation sequence of the video part; the measurement process of each feature vector includes: step one, calculating the similarity of the feature vectors in every two nodes to obtain the initial similarity of every two nodes, denoted as represents the initial similarity of node w and node e on the r-th feature vector, 0 represents the initial index of the similarity sequence, and r includes two feature vectors of the audio part and the video part in the node.

[0025] Step two: according to the same feature vector on different nodes, a feature vector matrix is constructed; by using a preset cumulative explained variance threshold, the feature vector matrix is dimensionally reduced by principal component analysis to obtain the feature vector of each node after dimension reduction; by using the feature vector of each node after dimension reduction, the similarity of each two nodes is calculated to obtain the similarity of each two nodes after the first dimension reduction: ;

[0026] Step three: continuously repeat step two until the change amplitude of the similarity value after two consecutive dimension reductions is less than the preset minimum fluctuation threshold; a sequence is obtained: ;

[0027] Wherein, n represents the number of dimension reduction operations, and the repeated part of the similarity is removed by n-2;

[0028] The analysis of the time sequence correlation includes inputting the feature representation F of each node into a pre-trained time sequence modeling network, and respectively outputting the correlation probability of the four feature vector pairing results between two nodes: ; wherein, Represents the correlation probability of the pairing between the feature vectors of the audio part of node w and node e; Represents the correlation probability of the pairing between the feature vector of the audio part of node w and the feature vector of the video part of node e; Represents the correlation probability of the pairing between the feature vector of the audio part of node e and the feature vector of the video part of node w; Represents the correlation probability of the pairing between the feature vectors of the video part of node w and node e.

[0029] As a preferred scheme of the AI-driven video content automatic auditing method, wherein: the prior probability of and in the feature representation F of each node is calculated respectively to generate and The probability distribution of each violation as the usual feature of each node.

[0030] As a preferred scheme of the AI-driven video content automatic auditing method, wherein: the auditing process of each agent includes two rounds of auditing stages: in the first stage, the agent traverses each node to obtain and If the probability of the existence of each violation is greater than a preset value, the video content at the node is marked and a warning is issued; otherwise, the second stage is entered: the video content of each node in the knowledge graph is audited by using a pre-trained neural network to generate the probability of violation of each node.

[0031] The voting according to the audit result of each intelligent agent includes that if any intelligent agent identifies that there is a node violation, the identified node of the violation is voted +1;If the number of votes at the node is greater than 1, it is judged that the unstable part in the video content;If the number of votes at the node is 0, it is judged that the video content passes the audit;

[0032] The preference library is the preference of each edge relationship The preset weight value range, and the weight sum of the four elements matched each time is 1;

[0033] The intelligent agent finds the maximum value and the minimum value in each sequence When auditing each two nodes, and generates a guide vector of node features according to the maximum value and the minimum value;The two guide vectors obtained by the audio and video parts are spliced;Two splicing vectors are obtained by different splicing methods, and And are used as weights;

[0034] The pre-trained neural network is used to determine the probability distribution of each guide vector and each violation condition, and the probability distribution of each splicing vector and each violation condition, and the weight preference of the intelligent agent is used for weighted summation to obtain the final probability distribution result.

[0035] As a preferred scheme of the AI-driven video content automatic audit method, wherein: the analysis of the violation space includes, in the case of determining the node content violation, according to the voting result, the proportion of different violations is counted;The distribution of the violation content in different violation conditions is generated, and the corresponding violation label is inserted into the node of the violation, and a warning is issued;

[0036] When the video content audit report is generated, the node range of the violation label is narrowed down, and only the lowest node indicated by each violation label is retained.

[0037] In the second aspect, the application provides an AI-driven video content automatic audit system, which comprises a collection unit, which acquires video information to be audited and performs multi-level segmentation on the video information;

[0038] The simulation unit simulates nodes in the slice according to the content independence of the video information;The attribute characteristics of the nodes are generated by analyzing the associated branches and historical characteristics of each node;

[0039] The analysis unit performs content audit through the audit strategy of multiple virtual intelligent agents, and analyzes the violation space according to the unstable part in the audit result;

[0040] The output unit generates a final video content review report according to the review result and the analysis result of the violation space.

[0041] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program, when executed by the processor, implements any step of the AI-driven video content automatic review method according to the first aspect of the present application.

[0042] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the AI-driven video content automatic review method according to the first aspect of the present application.

[0043] The present application has the following beneficial effects: through multi-level segmentation and structured node modeling, the complex relationship of video content in time and semantics can be comprehensively captured, and fine-grained management of video segments can be realized. Through multi-modal feature fusion and knowledge graph analysis, the collaborative recognition ability of multi-source information such as audio and video is significantly improved. Combined with principal component analysis and time sequence network, the discrimination accuracy of content change trend, abnormal behavior and cross-segment association is effectively enhanced. The multi-virtual intelligent agent collaborative review and voting mechanism not only improves the robustness of the review result, but also can adaptively cope with diversified violation behaviors. Through fault tolerance and violation space analysis, the false negative rate and the false positive rate are greatly reduced, efficient, accurate and intelligent automatic review of complex video content is realized, and the content safety management and platform compliance operation ability are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0045] Figure 1 The flowchart of the AI-driven video content automatic review method. DETAILED DESCRIPTION

[0046] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0047] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present application.

[0048] It should also be noted that, as used in the specification and in the claims, the article "a", "an", or "the" is intended to mean that there are one or more of the features or elements. As used in this specification and the claims, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or". That is, unless specified otherwise, or clear from the context, the designation "X employs A or B" means that X employs A or B or both A and B. In addition, the articles "a", "an", and "the" are intended to mean that there are one or more (for example, one) of the features or elements, unless otherwise indicated or unless it would be clear from the context.

[0049] Reference will now be made to the drawings, in which Figure 1 For one embodiment of the present application, the embodiment provides an AI-driven video content automatic review method, comprising the following steps:

[0050] S1: Obtain the video information to be reviewed, and perform multi-level segmentation on the video information.

[0051] Further, the video information includes video frame sequence data and audio stream data in the video to be reviewed. After aligning the video frame sequence data and the audio stream data according to the timestamp, the insertion of the segmentation point is performed on the time axis of the video to be reviewed, and the multi-level segmentation of the video information is realized according to the inserted segmentation point.

[0052] Step one: According to the out point of each program in the video content, the first-level segmentation point is inserted. The "out point of each program" generally refers to the start and end boundary point of each independent program unit in the video content stream, which can be the start point, end point or content switching boundary point of the program. In the present application, it is measured as the start point of the program.

[0053] Step two: In the video frame sequence data, the time of the shot cut (the time between two frames) is identified, and the second-level segmentation point is inserted.

[0054] Step three: Using an AI-driven large language model, each sentence in the audio stream data is identified, and according to the identification result of the sentence semantics, the continuous sentences in the semantic series are identified; the second-level segmentation point is inserted at the time after the continuous sentences; and the third-level segmentation point is inserted at the time after each sentence.

[0055] In fact, each sentence is encoded into a feature vector by an AI-driven large language model. Each feature vector represents a sentence. Further, the relationship between the feature vectors is used to identify the semantic concatenation relationship. In this embodiment, the AI-driven large language model is preferably a large pre-trained natural language processing model based on the Transformer architecture, which is used for sentence-level understanding and feature encoding of the transcribed text from the audio stream. Optional models include but are not limited to BERT, RoBERTa, ALBERT, ERNIE, XLNet, T5, etc. general language models, or models optimized for dialogue and speech-to-text fields, such as ChatGPT, Llama, Baidu-ERNIE-Bot, etc. By inputting each sentence into the above model, a sentence feature vector representing its semantic information is obtained. In a preferred embodiment, a long short-term memory (LSTM) layer is introduced into the neural network structure to model the sequence of sentence-level feature vectors. The LSTM layer can effectively capture the context dependence and semantic cohesion information between sentences, thereby automatically learning and distinguishing the semantic concatenation relationship between consecutive sentences. Specifically, the sentence feature vectors arranged in chronological order are input into the LSTM layer one by one, and the relevance of the current sentence and the context sentence is analyzed using the hidden state output.

[0056] Step four: according to the primary segmentation points, the time axis is divided to obtain the primary slices of the video. The secondary segmentation points obtained in steps two and three are divided on the time axis of the primary slices to obtain secondary slices under each primary slice; and the tertiary segmentation points are divided on the time axis of the secondary slices to obtain tertiary slices under each secondary slice.

[0057] First, the primary segmentation points are set by the "out point" of the program unit, which can effectively divide the video stream into logically independent program segments, improving the accuracy of content attribution and structure recognition. Second, the secondary segmentation points are identified by shot change points, which further segment the visual scene changes within the program, making it easier to capture important picture changes, abnormal scenes, and potential risks. Further, the large language model is used to identify sentences and semantic levels of audio content, and the tertiary and auxiliary segmentation points are set based on the semantic concatenation relationship, achieving accurate positioning and context tracing of semantic information under audio-visual synchronization. The multi-level segmentation mechanism organically combines the temporal structure, semantic structure, and visual structure of the video content, providing a solid foundation for subsequent node modeling, feature extraction, intelligent review, and violation detection, effectively improving the granularity and intelligence level of content analysis.

[0058] S2: In the slice, nodes are simulated according to the content independence of the video information; and the attribute features of the nodes are generated by analyzing the associated branches and historical features of each node.

[0059] It is to be understood that the attribute features include constructing the association branch using a knowledge graph, analyzing the node features in the video information environment; by analyzing the features of similar nodes in the history record, the habitual features of the current node are analyzed. By modeling the structured correlation of the multi-modal features between nodes, the deep semantic connection of the node-level video content in the audio and visual dimensions is realized, thereby providing high robustness and high resolution knowledge support for intelligent review and anomaly detection.

[0060] Using the AI-driven large language model, the audio part in each node is encoded to obtain the feature vector of the audio part (using the vector generated by the large language model in S1, saving additional operations); using the time series Transformer network, the video part in each node is encoded to obtain the feature vector of the video part; the feature representation of each node is: ; wherein, represents the feature vector of the node audio part, represents the feature vector of the node video part; by analyzing the correlation between the feature representations of the nodes, the edges between the nodes are generated to construct the knowledge graph.

[0061] Specifically, the correlation analysis includes generating edges between each two nodes by principal component analysis strategy and time series correlation analysis, respectively; wherein, the edge relationship includes: two similarity correlation sequences and four time series correlations of feature vector pairing results.

[0062] Further, the principal component analysis strategy includes: for the two feature vectors of the audio part and the video part in the node, the similarity is measured respectively to obtain the similarity correlation sequence of the audio part and the similarity correlation sequence of the video part; the measurement process for each feature vector includes (that is, the analysis of the audio or video feature vector in the following is carried out): Step one, calculate the similarity of the feature vectors in each two nodes to obtain the initial similarity of each two nodes, denoted as represents the initial similarity of node w and node e in the rth feature vector, 0 represents the initial index of the similarity sequence, and r includes the two feature vectors of the audio part and the video part in the node.

[0063] Step two: construct a feature vector matrix according to the same feature vector on different nodes; use the preset cumulative explained variance threshold to reduce the dimension of the feature vector matrix by principal component analysis to obtain the feature vector of each node after dimension reduction; use the feature vector of each node after dimension reduction to calculate the similarity of each two nodes to obtain the similarity of each two nodes after the first dimension reduction: .

[0064] Step three: continuously repeat step two until the change amplitude of the similarity value after dimensionality reduction for two consecutive times is less than the preset minimum fluctuation threshold; obtain the sequence: .

[0065] Wherein, n represents the number of dimensionality reduction operations, and n-2 is used to remove the repeated part of the similarity of the sequence, thereby reducing the calculation amount.

[0066] It is to be known that the analysis of the time sequence correlation includes inputting the feature representation F of each node into a pre-trained time sequence modeling network (in this embodiment, preferably a long short-term memory neural network (LSTM). In other optional embodiments, it can also be a gated recurrent unit (GRU), a time sequence Transformer, etc. Model, globally model the node sequence, and learn the time sequence dependency relationship and context dynamic association between nodes): ; wherein, represents the correlation probability of the feature vectors of the audio part of the node w and the node e after pairing; represents the correlation probability of the feature vectors of the audio part of the node w and the node e after pairing; represents the correlation probability of the feature vectors of the audio part of the node w and the node e after pairing; represents the correlation probability of the feature vectors of the audio part of the node w and the node e after pairing.

[0067] Specifically, the AI-driven large language model and the time sequence Transformer are used to encode the audio part and the video part of the node respectively, obtain a unified feature representation, and realize efficient fusion of multi-modal information. The principal component analysis strategy can remove the redundancy and noise in the high-dimensional features, and only retain the representative main information, thereby improving the effectiveness and interpretability of the feature similarity measurement between different nodes. At the same time, the feature space is dynamically optimized through the adaptive dimensionality reduction process. Further combined with the time sequence modeling network (such as LSTM / GRU / time sequence Transformer), the complex dependency relationship and dynamic coupling features between nodes are captured in the global time sequence context, and the correlation probability in multiple pairing situations is output, thereby significantly enhancing the expression ability of the knowledge graph edge weight. Through the above multi-dimensional and fine-grained structured correlation analysis, the deep logical relationship and abnormal link between video content segments can be accurately revealed, thereby providing a solid decision basis for subsequent agent collaborative review, violation space identification, and content security control.

[0068] For each node, the feature representation F includes and , respectively, to calculate the prior probability, respectively, to generate and For each probability distribution of the violation, as the usual feature of each node.

[0069] By modeling the prior probability of the audio feature vector and the video feature vector of each node respectively, the corresponding probability distribution is generated for multiple violation cases, thereby providing efficient, interpretable and traceable risk discrimination basis for subsequent content review process. Using large-scale historical data or expert knowledge, prior probability distribution models are established in advance for different violation types (such as pornography, violence, sensitive topics, copyright risks, etc.). The node features are matched with known violation labels to realize the rapid quantitative evaluation of the usual risk properties of the node content. This mechanism not only improves the early warning capability of the review link to potential violation content, but also reduces the dependence on real-time large model reasoning and reduces the system computing pressure. Through the supplement of the usual feature probability distribution, the discrimination robustness and fault tolerance of the node are further improved, so that the entire review system can make joint decisions combining prior knowledge and real-time features when facing new or complex violation behaviors, enhancing the intelligent and refined level of content security management.

[0070] S3: Through the multi-virtual agent review strategy, the content is reviewed, and the unstable part in the review result is analyzed to analyze the violation space.

[0071] It should be noted that the multi-virtual agent review strategy includes matching different preferences for all agents, voting according to the review results of each agent to generate a review result; wherein the preference matching process of the agent is: in the preference library, the attention weight of each agent is randomly matched.

[0072] Further, the review process of each agent includes two rounds of review stages: in the first stage, the agent traverses each node to obtain and For each probability distribution of the violation, if the probability of the violation exists is greater than a preset value, the video content at the node is marked and a warning is issued (here is a review result); otherwise, the second stage is entered: using a pre-trained neural network, the video content of each node in the knowledge graph is reviewed to generate the probability of violation of each node.

[0073] The voting according to the review results of each agent includes: if any agent identifies that there is a node violation (generally, the violation probability is greater than a certain threshold), the identified node of the violation is voted +1; if the number of votes at the node is greater than 1, the video content is judged as an unstable part; if the number of votes at the node is 0, the video content is judged as passing the review.

[0074] The preference library is for each edge relationship The preset weight value range, and the weight sum of the four elements matched each time is 1.

[0075] The agent finds the maximum value and the minimum value in each sequence of each two nodes, and generates a guide vector of the node features according to the maximum value and the minimum value. Specifically, the vector difference corresponding to the maximum value is taken as a translation amount, and the start and end positions of the vector difference corresponding to the minimum value are changed (actually, the vector difference corresponding to the minimum value is subtracted from the vector difference corresponding to the maximum value), so as to obtain a guide vector representing the change of the features. Through fine analysis of the similarity extreme value of each two nodes in the feature sequence, the directionality modeling of the risk evolution trend between the nodes is realized. Generally, the maximum similarity in the sequence often corresponds to a low-risk section with highly similar features and close semantic connection between the nodes, at this time, the vector difference is small and cannot effectively reflect the abnormal change. In contrast, the minimum similarity often reveals that there is a significant difference or mutation in the multi-modal attributes or semantic level between the contents of the nodes, which is exactly the "demarcation zone" where high-risk behaviors, abnormal evolution, and potential risks occur frequently. Specifically, the two node feature vectors corresponding to the minimum similarity are selected, and the change direction of the "semantically farthest level" is obtained through the vector difference. This change direction is the dominant path of the dramatic differentiation or abnormal jump of the content attributes between the nodes, and is easy to carry the features of violation, sensitivity, and deviation. Compared with single extreme value or mean value statistics, the guide vector takes the "globally farthest point" as the starting point of risk analysis, which can effectively amplify the performance of potential illegal behaviors in the feature space, and significantly improve the recognition ability of the system to complex and highly concealed contents. The vector difference corresponding to the maximum similarity serves as a translation reference, which can ensure that the extracted guide vector focuses on the actual content change between the nodes, rather than local noise or irrelevant disturbance. Finally, combined with the mapping of the guide vector and the probability distribution of various illegal behaviors by the neural network, the system can intelligently determine whether such change is directed to illegal risk, realize the directional guidance and high-confidence warning of the content evolution process.

[0076] The vector difference of the nodes corresponding to the minimum similarity is modeled as the semantic farthest differentiation direction between the nodes, which is used as a guide vector for discriminating the evolution trend of the content illegal behavior. The vector can accurately reflect the abnormal jump of the attribute and the risk mutation between the nodes, and effectively improve the intelligent recognition and early warning ability of the content review system to complex illegal behaviors

[0077] The two guide vectors (denoted as H and T) obtained from the audio and video parts are spliced; two spliced vectors [T, H] and [H, T] are obtained through different splicing sequences, and and are used as weights.

[0078] With the pre-trained neural network, the probability distribution of each guidance vector and each violation case, and the probability distribution of each splicing vector and each violation case are determined, and a weighted sum is performed with the weight preferences of the agents to obtain a final probability distribution result.

[0079] It is to be noted that, by matching the diversified preferences and attention weights, the differences in the attention points of the multi-agent in the real world for judging the risk of video content are simulated, and the adaptability and robustness of the system for complex and variable violation behaviors are effectively improved. The two-stage review mechanism ensures the hierarchical discrimination of explicit and implicit violation risks, which can not only efficiently capture high-risk nodes, but also conduct in-depth and context-enhanced detailed analysis on difficult samples. By dynamically assigning different attention weights to each agent in the preference library, the agents form complementary judgments in the multi-dimensional feature space. The design of the guidance vector (based on the extreme value in the feature sequence and its variation difference) can effectively capture the representative direction and abnormal features of the attribute changes between nodes, and enhance the sensitivity to the trend and abnormal behavior of the video content changes. The splicing and weighting mechanism further integrates multi-modal features and their variation relationships, making the probability judgment more holistic and targeted. Finally, the integration of voting and probability weighted judgment effectively improves the accuracy, robustness and traceability of the review results, significantly reduces the risk of missed and mistaken judgments, and meets the high-standard demand for intelligent and refined video content safety review in practical application scenarios.

[0080] Further, in the case of determining the content violation of the node, the proportion of different violations is counted according to the voting results; the distribution of the violation content in different violation cases is generated, and the corresponding violation label is inserted into the node, and a warning is issued. When the node content is found to be in violation, the type of violation can be statistically analyzed and labeled, realizing the hierarchical management and intelligent tracing of multiple violation risks. By counting the voting distribution in the review results of all agents, the system can not only determine whether the node is in violation, but also quantify the proportion of different violation types in the same node, thereby depicting the diversity and complexity of content risks. This mechanism can provide a structured data basis for subsequent content governance, hierarchical disposal, and automatic warning applications. Further, by automatically inserting specific violation labels into the violation nodes, the system can realize accurate tracing and positioning, facilitate the explanation and review of the review decision, and provide a basis for platform compliance operation and user protection. At the same time, the real-time warning mechanism helps the platform respond to potential risks in the first time, reduces the spread and impact of violation content, and improves the overall content safety and management intelligence level.

[0081] S4: generating a final video content review report according to the review results and the analysis results of the violation space.

[0082] In the generation of the video content review report, the node range of the violation label is narrowed down, and only the most specific node referred by each violation label is retained.

[0083] In the generation of the video content review report, the system collects all inserted violation labels and their distribution on all nodes (including nodes at all levels in the multi-level structure). For each type of violation label, the top-level node (such as the first slice) is checked in turn to check the distribution of the same type of violation label on all its subordinate nodes (such as the second and third slices).

[0084] If the violation label on a first-level node has no "same label" distributed in all its direct child nodes (the next level), the label is retained on the current node.

[0085] If the next level has a violation label consistent with the current label, the label on the current node is "erased", and only the label on the next level node is retained.

[0086] Recursion in turn until the bottom layer (the most specific node).

[0087] Finally, only the most specific node that does not have a violation label of the same type appears on the most specific node, the violation label is retained, and the minimum coverage and unique marking of the violation are achieved. However, it should be noted that in the process of narrowing down the node range, the "erasing" operation of the violation label adopts a local recursive and on-demand erasing mechanism. That is, the system only compares the node with all the lower nodes directly contained in it for each node with a violation label. If there is a violation label consistent with the upper node A (such as being determined to be "consistent" or "within the preset fluctuation range") in the direct lower nodes of A, only the label of node A is erased (i.e. the label is not displayed on the A layer, only the label on the lower node is retained); if there is no consistent label in the lower nodes of A, the label on A is retained. This erasing and non-erasing determination and operation is limited to A and its subordinates, and does not affect the label processing on other nodes (such as node B) in the same layer. In other words, if the lower nodes of node B in the same layer have no consistent violation label, the label on node B will be retained, and only the label on node A will be erased. This "node-by-node, branch-by-branch" recursive processing ensures that the label positioning is both fine and avoids unnecessary loss, and does not cause important information to be missed due to global erasing.

[0088] By recursive blanking of multi-level node labels and reservation of the lowest unique node, the system can effectively avoid the same violation behavior being repeatedly marked at the upper node and all child nodes, simplify the violation trace link, and improve the clarity and operability of the content review results. This mechanism helps the platform or regulatory personnel quickly lock the most representative and most refined violation segment when facing complex multi-level content, improves the processing efficiency and scientific nature of compliance decisions, and prevents misjudgment and multiple judgments. At the same time, it also provides high-quality structured review data support for downstream applications such as automated content distribution, risk visualization, and big data governance.

[0089] The embodiment also provides an AI-driven video content automatic review system, comprising:

[0090] A collection unit acquires video information to be reviewed and performs multi-level segmentation on the video information.

[0091] An analog unit performs node simulation according to the content independence of the video information in the slice, and analyzes the associated branches and historical characteristics of each node to generate attribute characteristics of the node.

[0092] An analysis unit performs content review through the review strategies of multiple virtual agents, and analyzes the violation space according to the unstable part in the review result.

[0093] An output unit generates a final video content review report according to the review result and the analysis result of the violation space.

[0094] The embodiment also provides a computer device suitable for the AI-driven video content automatic review method, comprising a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the AI-driven video content automatic review method proposed in the above embodiment.

[0095] The computer device can be a terminal, which includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, a carrier network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0096] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the method for automatically auditing video content based on AI driving as described in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0097] To sum up, the application realizes multi-dimensional and fine-grained correlation analysis and knowledge graph construction among nodes by multi-level video content segmentation and structured node modeling, deep coding of audio and video multi-modal features by combining large language models and time sequence Transformer networks, and principal component analysis and time sequence correlation modeling methods. Further, the identification ability and determination robustness of content violation behavior are improved through multi-virtual agent collaborative review, dynamic allocation of preference weight, guide vector and two-stage review judgment mechanism. Finally, based on violation space statistics and recursive blanking label mechanism, the system can accurately locate high-risk nodes in video content, realize automatic, intelligent and traceable content safety review, and effectively improve the intelligent level and governance efficiency of video content compliance management.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. An AI-driven based method for automatic video content review, the method comprising: The method comprises the following steps: acquiring video information to be audited, and performing multi-level segmentation on the video information; ​ In the slice, the content independence of the video information is simulated according to the nodes, and the attribute characteristics of each node are analyzed to generate the attribute characteristics of the nodes; Through the multi-virtual agent auditing strategy, the content is audited, and the violation space is analyzed according to the unstable part of the auditing result; According to the auditing result and the analysis result of the violation space, a final video content auditing report is generated; The multi-virtual agent auditing strategy comprises the following steps: matching different preferences for all agents, voting according to the auditing result of each agent, and generating an auditing result; wherein the preference matching process of the agent is as follows: in the preference library, the attention weight of each agent is randomly matched; The video information comprises video frame sequence data and audio stream data in the video to be audited; After aligning the video frame sequence data and the audio stream data according to the time stamp, the insertion of the segmentation point is performed on the time axis of the video to be audited, and the multi-level segmentation of the video information is realized according to the inserted segmentation point: Step 1: inserting a first-level segmentation point according to the starting point of each program in the video content; Step 2: identifying the moment of shot change in the video frame sequence data, and inserting a second-level segmentation point; Step 3: using an AI-driven large language model to identify each sentence in the audio stream data, and identifying continuous sentences according to the recognition result of the sentence semantics; inserting a second-level segmentation point at the moment after the continuous sentences; and inserting a third-level segmentation point at the moment after each sentence; Step 4: cutting the time axis according to the first-level segmentation point to obtain a first-level slice of the video; The second-level segmentation points obtained in steps 2 and 3 are cut on the time axis of the first-level slice to obtain second-level slices under each first-level slice; and the third-level segmentation points are cut on the time axis of the second-level slices to obtain third-level slices under each second-level slice. 2.The AI-driven video content automatic review method based on claim 1, characterized in that: The node simulation comprises: taking each slice of the video information on a time axis as a node; and synchronizing the hierarchical relationship of each node according to the hierarchical relationship between the slices to obtain three levels of nodes. ; wherein, represents a first-level node corresponding to the i th first-level slice; represents a second-level node corresponding to the j th second-level slice under the i th first-level slice; represents a third-level node corresponding to the third-level slice contained in the j th second-level slice under the i th first-level slice. 3.The AI-driven video content automatic review method of claim 2, wherein: The attribute characteristics comprise constructing the associated branch using a knowledge graph and analyzing the node characteristics in the video information environment; The habitual characteristics of the current node are analyzed through the characteristics of similar nodes in the historical record; The AI-driven large language model is used to encode the audio part in each node to obtain a feature vector of the audio part; The time sequence Transformer network encodes the video part in each node to obtain a feature vector of the video part; a feature representation of each node is represented as: ; wherein, represents a feature vector of the audio part of the node, represents a feature vector of the video part of the node; edges between nodes are generated by analyzing the correlation of the feature representations between nodes, and the knowledge graph is constructed. The correlation analysis comprises generating edges between each two nodes through principal component analysis strategy and time sequence correlation analysis respectively; wherein the edge relationship comprises two similarity correlation sequences and time sequence correlation of four feature vector pairing results; The principal component analysis strategy comprises: similarity measurement is performed on two kinds of feature vectors of the audio part and the video part in the nodes respectively to obtain a similarity correlation sequence of the audio part and a similarity correlation sequence of the video part; the measurement process for each kind of feature vector comprises: step one, calculating the similarity of the feature vectors in each two nodes to obtain the initial similarity of each two nodes, denoted as denotes the initial similarity of the node w and the node e on the rth feature vector, 0 denotes the initial index of the similarity sequence, and r comprises two kinds of feature vectors of the audio part and the video part in the nodes. Step two: according to the same feature vectors on different nodes, a feature vector matrix is constructed; by using a preset cumulative explained variance threshold, dimensionality reduction is performed on the feature vector matrix through principal component analysis to obtain the feature vectors of each node after dimensionality reduction; by using the feature vectors of each node after dimensionality reduction, the similarity of each two nodes is calculated to obtain the similarity of each two nodes after the first dimensionality reduction: ; Step three: continuously repeat step two until the variation amplitude of the similarity value after dimension reduction for two times in succession is less than the preset minimum fluctuation threshold value; and obtain the sequence: ; Wherein n represents the number of dimension reduction operations, and n-2 is used to remove the repeated similarity part of the sequence; The analysis of the time sequence correlation includes inputting the feature representation F of each node into a pre-trained time sequence modeling network, and respectively outputting the correlation probability of the four feature vector pairing results between two nodes: ; wherein, represents the correlation probability of the pairing between the feature vectors of the audio part of the node w and the node e; represents the correlation probability of the pairing between the feature vector of the audio part of the node w and the feature vector of the video part of the node e; represents the correlation probability of the pairing between the feature vector of the audio part of the node e and the feature vector of the video part of the node w; represents the correlation probability of the pairing between the feature vectors of the video part of the node w and the node e. 4.The AI-driven video content automatic review method of claim 3, wherein: the feature representation F of each node and respectively, to generate and the probability distribution for each violation case as the usual features of each node. 5.The AI-driven video content automatic review method based on claim 4, characterized in that: The auditing process of each of the intelligent agents includes two-round auditing stages: in a first stage, the intelligent agent traverses each node to obtain and For the probability distribution of each violation, if the probability of the existence of the violation is greater than a preset value, the video content at the node is marked, and a warning is issued. Conversely, the second stage is entered: using a pre-trained neural network to audit the video content of each node in the knowledge graph to generate the probability of violation of each node; The voting according to the audit result of each agent includes: if any agent identifies that there is a node violation, voting +1 for the identified node in violation; if the number of votes at the node is greater than 1, it is judged that the unstable part in the video content; if the number of votes at the node is 0, it is judged that the video content passes the audit; The preference library is for each edge relationship The preset weight value range, and the weight sum of the four elements matched each time is 1. The intelligent agent finds the maximum value and the minimum value in each sequence , and generates a guide vector of node features according to the maximum value and the minimum value; splices the two guide vectors obtained from the audio and video parts; obtains two spliced vectors through different splicing modes, and uses and as weights, respectively; Using the pre-trained neural network, the probability distribution of each guide vector and each violation condition, and the probability distribution of each splicing vector and each violation condition are determined, and the weight preferences of the agents are used for weighted summation to obtain the final probability distribution result. 6.The AI-driven video content automatic review method based on claim 5, characterized in that: The analysis of the violation space includes, in the case of determining that the node content is in violation, according to the voting result, the proportion of different violations is counted; The distribution of the violation content in different violation conditions is generated, and the corresponding violation label is inserted into the node of the violation, and a warning is issued; When the video content audit report is generated, the node range of the violation label is narrowed down, and only the lowest node indicated by each violation label is retained.

7. An AI-driven video content automatic review system based on any one of claims 1-6, characterized in that: It includes a collection unit that acquires video information to be audited and performs multi-level segmentation on the video information; The simulation unit simulates nodes in the slice according to the content independence of the video information, analyzes the associated branches and historical characteristics of each node, and generates the attribute characteristics of the nodes; The analysis unit performs content audit through the audit strategy of multiple virtual agents, and analyzes the violation space according to the unstable part in the audit result; The output unit generates a final video content audit report according to the audit result and the analysis result of the violation space.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The processor executes the computer program to realize the steps of the AI-driven video content automatic audit method according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the AI-driven video content automatic audit method according to any one of claims 1-6.

Images