Real-time detection system for illegal content of network live broadcast

Abstract

The invention discloses a network live broadcast illegal content real-time detection system, and relates to the field of network live broadcast, and the system comprises a data collection module which is used for collecting a to-be-detected data sample in a live broadcast room in real time, and carrying out the preprocessing of the to-be-detected data sample; the priori database is used for uploading the violation data samples and storing the violation data samples; the analysis module is used for receiving the preprocessed to-be-detected data sample acquired by the data acquisition module, and performing similarity analysis on the to-be-detected data sample and an illegal data sample in a prior database to obtain an illegal tendency of the to-be-detected data sample from the live broadcast room; the method is different from a single detection mode in the prior art, the timeliness, comprehensiveness and accuracy of violation detection are improved through multi-source data fusion and algorithm application, and technical support is provided for live broadcast supervision.

Description

Technical Field

[0001] The present invention relates to the technical field of network live broadcasting, and in particular to a real-time detection system for illegal content in network live broadcasting. Background Art

[0002] Live broadcast violation detection is crucial. With the help of image, text, and video detection technology, illegal content can be automatically identified to protect the health of the live broadcast environment.

[0003] The invention patent application with application number 201710485750.6 discloses a method for comprehensive status perception and real-time content supervision of a live broadcast platform, comprising the following steps: setting a dynamic traffic threshold for each live broadcast room based on the room's historical traffic data, obtaining the room's current traffic data in real time, and combining the rate of change of the current traffic data with the dynamic traffic value to determine the room's traffic suspicion value; extracting a library of illegal bullet comments based on the room's historical bullet comment data, and setting corresponding weights based on the frequency of occurrence of each illegal bullet comment; obtaining the room's current bullet comment data in real time, fuzzy matching it with the illegal bullet comment library, and determining the room's bullet comment suspicion value based on the matched illegal bullet comments and the corresponding weights; segmenting the live broadcast video scene, and performing scene mutation detection on the segmented live video scene, determining the room's scene mutation suspicion value based on the degree of scene mutation; comprehensively analyzing the traffic suspicion value, bullet comment suspicion value, and scene mutation suspicion value to determine a suspicious live broadcast room, and the administrator checking the suspicious live broadcast room to determine whether it violates the law; and updating the dynamic traffic threshold and illegal bullet comment library based on the violation judgment result. This application aims to address the problem of "huge data volume on live broadcast platforms and inefficient manual supervision."

[0004] However, the existing technology for managing violations in live broadcasts often involves a step-by-step review of various types of violations, and it is impossible to comprehensively and accurately determine that a live broadcast room that "skirts the line" with various types of violations is a violation.

[0005] Therefore, a real-time detection system for illegal content in live streaming is proposed. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the prior art, the present invention provides a real-time detection system for illegal content in online live broadcasts, which can effectively solve the problems of the prior art.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0008] The present invention discloses a real-time detection system for illegal content in live webcasts, comprising:

[0009] The data collection module is used to collect data samples to be tested in real time in the live broadcast room and pre-process the data samples to be tested; the priori database is used to upload illegal data samples and store the illegal data samples; the analysis module is used to receive the pre-processed data samples to be tested collected by the data collection module, and perform similarity analysis on the data samples to be tested and the illegal data samples in the priori database to obtain the violation tendency of the live broadcast room from which the data samples to be tested come; the correction and judgment module is used to receive the violation tendency of the live broadcast room obtained based on the similarity analysis in the analysis module, correct the violation tendency of the live broadcast room, and set a judgment threshold, and compare the judgment threshold with the corrected violation tendency of the live broadcast room to determine whether the live broadcast room is in violation; the management and control module is used to obtain the judgment result of whether the live broadcast room is in violation in the correction and judgment module, and control the live broadcast room to be blocked when the judgment result is yes, and refresh the system operation when the judgment result is no; the message module is used to continuously obtain the correction result of the violation tendency of the live broadcast room and the judgment result of whether the live broadcast room is in violation in the correction module, and generate daily management messages of the live broadcast room based on the obtained information.

[0010] Furthermore, the types of data samples to be detected collected by the data acquisition module include: live images, and live audio segments corresponding to the live images. The data acquisition module is provided with a control unit and a pre-processing unit at a lower level. The control unit is used to control the operating frequency of the data acquisition module, so that the data acquisition module runs in real time to collect data samples to be detected based on the operating frequency controlled by the control unit. The pre-processing unit is used to receive live image data in the data samples to be detected collected in real time, identify dynamic targets in the live image data, and capture the motion trajectory of the dynamic target in the live image data, synchronously represent the motion trajectory with a trajectory line, and after the motion trajectory of the dynamic target is captured, capture the live picture frame based on the specified interval time area in the live image data;

[0011] Among them, the pre-processing unit runs to capture no less than three live picture frames. After capturing the motion trajectory of the dynamic target and the live picture frames, the pre-processing unit distinguishes and stores the captured motion trajectory of the dynamic target, the live picture frames and the live audio segments in the data samples to be detected.

[0012] Furthermore, during the operation phase of the control unit, the instantaneous growth rate of visitors to the live broadcast room, the instantaneous growth rate of revenue, and the instantaneous refresh rate of the bullet screen are monitored in real time;

[0013]

[0014] Where: F viit is the instantaneous growth rate of visitors; m now is the number of online visitors in the live broadcast room at the current moment; m before Compared to m nowThe number of online visitors in the live broadcast room at the last moment; △t is the sampling interval;

[0015] Among them, the calculation logic of the instantaneous growth rate of revenue, the instantaneous refresh rate of barrage and the instantaneous growth rate of visitors are the same. When calculating the instantaneous growth rate of revenue, the current moment revenue of the live broadcast room and the revenue of the previous moment of the live broadcast room are used for calculation. When calculating the instantaneous refresh rate of barrage, the current moment barrage volume of the live broadcast room and the previous moment barrage volume of the live broadcast room are used for calculation, which are respectively denoted as F viit 、F dk 、F gi , set the range and initial frequency of the control unit to control the operating frequency of the data acquisition module, so that the control unit controls the operating frequency of the data acquisition module within the set range based on the initial frequency, in F viit 、F dk 、F gi When any of the above items shows an upward or downward trend, the initial frequency is controlled to increase or decrease based on the specified ratio, and is maintained after it is increased or decreased to the maximum or minimum value of the set range.

[0016] Furthermore, the illegal data samples stored in the a priori database are manually uploaded by the system end user, and the a priori database stores the illegal data samples in a differentiated manner based on the types of illegal data samples;

[0017] Among them, the types of illegal data samples include: historical dynamic target motion trajectories, live broadcast frames with illegal object annotation boxes, and text information extracted from historical live broadcast audio segments. The illegal data samples stored in the priori database are manually updated, deleted, and modified based on the system end users.

[0018] Furthermore, during the operation phase of the analysis module, the similarity analysis results are used to indicate the violation tendency of the live broadcast room from which the data sample to be detected originates;

[0019] The similarity analysis targets in the analysis module are: dynamic target motion trajectories and historical dynamic target motion trajectories, text information extracted based on live audio segments and text information extracted from historical live audio segments, and contour images extracted from live screen frames and live screen frames with illegal object annotation boxes;

[0020] When the similarity between the dynamic target's motion trajectory and the historical dynamic target's motion trajectory is used to represent the violation tendency:

[0021]

[0022] Where: S1 is the violation tendency based on the motion trajectory performance; α, (1-α) are weight coefficients; DTW (T a ,T b) is the time series alignment distance between the dynamic target motion trajectory and the historical dynamic target motion trajectory based on dynamic time warping; F(C a ,C b ) is the Fréchet distance between the two main curves after extracting the main curves from the dynamic target motion trajectory and the historical dynamic target motion trajectory using the local linear embedding algorithm;

[0023] Among them, α and (1-α) are both positive numbers, and α>(1-α). The larger S1 is, the more likely there is a violation in the live broadcast room where the dynamic target motion trajectory comes from. When calculating S1, each historical dynamic target motion trajectory is used as a similarity analysis target and a similarity analysis is performed with the dynamic target motion trajectory.

[0024]

[0025] Where: d(C a (s),C b (t)) is the main curve C a The point with parameter s on the principal curve C b Euclidean distance between points with upper parameter t.

[0026] Furthermore, when the similarity between text information extracted from live audio segments and text information extracted from historical live audio segments is used to indicate the tendency of violations:

[0027]

[0028] Where: S2 is the violation tendency expressed based on text information; σ is the global scaling factor; λ and (1-λ) are weight coefficients; SemSim is the semantic similarity of the two text messages; |C| is the size of the common word set after removing stop words; |T1| and |T2| are the number of words in the text information T1 extracted from the live audio segment and the text information T2 extracted from the historical live audio segment after removing stop words; StrustSim is the structural consistency of the two text messages;

[0029] Among them, λ and (1-λ) are both positive numbers, and λ>(1-λ). When calculating S2, the text information extracted from each historical live audio segment is used as the similarity analysis target, and similarity analysis is performed with the text information extracted based on the live audio segment.

[0030] Furthermore, when the violation tendency is represented by the similarity between the contour image extracted from the live broadcast frame and the live broadcast frame with the illegal object labeling box:

[0031] Calculate the similarity between each contour in the contour image and the contour of the object in each illegal object annotation box;

[0032]

[0033] Where: s3 is the violation tendency based on the live frame performance; m(sim>90%) is the number of target contours in the contour image whose contours are more than 90% similar to the contours of the objects in the violation object annotation box; m all is the total amount of contours in the contour image; ω and ε are weights; D Polar is the polar coordinate distribution distance of the two contours; D Fourier is the Fourier descriptor distance between two contours;

[0034] The value ranges of ω and ε are 0.1-3 and 0.5-1 respectively, and their default values ​​are 1 and 0.8.

[0035] Furthermore, the correction and determination module is provided with a configuration unit at a lower level, and the configuration unit is used to configure an electronic label for the correction result of the live broadcast violation tendency. The correction logic for the live broadcast violation tendency in the correction and determination module is expressed as follows:

[0036] S=S1×k1+S2×k2+S3×k3;

[0037] Where: S is the corrected live broadcast violation tendency; k1, k2, and k3 are the reference proportions of violation tendency; S1, S2, and S3 are the violation tendency based on motion trajectory, text information, and live broadcast frame performance, respectively.

[0038] Among them, the electronic tag content configured for the live broadcast violation tendency correction result is the correction time of the correction result.

[0039] Furthermore, the values ​​of the violation tendency reference ratios k1, k2, and k3 are subject to:

[0040] k1, k2, and k3 are all positive numbers, and their sum is 1. When the live broadcast content in the live broadcast room is mainly based on the live broadcast user's image, k1>k2, k3>k2; when the live broadcast content in the live broadcast room is mainly based on the projection screen content, k1<k2, k3<k2;

[0041] The contents of the daily management messages of the live broadcast room generated by the message module are sorted based on the corresponding time of their respective electronic tags.

[0042] Furthermore, the data acquisition module is interactively connected to a control unit and a preprocessing unit at a lower level through a wireless network, the data acquisition module is interactively connected to a priori database and an analysis module through a wireless network, the analysis module is interactively connected to the data acquisition module through a wireless network, the analysis module is interactively connected to a correction and determination module through a wireless network, the correction and determination module is interactively connected to a configuration unit at a lower level through a wireless network, and the correction and determination module is interactively connected to a management and control module and a message module through a wireless network.

[0043] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0044] The present invention provides a real-time detection system for illegal content in online live broadcasts. During operation, the system collects data such as live broadcast images and audio segments, and performs multi-dimensional similarity analysis of motion trajectories, text information, and contour images with historical violation data. The system dynamically adjusts the detection frequency in combination with the instantaneous growth rate of visitor volume, revenue, and barrage volume. It adopts dynamic time warping, local linear embedding, Fourier descriptor and other algorithms to quantify the violation tendency. According to the main type of live broadcast content, that is, user image or projection content, it intelligently allocates weights to correct the violation tendency, sets a judgment threshold for comparison and determines the violation status, and uses the detection results to control the suspension of the live broadcast room or system refresh, and generates management messages sorted by time. This mechanism breaks through the traditional single detection mode, improves the timeliness, comprehensiveness and accuracy of violation detection through multi-source data fusion and algorithm application, and provides technical support for live broadcast supervision. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0046] Figure 1 This is a structural diagram of a real-time detection system for illegal content in live streaming. DETAILED DESCRIPTION

[0047] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] The present invention will be further described below with reference to the embodiments.

[0049] Example:

[0050] A real-time detection system for illegal content in live webcasts of this embodiment is as follows: Figure 1 As shown, including:

[0051] The data acquisition module is used to collect the data samples to be tested in real time in the live broadcast room and pre-process the data samples to be tested;

[0052] The types of data samples to be detected collected by the data acquisition module include: live images and live audio segments corresponding to the live images. The data acquisition module is provided with a control unit and a pre-processing unit at the lower level. The control unit is used to control the operating frequency of the data acquisition module so that the data acquisition module can collect the data samples to be detected in real time based on the operating frequency controlled by the control unit. The pre-processing unit is used to receive the live image data in the data samples to be detected collected in real time, identify dynamic targets in the live image data, and capture the motion trajectory of the dynamic targets in the live image data, and synchronously represent the motion trajectory with a trajectory line. After the motion trajectory of the dynamic target is captured, the live picture frame is captured based on a specified interval time area in the live image data.

[0053] The pre-processing unit captures no less than three live picture frames, and after capturing the motion trajectory of the dynamic target and the live picture frames, the pre-processing unit distinguishes and stores the captured motion trajectory of the dynamic target, the live picture frames, and the live audio segment in the data sample to be detected;

[0054] During the control unit operation phase, real-time monitoring is performed on the instantaneous growth rate of visitors to the live broadcast room, the instantaneous growth rate of revenue, and the instantaneous refresh rate of bullet comments;

[0055]

[0056] Where: F viit is the instantaneous growth rate of visitors; m now is the number of online visitors in the live broadcast room at the current moment; m before Compared to m now The number of online visitors in the live broadcast room at the last moment; △t is the sampling interval;

[0057] Among them, the calculation logic of the instantaneous growth rate of revenue, the instantaneous refresh rate of barrage and the instantaneous growth rate of visitors are the same. When calculating the instantaneous growth rate of revenue, the current moment revenue of the live broadcast room and the revenue of the previous moment of the live broadcast room are used for calculation. When calculating the instantaneous refresh rate of barrage, the current moment barrage volume of the live broadcast room and the previous moment barrage volume of the live broadcast room are used for calculation, which are respectively denoted as F viit 、F dk 、F gi , set the range and initial frequency of the control unit to control the operating frequency of the data acquisition module, so that the control unit controls the operating frequency of the data acquisition module within the set range based on the initial frequency, in F viit 、F dk 、F gi When any of the above items shows an upward or downward trend, the initial frequency is controlled to be increased or decreased based on the specified ratio, and maintained after being increased or decreased to the maximum or minimum value of the set range;

[0058] Through the above logic, the frequency of data acquisition module operation is regulated, so that the data samples to be tested collected by the data acquisition module are more representative, thereby improving the accuracy of the system operation output results;

[0059] A priori database, used to upload and store illegal data samples;

[0060] The illegal data samples stored in the priori database are manually uploaded by the system end user. When the priori database stores the illegal data samples, it differentiates and stores them based on the type of illegal data samples;

[0061] Among them, the types of illegal data samples include: historical dynamic target motion trajectories, live broadcast frames with illegal object annotation boxes, and text information extracted from historical live broadcast audio segments. The illegal data samples stored in the prior database are based on manual updates, deletions, and modifications by system users;

[0062] An analysis module is used to receive the pre-processed data samples to be tested collected by the data collection module, and perform similarity analysis between the data samples to be tested and the illegal data samples in the prior database to obtain the illegal tendency of the live broadcast room from which the data samples to be tested originated;

[0063] During the analysis module operation phase, similarity analysis results are used to indicate the violation tendency of the live broadcast room where the data sample to be detected originates;

[0064] The similarity analysis targets in the analysis module are: dynamic target motion trajectories and historical dynamic target motion trajectories, text information extracted from live audio segments and historical text information extracted from live audio segments, and contour images extracted from live screen frames and live screen frames with illegal object annotation boxes;

[0065] When the similarity between the dynamic target's motion trajectory and the historical dynamic target's motion trajectory is used to represent the violation tendency:

[0066]

[0067] Where: S1 is the violation tendency based on the motion trajectory performance; α, (1-α) are weight coefficients; DTW (T a ,T b ) is the time series alignment distance between the dynamic target motion trajectory and the historical dynamic target motion trajectory based on dynamic time warping; F(C a ,C b ) is the Fréchet distance between the two main curves after extracting the main curves from the dynamic target motion trajectory and the historical dynamic target motion trajectory using the local linear embedding algorithm;

[0068] Among them, α and (1-α) are both positive numbers, and α>(1-α). The larger S1 is, the more likely there is a violation in the live broadcast room where the dynamic target motion trajectory comes from. When calculating S1, each historical dynamic target motion trajectory is used as a similarity analysis target and a similarity analysis is performed with the dynamic target motion trajectory.

[0069]

[0070] Where: d(C a (s),C b (t)) is the main curve C a The point with parameter s on the principal curve C b The Euclidean distance between points with upper parameter t;

[0071] When the similarity between text information extracted from live audio segments and text information extracted from historical live audio segments is used to indicate violation trends:

[0072]

[0073] Where: S2 is the violation tendency expressed based on text information; σ is the global scaling factor; λ and (1-λ) are weight coefficients; SemSim is the semantic similarity of the two text messages; |C| is the size of the common word set after removing stop words; |T1| and |T2| are the number of words in the text information T1 extracted from the live audio segment and the text information T2 extracted from the historical live audio segment after removing stop words; StrustSim is the structural consistency of the two text messages;

[0074] Wherein, λ and (1-λ) are both positive numbers, and λ>(1-λ). When calculating S2, the text information extracted from each historical live audio segment is used as the similarity analysis target, and similarity analysis is performed with the text information extracted based on the live audio segment;

[0075] When the violation tendency is represented by the similarity between the contour image extracted from the live video frame and the live video frame with the illegal object labeling box:

[0076] Calculate the similarity between each contour in the contour image and the contour of the object in each illegal object annotation box;

[0077]

[0078] Where: s3 is the violation tendency based on the live frame performance; m(sim>90%) is the number of target contours in the contour image whose contours are more than 90% similar to the contours of the objects in the violation object annotation box; m all is the total amount of contours in the contour image; ω and ε are weights; D Polar is the polar coordinate distribution distance of the two contours; D Fourieris the Fourier descriptor distance between two contours;

[0079] Among them, the value ranges of ω and ε are 0.1~3 and 0.5~1 respectively, and the default values ​​of the two are 1 and 0.8;

[0080] Through the above logical formula, the violation tendency of each type of data sample to be detected is quantitatively calculated. The quantitative calculation results provide data support for the final live broadcast violation tendency, ensuring the stable output of the corrected live broadcast violation tendency.

[0081] The correction and determination module is used to receive the violation tendency of the live broadcast room obtained based on the similarity analysis in the analysis module, correct the violation tendency of the live broadcast room, set a determination threshold, and compare the determination threshold with the corrected violation tendency of the live broadcast room to determine whether the live broadcast room violates the law;

[0082] The correction and determination module is provided with a configuration unit at the lower level. The configuration unit is used to configure an electronic label for the correction result of the live broadcast violation tendency. The correction logic for the live broadcast violation tendency in the correction and determination module is expressed as follows:

[0083] S=S1×k1+S2×k2+S3×k3;

[0084] Where: S is the corrected live broadcast violation tendency; k1, k2, and k3 are the reference proportions of violation tendency; S1, S2, and S3 are the violation tendency based on motion trajectory, text information, and live broadcast frame performance, respectively.

[0085] The electronic tag content configured for the live broadcast violation tendency correction result is the correction time of the correction result;

[0086] Through the above logical formula, the violation tendency results of the quantified calculation are applied to comprehensively calculate the violation tendency of the live broadcast, so as to make a final judgment on whether the live broadcast room is in violation of the rules, and use the judgment results to provide further operation support for subsequent modules in this system;

[0087] The values ​​of the reference proportions of violation tendency k1, k2, and k3 are subject to:

[0088] k1, k2, and k3 are all positive numbers, and their sum is 1. When the live broadcast content in the live broadcast room is mainly based on the live broadcast user's image, k1>k2, k3>k2; when the live broadcast content in the live broadcast room is mainly based on the projection screen content, k1<k2, k3<k2;

[0089] The control module is used to obtain the judgment result of whether the live broadcast room has violated the regulations in the correction and judgment module. If the judgment result is yes, the live broadcast room will be blocked. If the judgment result is no, the system will be refreshed.

[0090] The message module is used to continuously obtain the correction results of the correction module on the violation tendency of the live broadcast room and the judgment result of whether the live broadcast room has violated the rules, and generate daily management messages for the live broadcast room based on the obtained information;

[0091] The content of the daily management message of the live broadcast room generated by the message module is sorted based on the time corresponding to their respective electronic tags;

[0092] The data acquisition module is interactively connected to the control unit and the preprocessing unit at the lower level through a wireless network, the data acquisition module is interactively connected to the priori database and the analysis module through a wireless network, the analysis module is interactively connected to the data acquisition module through a wireless network, the analysis module is interactively connected to the correction and judgment module through a wireless network, the correction and judgment module is interactively connected to the configuration unit at the lower level through a wireless network, and the correction and judgment module is interactively connected to the management and control module and the message module through a wireless network.

[0093] In this embodiment, the data acquisition module runs in the live broadcast room to collect data samples to be detected in real time, pre-processes the data samples to be detected, and the control unit synchronously controls the operating frequency of the data acquisition module so that the data acquisition module runs in real time to collect data samples to be detected based on the operating frequency controlled by the control unit. The pre-processing unit receives the live image data in the real-time collected data samples to be detected, identifies the dynamic target in the live image data, and captures the motion trajectory of the dynamic target in the live image data, and synchronously represents the motion trajectory with a trajectory line. After the motion trajectory of the dynamic target is captured, the live picture frame is captured based on the specified interval time area in the live image data, the priori database is post-operated to upload the illegal data samples, and the illegal data samples are stored. The analysis module then receives the pre-processed data samples to be detected collected by the data acquisition module, and the data samples to be detected are pre-processed. A similarity analysis is performed on the data sample and the illegal data sample in the prior database to obtain the violation tendency of the live broadcast room from which the data sample to be detected comes, and the correction and judgment module receives the violation tendency of the live broadcast room obtained based on the similarity analysis in the analysis module, corrects the violation tendency of the live broadcast room, and sets a judgment threshold. The judgment threshold is applied to compare with the corrected violation tendency of the live broadcast room to determine whether the live broadcast room is in violation. The configuration unit simultaneously configures an electronic tag for the correction result of the live broadcast violation tendency. The management and control module further obtains the judgment result of whether the live broadcast room is in violation in the correction and judgment module, and controls the live broadcast room to be blocked when the judgment result is yes. When the judgment result is no, the system is refreshed to run. Finally, the message module continuously obtains the correction result of the violation tendency of the live broadcast room and the judgment result of whether the live broadcast room is in violation in the correction module, and generates a daily management message for the live broadcast room based on the obtained information.

[0094] Through the operation of the system in the above embodiment, a more comprehensive supervision of violations is provided for the network live broadcast environment, ensuring a healthier network live broadcast environment;

[0095] The following is an application example of the system in the above embodiment:

[0096] Scene background:

[0097] An e-commerce platform needs to monitor livestreams for violations such as the display of prohibited products and inappropriate speech. Taking a beauty livestream as an example, the system collects and analyzes multi-dimensional data to quickly identify and address violations.

[0098] 1. Data Collection and Preprocessing

[0099] Trigger condition: When the anchor starts to showcase a new product, the system automatically starts data collection.

[0100] Dynamically adjust the acquisition frequency: According to the instantaneous changes in the number of visitors and barrage in the live broadcast room (such as a sudden increase in the number of visitors; 50%), the picture acquisition frequency is automatically increased (from 10 frames / second to 20 frames / second) to ensure that key pictures are captured.

[0101] Multimodal data capture:

[0102] Image and motion: Identify dynamic targets such as the host's hands, track their motion trajectory (such as the path of their hands) when picking up products, and extract multiple image frames (including close-ups of product displays) at fixed intervals (every 0.5 seconds).

[0103] Audio capture: Synchronously record live audio and convert it into text for analysis.

[0104] Data classification storage: The captured motion trajectories, picture frames, and audio clips are saved separately to facilitate subsequent dimensional detection.

[0105] II. Comparative Analysis of Illegal Samples

[0106] Pre-set violation database: Historical violation case data has been stored, including:

[0107] Dangerous action trajectory: such as the rapid waving trajectory of the hand in the act of "displaying a controlled knife";

[0108] Illegal screen marking: Screens containing unregistered cosmetics and infringing logos (marked with a box to mark the illegal objects);

[0109] Inappropriate speech text: such as "absolute language" and "guiding private transactions"; the text of the voice transcription.

[0110] Real-time detection logic:

[0111] Action compliance judgment:

[0112] The live streamer's trajectory of picking up the product is compared with the trajectory of historical illegal actions, and the trajectory similarity is calculated through an algorithm (combining time series alignment and curve shape matching). If the similarity exceeds a threshold (e.g., 80%), it is marked as "suspected dangerous action."

[0113] Compliance judgment of voice content:

[0114] The live broadcast audio is transcribed into text (e.g., "This essence works immediately") and compared with the illegal text database to analyze semantic similarity and structural consistency (e.g., detecting absolute terms such as "immediately"). If a match is found for illegal keywords, it will be marked as "suspected false advertising."

[0115] Screen content compliance judgment:

[0116] Extract the product outline and compare it with the illegal product labeling box (such as the outline of unregistered cosmetics). Calculate the outline similarity (polar coordinate distribution and Fourier feature matching). If the similarity exceeds 90%, it will be marked as "suspected illegal product display."

[0117] III. Comprehensive Assessment and Violation Determination

[0118] Multi-dimensional weight calculation: According to the live broadcast type (such as mainly based on the host's image), weights are assigned to the three types of detection results: action, voice, and picture (such as action accounts for 60%, voice accounts for 30%, and picture accounts for 10%), and the comprehensive violation score is calculated.

[0119] Example: If the action score is 0.7, the voice score is 0.8, and the picture score is 0.9, the overall score is:

[0120] 0.6×0.7+0.3×0.8+0.1×0.9=0.75.

[0121] Threshold comparison: The preset violation threshold is 0.6. If the comprehensive score exceeds the threshold (such as 0.75), the live broadcast room is judged to be in violation.

[0122] IV. Emergency Response and Data Recording

[0123] Real-time control: After determining a violation, the system will automatically interrupt the live broadcast, pop up a violation prompt, and notify manual review.

[0124] Management data generation: Generate reports containing violation types, scores, and timestamps in chronological order for platform archiving and subsequent compliance trend analysis (such as statistics on high-frequency violation scenarios).

[0125] In summary, in the above embodiments, during operation, the system collects live images, audio segments and other data, and conducts multi-dimensional similarity analysis of motion trajectories, text information, and contour images with historical violation data. The detection frequency is dynamically adjusted in combination with the instantaneous growth rate of visitor volume, revenue, and barrage volume. Dynamic time warping, local linear embedding, Fourier descriptor and other algorithms are used to quantify violation tendencies. According to the main type of live content, that is, user images or projection content, weights are intelligently allocated to correct violation tendencies. A judgment threshold is set for comparison to determine the violation status, and the detection results are used to control the suspension of the live broadcast room or system refresh, and generate management messages sorted by time. This mechanism breaks through the traditional single detection mode and improves the timeliness, comprehensiveness and accuracy of violation detection through multi-source data fusion and algorithm application, providing technical support for live broadcast supervision.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A real-time detection system for illegal content in live webcasts, characterized by: include: The data acquisition module is used to collect the data samples to be tested in real time in the live broadcast room and pre-process the data samples to be tested; A priori database, used to upload and store illegal data samples; An analysis module is used to receive the pre-processed data samples to be tested collected by the data collection module, and perform similarity analysis between the data samples to be tested and the illegal data samples in the prior database to obtain the illegal tendency of the live broadcast room from which the data samples to be tested originated; The correction and determination module is used to receive the violation tendency of the live broadcast room obtained based on the similarity analysis in the analysis module, correct the violation tendency of the live broadcast room, set a determination threshold, and compare the determination threshold with the corrected violation tendency of the live broadcast room to determine whether the live broadcast room violates the law; The control module is used to obtain the judgment result of whether the live broadcast room has violated the regulations in the correction and judgment module, and control the suspension of the live broadcast room if the judgment result is yes, and refresh the system operation if the judgment result is no; The message module is used to continuously obtain the correction results of the live broadcast room's violation tendencies and the judgment results of whether the live broadcast room is in violation in the correction module, and generate daily management messages for the live broadcast room based on the obtained information.

2. A real-time detection system for illegal content in live streaming according to claim 1, characterized in that: The types of data samples to be detected collected by the data acquisition module include: live images, and live audio segments corresponding to the live images. The data acquisition module is provided with a control unit and a pre-processing unit at a lower level. The control unit is used to control the operating frequency of the data acquisition module, so that the data acquisition module runs in real time to collect data samples to be detected based on the operating frequency controlled by the control unit. The pre-processing unit is used to receive live image data in the data samples to be detected collected in real time, identify dynamic targets in the live image data, and capture the motion trajectory of the dynamic target in the live image data, synchronously represent the motion trajectory with a trajectory line, and after the motion trajectory of the dynamic target is captured, capture a live picture frame based on a specified interval time area in the live image data; Among them, the pre-processing unit runs to capture no less than three live picture frames. After capturing the motion trajectory of the dynamic target and the live picture frames, the pre-processing unit distinguishes and stores the captured motion trajectory of the dynamic target, the live picture frames and the live audio segments in the data samples to be detected.

3. A real-time detection system for illegal content in live streaming according to claim 2, characterized in that: During the operation phase of the control unit, the instantaneous growth rate of visitors to the live broadcast room, the instantaneous growth rate of revenue, and the instantaneous refresh rate of the bullet screen are monitored in real time; Where: F viit is the instantaneous growth rate of visitors; m now is the number of online visitors in the live broadcast room at the current moment; m before Compared to m now The number of online visitors in the live broadcast room at the last moment; △t is the sampling interval; Among them, the calculation logic of the instantaneous growth rate of revenue, the instantaneous refresh rate of barrage and the instantaneous growth rate of visitors are the same. When calculating the instantaneous growth rate of revenue, the current moment revenue of the live broadcast room and the revenue of the previous moment of the live broadcast room are used for calculation. When calculating the instantaneous refresh rate of barrage, the current moment barrage volume of the live broadcast room and the previous moment barrage volume of the live broadcast room are used for calculation, which are respectively denoted as F viit 、F dk 、F gi , set the range and initial frequency of the control unit to control the operating frequency of the data acquisition module, so that the control unit controls the operating frequency of the data acquisition module within the set range based on the initial frequency, in F viit 、F dk 、F gi When any of the above items shows an upward or downward trend, the initial frequency is controlled to increase or decrease based on the specified ratio, and is maintained after it is increased or decreased to the maximum or minimum value of the set range.

4. A real-time detection system for illegal content in live webcasts according to claim 1, characterized in that: The illegal data samples stored in the a priori database are manually uploaded by the system end user. When the a priori database stores the illegal data samples, it differentiates and stores them based on the types of illegal data samples; Among them, the types of illegal data samples include: historical dynamic target motion trajectories, live broadcast frames with illegal object annotation boxes, and text information extracted from historical live broadcast audio segments. The illegal data samples stored in the priori database are manually updated, deleted, and modified based on the system end users.

5. A real-time detection system for illegal content in live webcasts according to claim 1, characterized in that: During the operation phase of the analysis module, the similarity analysis results are used to indicate the violation tendency of the live broadcast room where the data sample to be detected comes from; The similarity analysis targets in the analysis module are: dynamic target motion trajectories and historical dynamic target motion trajectories, text information extracted based on live audio segments and text information extracted from historical live audio segments, and contour images extracted from live screen frames and live screen frames with illegal object annotation boxes; When the similarity between the dynamic target's motion trajectory and the historical dynamic target's motion trajectory is used to represent the violation tendency: Where: S1 is the violation tendency based on the motion trajectory performance; α, (1-α) are weight coefficients; DTW (T a ,T b ) is the time series alignment distance between the dynamic target motion trajectory and the historical dynamic target motion trajectory based on dynamic time warping; F(C a ,C b ) is the Fréchet distance between the two main curves after extracting the main curves from the dynamic target motion trajectory and the historical dynamic target motion trajectory using the local linear embedding algorithm; Among them, α and (1-α) are both positive numbers, and α>(1-α). The larger S1 is, the more likely there is a violation in the live broadcast room where the dynamic target motion trajectory comes from. When calculating S1, each historical dynamic target motion trajectory is used as a similarity analysis target and a similarity analysis is performed with the dynamic target motion trajectory. Where: d(C a (s),C b (t)) is the main curve C a The point with parameter s on the principal curve C b Euclidean distance between points with upper parameter t.

6. A real-time detection system for illegal content in live webcasts according to claim 5, characterized in that: When the similarity between text information extracted from live audio segments and text information extracted from historical live audio segments is used to indicate violation trends: Where: S2 is the violation tendency expressed based on text information; σ is the global scaling factor; λ and (1-λ) are weight coefficients; SemSim is the semantic similarity of the two text messages; |C| is the size of the common word set after removing stop words; |T1| and |T2| are the number of words in the text information T1 extracted from the live audio segment and the text information T2 extracted from the historical live audio segment after removing stop words; StrustSim is the structural consistency of the two text messages; Among them, λ and (1-λ) are both positive numbers, and λ>(1-λ). When calculating S2, the text information extracted from each historical live audio segment is used as the similarity analysis target, and similarity analysis is performed with the text information extracted based on the live audio segment.

7. A real-time detection system for illegal content in live webcasts according to claim 5, characterized in that: When the violation tendency is represented by the similarity between the contour image extracted from the live video frame and the live video frame with the illegal object labeling box: Calculate the similarity between each contour in the contour image and the contour of the object in each illegal object annotation box; Where: s3 is the violation tendency based on the live frame performance; m(sim>90%) is the number of target contours in the contour image whose contours are more than 90% similar to the contours of the objects in the violation object annotation box; m all is the total amount of contours in the contour image; ω and ε are weights; D Polar is the polar coordinate distribution distance of the two contours; D Fourier is the Fourier descriptor distance between two contours; The value ranges of ω and ε are 0.1-3 and 0.5-1 respectively, and their default values ​​are 1 and 0.

8.

8. A real-time detection system for illegal content in live webcasts according to claim 1, characterized in that: The correction and determination module is provided with a configuration unit at a lower level, and the configuration unit is used to configure an electronic label for the correction result of the live broadcast violation tendency. The correction logic for the live broadcast violation tendency in the correction and determination module is expressed as follows: S=S1×k1+S2×k2+S3×k3; Where: S is the corrected live broadcast violation tendency; k1, k2, and k3 are the reference proportions of violation tendency; S1, S2, and S3 are the violation tendency based on motion trajectory, text information, and live broadcast frame performance, respectively. Among them, the electronic tag content configured for the live broadcast violation tendency correction result is the correction time of the correction result.

9. A real-time detection system for illegal content in live webcasts according to claim 1 or 8, characterized in that: The values ​​of the reference proportions of violation tendency k1, k2, and k3 are subject to: k1, k2, and k3 are all positive numbers, and their sum is 1. When the live broadcast content in the live broadcast room is mainly based on the live broadcast user's image, k1>k2, k3>k2; when the live broadcast content in the live broadcast room is mainly based on the projection screen content, k1<k2, k3<k2; The contents of the daily management messages of the live broadcast room generated by the message module are sorted based on the corresponding time of their respective electronic tags.

10. A real-time detection system for illegal content in live webcasts according to claim 1, characterized in that: The data acquisition module is interactively connected to a control unit and a preprocessing unit at the lower level through a wireless network, the data acquisition module is interactively connected to a priori database and an analysis module through a wireless network, the analysis module is interactively connected to the data acquisition module through a wireless network, the analysis module is interactively connected to a correction and determination module through a wireless network, the correction and determination module is interactively connected to a configuration unit at the lower level through a wireless network, and the correction and determination module is interactively connected to a management and control module and a message module through a wireless network.

Images