A multi-level video risk mining method, device and medium

Abstract

The present invention discloses a multi-level video risk mining method, device and medium, including: obtaining video data, environmental data and comment data; extracting key frames in the video data and performing risk detection to obtain detection information of each key frame; obtaining user characteristics, and cleaning each key frame according to user characteristics, environmental data and detection information to obtain key frames in a standardized format; pre-configuring risk identification rules; matching key frames in a standardized format with risk identification rules, and counting violations according to matching results; determining whether a hit is found according to the violation counting result, and performing corresponding disposal on the hit video data; and reviewing the video data through the comment data; and marking and grading the video data according to the disposal and review results. The present invention can improve the quality of risk mining.

Description

Technical Field

[0001] The present invention relates to the technical field of video risk mining, and in particular to a multi-level video risk mining method, device and medium. Background Art

[0002] With the surge in the amount of live streaming content, how to effectively manage and monitor these live streaming contents to ensure that they comply with legal regulations and social ethical standards has become an important challenge facing live streaming platforms.

[0003] Traditional methods for mining risks in live / video files generally use a frame-by-frame + manual approach, that is, screenshots of live / video files frame by frame. Since each frame captured manually only includes one picture, the text in the picture is identified by the human eye, and then the content of each frame is analyzed based on the text information to determine whether there are risks that do not comply with legal norms and social moral standards. Although the manual mining method has a great inhibitory effect on risks, it consumes a lot of manpower costs; and the manual judgment method is prone to misjudgment and risk omissions when the text information is changeable and in the form of separate fragments. Therefore, how to identify risks through multiple levels is a great technical difficulty. Summary of the invention

[0004] In view of this, the purpose of the present invention is to propose a multi-level video risk mining method, which uses a set of multi-level and multi-dimensional rules to conduct risk mining on live broadcast / video files in social media, and achieves the goal of high-quality risk mining through machine detection, rule matching and user behavior analysis without affecting business processes.

[0005] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0006] The present invention provides a multi-level video risk mining method, comprising the following steps:

[0007] Step 1: Obtain video data, environment data, and comment data;

[0008] Step 2: extract key frames from the video data and perform risk detection to obtain detection information of each key frame;

[0009] Step 3: Obtain user features, and clean each key frame according to the user features, environmental data, and detection information to obtain key frames in a standardized format;

[0010] Step 4: Pre-configure risk identification rules;

[0011] Step 5: Match the key frames in the standardized format with the risk identification rules, and count the violations based on the matching results;

[0012] Step 6: determine whether the violation is a hit according to the violation counting result, and perform corresponding treatment on the hit video data; and review the video data through the comment data;

[0013] Step 7: Label and grade the video data based on the processing and review results.

[0014] Furthermore, the step 1 specifically includes:

[0015] Step 11: Use RTC and HTTP protocols to establish real-time communication between the acquisition end and multiple live broadcast ends. The acquisition end receives the video data address and collects corresponding video data from different live broadcast ends according to the video data address. The video data is a video stream or a video file.

[0016] Step 12: Acquire environmental data through the client, where the environmental data includes the user's location information;

[0017] Step 13: Collect the comment content corresponding to the video data and push it into the message list.

[0018] Furthermore, the step 2 specifically includes:

[0019] Step 21, converting video data in different formats into video data in a video coding format;

[0020] Step 22: inputting the video data after format conversion into the risk detection model;

[0021] Step 23, the risk detection model divides the video data into multiple segments according to a set time interval based on a computer vision algorithm, and the set time interval ranges from 1 to 5 seconds, and determines whether the image difference between the first video frame of the current segment and the first video frame of the next segment is within the allowable difference range. If so, the first video frame of the current segment is used as a key frame, if not, each video frame of the current segment is used as a key frame, and for the last segment, each video frame in the last segment is used as a key frame; each key frame includes a picture and / or corresponding audio;

[0022] Step 24: the risk detection model identifies multiple groups of texts from multiple key frames based on speech recognition technology and OCR text recognition software, where one key frame corresponds to one group of texts;

[0023] Step 25: The risk detection model matches each group of texts with pre-stored sensitive words, detects whether there is a risk through the matching results, and returns the detection information corresponding to each key frame; the detection information includes the frame capture time, risk result, risk label, identified text, violation status and content type.

[0024] Furthermore, the step 3 specifically includes:

[0025] Step 31: Obtain user characteristics from a user database, wherein the user characteristics include gender, user identity, and charm value;

[0026] Step 32, setting a standardized format to sequentially splice user features, user location information, and detection information;

[0027] Step 33, according to the set standardized format, the user features, user location information and specific contents of the detection information corresponding to each key frame are sequentially spliced ​​to obtain each key frame with a standardized format. The standardized format of each key frame is specifically: gender-user identity-charm value-user location information-frame capture time-risk result-risk label-identified text-violation status-content type.

[0028] Furthermore, the step 4 specifically includes:

[0029] Step 41: pre-configure risk identification rules, including rule content, user data rules, and hit results; wherein:

[0030] The rule content includes: content type, risk label, risk result, counting method and counting type; the content type includes framed picture, framed voice and empty voice, the risk label includes at least one preset label content, and the risk result includes rejection and suspected; the counting method includes label or rule, and the counting type includes continuous or cumulative and threshold N;

[0031] The user data rules include gender, user identity, charm value, user location information and violation status; the user data rules are to limit or not limit the specific content of gender, user identity, charm value, user location information and violation status according to the needs of the user;

[0032] The hit result includes a treatment type, a hit tag, and a reset count; the treatment type includes observation or treatment, the hit tag includes warning or advertising ticket, and the reset count includes clearing the count after the hit or not clearing the count after the hit;

[0033] Step 42: Upload the risk identification rules as configuration information to a configuration center;

[0034] Step 43: When the configuration information changes, the configuration center publishes the new configuration information, and multiple servers subscribe to and update the new configuration information according to the published information.

[0035] Furthermore, the step 5 specifically includes:

[0036] Step 51, selecting the first key frame in the standardized format as the current key frame according to the sequence of frame capture time;

[0037] Step 52: Match the current key frame with the content type, risk label, risk result and user data rules in the risk identification rules one by one. If the match is successful, the current key frame is used as the target frame; if the match fails, proceed to step 54;

[0038] Step 53: Count the target frame according to the counting method and counting type in the risk identification rule;

[0039] Step 54: Select the next key frame as the current key frame, and repeat steps 52 to 53 to complete the counting of all target frames.

[0040] Furthermore, the step 52 specifically includes:

[0041] Step 521, matching the content type in the current key frame with the content type in the risk identification rule, and determining whether the content type in the current key frame matches any one of the captured image, captured voice, and empty voice. If so, proceed to step 522; if not, abandon the current key frame, select the next key frame as the current key frame, and return to step 521;

[0042] Step 522: match the risk label in the current key frame with the risk label in the risk identification rule to determine whether the risk label in the current key frame meets any of the preset label contents. If so, proceed to step 523; if not, abandon the current key frame, select the next key frame as the current key frame, and return to step 521;

[0043] Step 523, matching the risk result in the current key frame with the risk result in the risk identification rule, and determining whether the risk result in the current key frame meets any one of rejection and suspicion. If so, proceed to step 524; if not, abandon the current key frame, select the next key frame as the current key frame, and return to step 521;

[0044] Step 524: determine whether there is a user data rule in the risk identification rule. If not, proceed to step 525. If yes, determine the limited content contained in the user data rule, and match the limited content in the user data rule with the corresponding content in the current key frame one by one. If they are completely matched, proceed to step 525. If they are not completely matched, abandon the current key frame, select the next key frame as the current key frame, and return to step 521.

[0045] Step 525, taking the current key frame as the target frame and proceeding to step 531;

[0046] The step 53 specifically includes:

[0047] Step 531: Determine the counting rule according to the counting method and counting type configured in the risk identification rule, including four combinations, namely, rule continuity, rule accumulation, label continuity and label accumulation;

[0048] Step 532: if the adopted counting rule is rule continuity, then generate a corresponding counting key1, wherein the counting key1 includes the counting method as rule, the risk identification rule ID and the video data ID, and the number of the counting key1 is 1; determine whether the current target frame is the first target frame, if so, then count the corresponding position of the counting key1 as 1; if not, determine whether the current target frame is continuous with the previous target frame according to the frame capture time and the set time interval, if so, perform incremental counting at the corresponding position of the counting key1, if not, reset the count and count again;

[0049] Step 533: if the adopted counting rule is rule accumulation, a corresponding counting key1 is generated, wherein the counting key1 includes the counting mode as rule, the risk identification rule ID and the video data ID, and it is determined whether the current target frame matches the counting key1, and the number of the counting key1 is 1; it is determined whether the current target frame is the first target frame, and if so, the corresponding position of the counting key1 is counted as 1; if not, an incremental count is performed at the corresponding position of the counting key1;

[0050] Step 534: if the adopted counting rule is label continuity, then generate a corresponding counting key2 according to different label contents in the risk label in the risk identification rule, the counting key2 includes a counting method of label, a risk identification rule ID, a video data ID and a label content, and the number of the counting key2 is consistent with the number of label contents preset in the risk label in the risk identification rule; match a corresponding counting key2 according to the risk label in the current target frame, and judge whether the current target frame is the first target frame on the paired counting key2, if yes, then count as 1 at the corresponding position of the paired counting key2; if not, judge whether the current target frame is continuous with the previous target frame in the paired counting key2 according to the frame cut time and the set time interval, if yes, then perform incremental counting at the corresponding position of the paired counting key2, if not, reset the count and count again;

[0051] Step 535: If the adopted counting rule is label accumulation, a corresponding counting key2 is generated according to different label contents in the risk label in the risk identification rule, and the counting key2 includes a counting method of a label, a risk identification rule ID, a video data ID and a label content, and the number of the counting key2 is consistent with the number of label contents preset in the risk label in the risk identification rule; a corresponding counting key2 is matched according to the risk label in the current target frame, and it is determined whether the current target frame is the first target frame on the paired counting key2, if yes, the count at the corresponding position of the paired counting key2 is 1; if not, an incremental count is performed at the corresponding position of the paired counting key2.

[0052] Furthermore, the step 6 specifically includes:

[0053] Step 61: determine whether the final counting result is greater than the threshold value N. If yes, proceed to step 62; if not, do not process;

[0054] Step 62: Process the video and / or user according to the treatment type and hit tag in the risk identification rule:

[0055] If the disposal type is observation, the video data is pushed to relevant personnel using websocket communication, and the priority of the video data is increased for priority processing, and the processing results are recorded;

[0056] If the disposal type is disposal, the video data will be deleted and the user will be punished;

[0057] If the hit tag is warning, the user is warned;

[0058] If the hit tag is an advertising ticket, an advertising ticket will be issued to the user;

[0059] Step 63: After executing a disposal, warning or advertising ticket, record the penalty time and store it in the penalty queue; when the next hit occurs, record the current hit time and store it in the penalty queue; in the penalty queue, determine whether the current hit time and the previous penalty time are within the set time period, if not, execute the corresponding penalty; if yes, do not repeat the penalty, and remove the current hit time from the penalty queue;

[0060] Step 64: Check whether there is user feedback and reporting content based on the comment content under the video data. If yes, review and process the published video data; if not, do not process it.

[0061] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a multi-level video risk mining method as described above is implemented.

[0062] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the multi-level video risk mining method as described above is implemented.

[0063] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0064] 1. Use RTC and HTTP protocols to achieve real-time communication and data capture, ensure real-time transmission and high-definition presentation of video streams, and enhance user interaction experience.

[0065] 2. Use key frame extraction: select key frames from video data at set time intervals to accelerate the risk detection process of video content.

[0066] 3. Obtain multi-dimensional user features by querying the user database, and combine data cleaning and standardization technology. According to the set standardized format, the user features corresponding to each key frame, the user's location information, and the specific content of the detection information are sequentially spliced ​​to obtain each key frame with a standardized format, ensuring the comprehensiveness and depth of the video content analysis.

[0067] 4. Design a preset rule mechanism. The risk identification rules include rule content, user data rules, and hit results. The rule content uses AND logic to combine content type, risk label, risk result, counting method, and counting type. The counting method includes rules or labels to implement counting logic at the rule or label level.

[0068] 5. The counting types include continuous or cumulative counting modes, providing flexible configuration options at the counting level, and supporting count reset or continuous accumulation after a hit, ensuring accurate monitoring and timely response to violations.

[0069] 6. Use a risk detection model, which is responsible for automatically advancing key frames, automatically identifying text, matching text with sensitive words, and outputting detection information. This series of operations is implemented by a deep learning network, reducing labor costs. In addition, the images and audio in the key frames can be recognized by the risk detection model using different expressions, thereby improving the accuracy of risk identification.

[0070] 7. The present invention performs risk detection on video data through risk detection models, key frames in standardized formats and risk identification rules, and reviews video data through comment data, making comprehensive assessments from different aspects. The results are more objective and reduce risk omissions caused by human misjudgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0072] Figure 1 It is an execution flow chart of a multi-level video risk mining method provided by an embodiment of the present invention.

[0073] Figure 2 It is a schematic diagram of an electronic device provided by an embodiment of the present invention.

[0074] Figure 3 It is a schematic diagram of a computer-readable storage medium provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The present invention will be further described in detail below in conjunction with the accompanying drawings and examples. It is particularly noted that the following examples are only used to illustrate the present invention, but are not intended to limit the scope of the present invention. Similarly, the following examples are only partial embodiments of the present invention rather than all embodiments, and all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0076] See also Figure 1 A multi-level video risk mining method of the present invention comprises the following steps:

[0077] Step 1: Obtain video data, environment data, and comment data;

[0078] In this embodiment, the step 1 specifically includes:

[0079] Step 11: Use RTC and HTTP protocols to establish real-time communication between the acquisition end and multiple live broadcast ends. The use of RTC and HTTP protocols as real-time communication technology can ensure that the acquisition end and the live broadcast end are in real-time communication, ensure that the video data can be transmitted in time and presented with high quality, and enhance the user interaction experience; the acquisition end receives the video data address, and collects corresponding video data from different live broadcast ends according to the video data address, and the video data is a video stream or a video file;

[0080] Step 12: Acquire environmental data through the client, where the environmental data includes the user's location information;

[0081] Step 13: Collect the comments corresponding to the video data. The comments serve as an auxiliary element for risk judgment. The comments are pushed into the message list to facilitate quick recall of the comments for processing later.

[0082] Step 2: extract key frames from the video data and perform risk detection to obtain detection information of each key frame;

[0083] In this embodiment, step 2 specifically includes:

[0084] Step 21, uniformly converting video data of different formats into video data of video coding format; specifically, first determine whether the format of the video data is a video coding format, if so, no format conversion is required; if not, convert the current format into a video coding format; uniformly converting different formats into a video coding format can facilitate subsequent unified processing of key frame extraction;

[0085] Step 22: inputting the video data after format conversion into the risk detection model;

[0086] Step 23, the risk detection model divides the video data into multiple segments according to a set time interval based on a computer vision algorithm, and the set time interval ranges from 1 to 5 seconds. It is determined whether the image difference between the first video frame of the current segment and the first video frame of the next segment is within the allowable difference range. If so, the first video frame of the current segment is used as a key frame. If not, each video frame of the current segment is used as a key frame. For the last segment, each video frame in the last segment is used as a key frame; each key frame includes a picture and / or corresponding audio; the key frame is automatically extracted through a computer vision algorithm to avoid errors and inefficiencies caused by manual extraction. By efficiently extracting key frames from video data, the consumption of computing resources is significantly reduced, and the risk detection process of video content is accelerated. Among them, the specific value of the set time interval can be set according to the detection requirements. Here, 2s is taken as an example, such as: every 2s. For example: the total length of video data is 11s, and the time interval is set to every 2s, then the video data can be divided into 6 segments, and a video frame is randomly selected from each segment as a key frame, and the pictures and / or audios of the 1st, 3rd, 5th, 7th, 9th and 11th seconds of the video data can be key frames.

[0087] Step 24, the risk detection model identifies multiple groups of texts from multiple key frames based on speech recognition technology (ASR) and OCR text recognition software, and one key frame corresponds to a group of texts; specifically, the image is converted to text by OCR text recognition software; the audio is converted to text by speech recognition technology;

[0088] If a key frame contains both images and audio, multiple characters are recognized by OCR text recognition software and speech recognition technology respectively and then integrated into a group of texts; if a key frame contains only images, multiple characters are recognized by OCR text recognition software and then integrated into a group of texts; if a key frame contains only audio, multiple characters are recognized by speech recognition technology and then integrated into a group of texts; if a key frame contains only audio and no text is recognized by speech recognition technology, it is defined as blank speech.

[0089] Step 25, the risk detection model matches each group of texts with pre-stored sensitive words, and detects whether there is a risk through the matching results. Specifically, if sensitive words are matched in the text, it is detected that there may be a risk; if sensitive words are not matched in the text, it is detected that there may be no risk; the process of risk detection through the risk detection model not only improves the efficiency of video data processing, but also ensures the accuracy and real-time nature of risk identification.

[0090] And return the detection information corresponding to each key frame; the detection information includes the cut-frame time, risk result, risk label, identified text, violation status and content type. Among them, the cut-frame time is a specific time point, the purpose is to distinguish the order of each key frame, so that the key frames can be selected in order and whether the two key frames are continuous during the subsequent counting; the risk results include pass, suspected and rejected; the risk label includes none and specific risky label content (such as: "face: no face: no face", "advertisement: marketing advertisement: recruitment advertisement", "prohibited: bad artist: bad star"); the identified text includes specific text content; the violation status includes mute and can speak; the content type includes cut-frame picture, cut-frame audio, blank audio and cut-frame picture + audio.

[0091] For example: Output the detection information of the first key frame: screenshot time = 20240906101501, risk result = "suspected", risk label = "face: no face: no face", recognized text = "Welcome to the live broadcast room", violation status = "banned", content type = "screenshot image";

[0092] Output the detection information of the second key frame: frame capture time = 20240906101503, risk result = "reject", risk label = "advertisement: marketing advertisement: recruitment advertisement", recognized text = "Welcome to the live broadcast room", violation status = "banned", content type = "frame capture audio";

[0093] Output the detection information of the third key frame: frame capture time = 20240906101505, risk result = "suspected", risk label = "prohibited: bad artist: bad star", identified text = "Welcome to the live broadcast room", violation status = "banned", content type = "frame capture image + audio";

[0094] Output the detection information of the 4th key frame: frame time = 20240906101507, risk result = "reject", risk label = "prohibited: bad artist: bad star", recognized text = none, violation status = "banned", content type = "blank audio";

[0095] Output the detection information of the fifth key frame: frame capture time = 20240906101509, risk result = "passed", risk label = none, recognized text = "Welcome to the live broadcast room", violation status = "can speak", content type = "frame capture audio";

[0096] Output the detection information of the 6th key frame: screenshot time = 20240906101511, risk result = "pass", risk label = none, recognized text = "Welcome to the live broadcast room", violation status = "can speak", content type = "screenshot image".

[0097] Step 3: Obtain user features, and clean each key frame according to the user features, environmental data, and detection information to obtain key frames in a standardized format;

[0098] In this embodiment, step 3 specifically includes:

[0099] Step 31: Obtain user characteristics from a user database, wherein the user characteristics include gender, user identity, and charm value;

[0100] Step 32, setting a standardized format to sequentially splice user features, user location information, and detection information;

[0101] Step 33, according to the set standardized format, the user features, user location information and specific contents of the detection information corresponding to each key frame are sequentially spliced ​​to obtain each key frame with a standardized format. The standardized format of each key frame is specifically: gender-user identity-charm value-user location information-frame capture time-risk result-risk label-identified text-violation status-content type.

[0102] The key frames of the video data are cleaned and standardized to analyze the video content more comprehensively and facilitate subsequent use. There are multiple key frames in a video data. The following is the result after cleaning:

[0103] For example: the first keyframe: Gender = Male, User ID = Anchor, Charm Value = 60, Location = Shanghai, Screenshot Time = 20240906101501, Risk Result = "Suspected", Risk Label = "Face: No Face: No Face", Recognized Text = "Welcome to the Live Broadcast Room", Violation Status = "Mute", Content Type = "Screenshot Image";

[0104] The second key frame: Gender = Male, User ID = Anchor, Charm Value = 60, Location = Shanghai, Screenshot Time = 20240906101503, Risk Result = "Rejected", Risk Label = "Advertisement: Marketing Advertisement: Recruitment Advertisement", Identified Text = "Welcome to the Live Broadcast Room", Violation Status = "Mute", Content Type = "Screenshot Audio";

[0105] The third key frame: Gender = Male, User ID = Anchor, Charm Value = 60, Location = Shanghai, Screenshot Time = 20240906101505, Risk Result = "Suspected", Risk Label = "Prohibited: Bad Artist: Bad Star", Identified Text = "Welcome to the Live Broadcast Room", Violation Status = "Mute", Content Type = "Screenshot Image + Audio";

[0106] The fourth key frame: Gender = Male, User ID = Anchor, Charm Value = 60, Location = Shanghai, Frame Time = 20240906101507, Risk Result = "Rejected", Risk Label = "Prohibited: Bad Artist: Bad Star", Identified Text = None, Violation Status = "Mute", Content Type = "Blank Audio";

[0107] The fifth key frame: Gender = Male, User ID = Anchor, Charm Value = 60, Location = Shanghai, Screenshot Time = 20240906101509, Risk Result = "Passed", Risk Label = None, Identified Text = "Welcome to the Live Broadcast Room", Violation Status = "Can Speak", Content Type = "Screenshot Audio";

[0108] The 6th key frame: Gender = Male, User identity = Anchor, Charm value = 60, Location = Shanghai, Screenshot time = 20240906101511, Risk result = "Pass", Risk label = None, Identified text = "Welcome to the live broadcast room", Violation status = "Can speak", Content type = "Screenshot image".

[0109] Step 4: Pre-configure risk identification rules;

[0110] In this embodiment, step 4 specifically includes:

[0111] Step 41: Pre-configure risk identification rules, including rule content, user data rules, and hit results. To improve the accuracy of risk mining, refine the standards and conditions of risk identification rules and design the rules into three dimensions: "rule content, user data rules, and hit results". Among them:

[0112] The rule content includes: content type, risk label, risk result, counting method and counting type; the content type includes framed picture, framed voice and empty voice, the risk label includes at least one preset label content, and the risk result includes rejection and suspected; the counting method includes label or rule, and the counting type includes continuous or cumulative and threshold N;

[0113] The user data rules include gender, user identity, charm value, user location information and violation status; the user data rules are to limit or not limit the specific content of gender, user identity, charm value, user location information and violation status according to the needs of the user;

[0114] The hit result includes a treatment type, a hit tag, and a reset count; the treatment type includes observation or treatment, the hit tag includes warning or advertising ticket, and the reset count includes clearing the count after the hit or not clearing the count after the hit;

[0115] Users can configure appropriate risk identification rules according to their needs. For example, they can be configured as:

[0116] Rule 1:

[0117] Rule content: a) Content type: "framed image, framed audio, blank audio";

[0118] b) Risk tags: "Advertisement: Marketing Advertisement: Recruitment Advertisement", "Face: No Face: No Face";

[0119] c) Risk outcome: rejected, suspected;

[0120] d) Counting method: rules;

[0121] e) Count type: continuous, the counting threshold is 1;

[0122] 2. User data rules: Male, no speech;

[0123] 3. Hit result:

[0124] a) Disposition type: Observation;

[0125] b) Hit tag: warning;

[0126] c) Reset count: the count is reset to zero after a hit;

[0127] Or rule 2:

[0128] 1. Rules content:

[0129] a) Content type: "frame capture image, frame capture audio, blank audio";

[0130] b) Risk tags: "Advertisement: Marketing Advertisement: Recruitment Advertisement", "Face: No Face: No Face";

[0131] c) Risk outcome: rejected, suspected;

[0132] d) Counting method: label;

[0133] e) Count type: cumulative, the counting threshold is 1;

[0134] 2. Hit result:

[0135] a) Disposal type: Disposal;

[0136] b) Hit tag: advertising ticket;

[0137] c) Reset count: The count is not reset after a hit.

[0138] Step 42: Upload the risk identification rules as configuration information to a configuration center;

[0139] Step 43, when the configuration information changes, the configuration center publishes the new configuration information, and multiple servers subscribe to and update the new configuration information based on the published information. The concept of "scenario-based" is introduced into the configuration system to adapt to the diverse needs of video services. In order to enhance the efficiency of risk matching and reduce the overhead of rule queries, the rules are pre-loaded into the memory when the system starts. In addition, in order to improve the flexibility and response speed of rule configuration synchronization to other servers, a front-end management interface is specially designed. By utilizing the publish-subscribe model of the message middleware, when the configuration changes, the update information can be broadcast in real time to ensure that all relevant servers synchronously refresh the rule sets in their memory, thereby achieving dynamic adjustment and immediate effectiveness.

[0140] Step 5: Match the key frames in the standardized format with the risk identification rules, and count the violations based on the matching results;

[0141] In this embodiment, step 5 specifically includes:

[0142] Step 51, selecting the first key frame in the standardized format as the current key frame according to the selection order of the frame capture time;

[0143] Step 52: Match the current key frame with the content type, risk label, risk result and user data rules in the risk identification rules one by one. If the match is successful, the current key frame is used as the target frame; if the match fails, proceed to step 54;

[0144] Step 53: Count the target frame according to the counting method and counting type in the risk identification rule;

[0145] Step 54: Select the next key frame as the current key frame, and repeat steps 52 to 53 to complete the counting of all target frames.

[0146] In this embodiment, step 52 specifically includes:

[0147] Step 521, matching the content type in the current key frame with the content type in the risk identification rule, and determining whether the content type in the current key frame matches any one of the captured image, captured voice, and empty voice. If so, proceed to step 522; if not, abandon the current key frame, select the next key frame as the current key frame, and return to step 521;

[0148] Step 522: match the risk label in the current key frame with the risk label in the risk identification rule to determine whether the risk label in the current key frame meets any of the preset label contents. If so, proceed to step 523; if not, abandon the current key frame, select the next key frame as the current key frame, and return to step 521;

[0149] Step 523, matching the risk result in the current key frame with the risk result in the risk identification rule, and determining whether the risk result in the current key frame meets any one of rejection and suspicion. If so, proceed to step 524; if not, abandon the current key frame, select the next key frame as the current key frame, and return to step 521;

[0150] Step 524: determine whether there is a user data rule in the risk identification rule. If not, proceed to step 525. If yes, determine the limited content contained in the user data rule, and match the limited content in the user data rule with the corresponding content in the current key frame one by one. If they are completely matched, proceed to step 525. If they are not completely matched, abandon the current key frame, select the next key frame as the current key frame, and return to step 521.

[0151] Step 525, taking the current key frame as the target frame and proceeding to step 531;

[0152] The step 53 specifically includes:

[0153] Step 531: Determine the counting rule according to the counting method and counting type configured in the risk identification rule, including four combinations, namely, rule continuity, rule accumulation, label continuity and label accumulation;

[0154] Step 532: If the adopted counting rule is rule continuity, a corresponding counting key1 is generated, wherein the counting key1 includes a counting method of a rule, a risk identification rule ID and a video data ID, and the number of the counting key1 is 1; it is determined whether the current target frame is the first target frame, and if so, the count at the corresponding position of the counting key1 is 1; if not, it is determined whether the current target frame is continuous with the previous target frame according to the frame capture time and the set time interval, and if so, an incremental count is performed at the corresponding position of the counting key1, and if not, the count is reset and then counted again; the counting function of the cache is utilized to accumulate the count value through a specific key, such as the INCR command of Redis for incremental counting to obtain the count value.

[0155] Step 533: if the adopted counting rule is rule accumulation, a corresponding counting key1 is generated, wherein the counting key1 includes the counting mode as rule, the risk identification rule ID and the video data ID, and it is determined whether the current target frame matches the counting key1, and the number of the counting key1 is 1; it is determined whether the current target frame is the first target frame, and if so, the corresponding position of the counting key1 is counted as 1; if not, an incremental count is performed at the corresponding position of the counting key1;

[0156] Step 534: if the adopted counting rule is label continuity, then generate a corresponding counting key2 according to different label contents in the risk label in the risk identification rule, the counting key2 includes a counting method of label, a risk identification rule ID, a video data ID and a label content, and the number of the counting key2 is consistent with the number of label contents preset in the risk label in the risk identification rule; match a corresponding counting key2 according to the risk label in the current target frame, and judge whether the current target frame is the first target frame on the paired counting key2, if yes, then count as 1 at the corresponding position of the paired counting key2; if not, judge whether the current target frame is continuous with the previous target frame in the paired counting key2 according to the frame cut time and the set time interval, if yes, then perform incremental counting at the corresponding position of the paired counting key2, if not, reset the count and count again;

[0157] Step 535: If the adopted counting rule is label accumulation, a corresponding counting key2 is generated according to different label contents in the risk label in the risk identification rule, and the counting key2 includes a counting method of a label, a risk identification rule ID, a video data ID and a label content, and the number of the counting key2 is consistent with the number of label contents preset in the risk label in the risk identification rule; a corresponding counting key2 is matched according to the risk label in the current target frame, and it is determined whether the current target frame is the first target frame on the paired counting key2, if yes, the count at the corresponding position of the paired counting key2 is 1; if not, an incremental count is performed at the corresponding position of the paired counting key2.

[0158] For example: Rule 1:

[0159] ①For the first key frame:

[0160] The content type = "screenshot image" in the first keyframe belongs to one of the content types in rule 1: "screenshot image, screenshot voice, blank voice"; the risk label = "face: no face: no face" in the first keyframe belongs to one of the risk labels in rule 1: "advertisement: marketing advertisement: recruitment advertisement", "face: no face: no face"; the risk result = "suspected" in the first keyframe belongs to one of the risk results in rule 1: "rejected, suspected";

[0161] Further judgment: Gender = Male and Violation Status = "Mute" in the first key frame both meet the user data rule in Rule 1: Male, Mute; and there is no need to judge the "user identity, charm value, and location" that are not limited.

[0162] The counting rule is: rule continuity (rule continuity here means that different label contents are not distinguished when counting), the generated count key1 = rule + rule 1 + video data ID. At this time, the current key frame is the first target frame, so the count key1 = rule + rule 1 + video data ID + quantity 1;

[0163] ②For the second key frame:

[0164] The content type = "framed audio" in the second keyframe belongs to the content type in rule 1: "framed image, framed voice, blank voice"; the risk label = "advertisement: marketing advertisement: recruitment advertisement" in the second keyframe belongs to the risk label in rule 1: "advertisement: marketing advertisement: recruitment advertisement", "face: no face: no face"; the risk result = "reject" in the second keyframe belongs to the risk result in rule 1: "reject, suspected";

[0165] Further judgment: Gender = Male and Violation Status = "Mute" in the second key frame both meet the user data rule in Rule 1: Male, Mute; and there is no need to judge the "user identity, charm value, and location" that are not limited.

[0166] The counting rule is: rule continuity (rule continuity here means that different label contents are not distinguished when counting). At this time, counting key1 = rule + rule 1 + video data ID + quantity 1. According to the current key frame capture time 20240906101503, the first key frame capture time 20240906101501 and the set time interval 10s, it is judged to be continuous, so counting key1 = rule + rule 1 + video data ID + quantity 2;

[0167] ③For the third key frame:

[0168] The content type = "screenshot image + audio" in the third keyframe belongs to the content types in rule 1: "screenshot image, screenshot voice, blank voice"; the risk label = "prohibited: bad artist: bad star" in the third keyframe does not belong to the risk labels in rule 1: "advertisement: marketing advertisement: recruitment advertisement", "face: no face: no face"; then the third keyframe is not the target frame and is not counted;

[0169] ④For the 4th key frame:

[0170] The content type = "blank audio" in the 4th keyframe belongs to one of the content types in rule 1: "framed picture, framed voice, blank voice"; the risk label = "prohibited: bad artist: bad star" in the 4th keyframe does not belong to the risk labels in rule 1: "advertisement: marketing advertisement: recruitment advertisement", "face: no face: no face"; then the 4th keyframe is not the target frame, and the 4th keyframe is not counted;

[0171] ⑤For the 5th key frame:

[0172] The content type = "Screenshot" in the 5th keyframe belongs to one of the content types in rule 1: "Screenshot, Screenshot Voice, Blank Voice"; the risk label in the 5th keyframe = None, which does not belong to the risk labels in rule 1: "Advertisement: Marketing Advertisement: Recruitment Advertisement", "Face: No Face: No Face"; therefore, the 5th keyframe is not a target frame and is not counted;

[0173] ⑥For the 6th key frame:

[0174] The content type = "Screenshot" in the 6th keyframe belongs to one of the content types in rule 1: "Screenshot, Speech, Blank Speech"; the risk label in the 6th keyframe = None, which does not belong to the risk labels in rule 1: "Advertisement: Marketing Advertisement: Recruitment Advertisement", "Face: No Face: No Face"; therefore, the 6th keyframe is not a target frame and is not counted;

[0175] The result of counting according to rule 1 is: count key1 = rule + rule 1 + video data ID + quantity 2.

[0176] Rule 2:

[0177] ①For the first key frame:

[0178] The content type = "screenshot image" in the first keyframe belongs to one of the content types in rule 1: "screenshot image, screenshot voice, blank voice"; the risk label = "face: no face: no face" in the first keyframe belongs to one of the risk labels in rule 1: "advertisement: marketing advertisement: recruitment advertisement", "face: no face: no face"; the risk result = "suspected" in the first keyframe belongs to one of the risk results in rule 1: "rejected, suspected";

[0179] No user data rules are configured in Rule 2, that is, gender, user identity, charm value, location and violation status are not limited, and there is no need to judge these contents.

[0180] The counting rule is: label accumulation (label accumulation here means distinguishing different label contents when counting). The two generated counting keys 2 are counting key 21 = label + rule 2 + video data ID + advertisement: marketing advertisement: recruitment advertisement, and counting key 22 = label + rule 2 + video data ID + face: no face: no face. At this time, the current key frame is the first target frame, and the corresponding risk label = "face: no face: no face" is paired with counting key 22 = label + rule 2 + video data ID, so counting key 22 = label + rule 2 + video data ID + quantity 1;

[0181] ②For the second key frame:

[0182] The content type = "framed audio" in the second keyframe belongs to the content type in rule 1: "framed image, framed voice, blank voice"; the risk label = "advertisement: marketing advertisement: recruitment advertisement" in the second keyframe belongs to the risk label in rule 1: "advertisement: marketing advertisement: recruitment advertisement", "face: no face: no face"; the risk result = "reject" in the second keyframe belongs to the risk result in rule 1: "reject, suspected";

[0183] The counting rule is: label accumulation. In the second keyframe, the risk label = "advertisement: marketing advertisement: recruitment advertisement". The matching is count key21 = label + rule 2 + video data ID. Then count key21 = label + rule 2 + video data ID + quantity 1.

[0184] ③For the third key frame:

[0185] The content type = "screenshot image + audio" in the third keyframe belongs to the content types in rule 1: "screenshot image, screenshot voice, blank voice"; the risk label = "prohibited: bad artist: bad star" in the third keyframe does not belong to the risk labels in rule 1: "advertisement: marketing advertisement: recruitment advertisement", "face: no face: no face"; then the third keyframe is not the target frame and is not counted;

[0186] ④For the 4th key frame:

[0187] The content type = "blank audio" in the 4th keyframe belongs to one of the content types in rule 1: "framed picture, framed voice, blank voice"; the risk label = "prohibited: bad artist: bad star" in the 4th keyframe does not belong to the risk labels in rule 1: "advertisement: marketing advertisement: recruitment advertisement", "face: no face: no face"; then the 4th keyframe is not the target frame, and the 4th keyframe is not counted;

[0188] ⑤For the 5th key frame:

[0189] The content type = "Screenshot" in the 5th keyframe belongs to one of the content types in rule 1: "Screenshot, Screenshot Voice, Blank Voice"; the risk label in the 5th keyframe = None, which does not belong to the risk labels in rule 1: "Advertisement: Marketing Advertisement: Recruitment Advertisement", "Face: No Face: No Face"; therefore, the 5th keyframe is not a target frame and is not counted;

[0190] ⑥For the 6th key frame:

[0191] The content type = "Screenshot" in the 6th keyframe belongs to one of the content types in rule 1: "Screenshot, Speech, Blank Speech"; the risk label in the 6th keyframe = None, which does not belong to the risk labels in rule 1: "Advertisement: Marketing Advertisement: Recruitment Advertisement", "Face: No Face: No Face"; therefore, the 6th keyframe is not a target frame and is not counted;

[0192] The counting result according to rule 2 is: counting key 21 = label + rule 2 + video data ID + quantity 1, counting key 22 = label + rule 2 + video data ID + quantity 1.

[0193] Step 6: determine whether the violation is a hit according to the violation counting result, and perform corresponding treatment on the hit video data; and review the video data through the comment data;

[0194] In this embodiment, step 6 specifically includes:

[0195] Step 61: determine whether the final counting result is greater than the threshold value N. If yes, proceed to step 62; if not, do not process;

[0196] Step 62: Process the video and / or user according to the treatment type and hit tag in the risk identification rule:

[0197] If the disposal type is observation, the video data is pushed to relevant personnel using websocket communication, and the priority of the video data is increased for priority processing, and the processing results are recorded;

[0198] If the disposal type is disposal, the video data will be deleted and the user will be punished;

[0199] If the hit tag is warning, the user is warned;

[0200] If the hit tag is an advertising ticket, an advertising ticket will be issued to the user;

[0201] For example, if the threshold N is set to 1, then the count number for rule 1 is 2, which meets the conditions, and the action type is executed: observation and the hit label is: warning; and the count is reset to zero after the hit;

[0202] The counting result for rule 1 is: count key21 = label + rule 2 + video data ID + quantity 1, count key22 = label + rule 2 + video data ID + quantity 1, each quantity is 1. If the conditions are not met, the disposal type: disposal and hit label: advertising ticket will not be executed; and the count will not be cleared after hitting.

[0203] Step 63. After executing a disposal, warning or advertising ticket, record the penalty time and store it in the penalty queue; when the next hit occurs, record the current hit time and store it in the penalty queue; in the penalty queue, determine whether the current hit time and the previous penalty time are within the set time period. If not, execute the corresponding penalty; if so, do not repeat the penalty, and remove the current hit time from the penalty queue; adopt a penalty deduplication mechanism to solve the problem of repeated hits in multiple rule matching during the review. This mechanism ensures a reasonable interval between the processing flow and enhances the stability and fairness of the system.

[0204] Step 64: Check whether there is user feedback and reporting content based on the comment content under the video data. If yes, review and process the published video data; if not, do not process it.

[0205] Step 7: Label and grade the video data based on the processing and review results to ensure that users of different ages and locations see appropriate content;

[0206] Step 8: Optimize the recommendation algorithm by analyzing user behavior data and combining labeling and grading results to avoid recommending risky content to users.

[0207] like Figure 2 As shown, an embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned multi-level video risk mining method when executing the program.

[0208] like Figure 3 As shown, an embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the above-mentioned multi-level video risk mining method is implemented.

[0209] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0210] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to perform all or part of the steps of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0211] The above descriptions are only some embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any equivalent device or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A multi-level video risk mining method, characterized in that: The steps include: Step 1: Obtain video data, environment data, and comment data; Step 2: extract key frames from the video data and perform risk detection to obtain detection information of each key frame; Step 3: Obtain user features, and clean each key frame according to the user features, environmental data, and detection information to obtain key frames in a standardized format; Step 4: Pre-configure risk identification rules; Step 5: Match the key frames in the standardized format with the risk identification rules and count the violations based on the matching results; specifically, the following steps are included: Step 51, selecting the first key frame in the standardized format as the current key frame according to the sequence of frame capture time; Step 52: Match the current key frame with the content type, risk label, risk result and user data rules in the risk identification rules one by one. If the match is successful, the current key frame is used as the target frame; if the match fails, proceed to step 54; step 52 specifically includes: Step 521, matching the content type in the current key frame with the content type in the risk identification rule, and determining whether the content type in the current key frame matches any one of the captured image, captured voice, and empty voice. If so, proceed to step 522; if not, abandon the current key frame, select the next key frame as the current key frame, and return to step 521; Step 522: match the risk label in the current key frame with the risk label in the risk identification rule to determine whether the risk label in the current key frame meets any of the preset label contents. If so, proceed to step 523; if not, abandon the current key frame, select the next key frame as the current key frame, and return to step 521; Step 523, matching the risk result in the current key frame with the risk result in the risk identification rule, and determining whether the risk result in the current key frame meets any one of rejection and suspicion. If so, proceed to step 524; if not, abandon the current key frame, select the next key frame as the current key frame, and return to step 521; Step 524: determine whether there is a user data rule in the risk identification rule. If not, proceed to step 525. If yes, determine the limited content contained in the user data rule, and match the limited content in the user data rule with the corresponding content in the current key frame one by one. If they are completely matched, proceed to step 525. If they are not completely matched, abandon the current key frame, select the next key frame as the current key frame, and return to step 521. Step 525, taking the current key frame as the target frame and proceeding to step 531; Step 53: Count the target frame according to the counting method and counting type in the risk identification rule; the step 53 specifically includes: Step 531: Determine the counting rule according to the counting method and counting type configured in the risk identification rule, including four combinations, namely, rule continuity, rule accumulation, label continuity and label accumulation; Step 532: if the adopted counting rule is rule continuity, then generate a corresponding counting key1, wherein the counting key1 includes the counting method as rule, the risk identification rule ID and the video data ID, and the number of the counting key1 is 1; determine whether the current target frame is the first target frame, if so, then count the corresponding position of the counting key1 as 1; if not, determine whether the current target frame is continuous with the previous target frame according to the frame capture time and the set time interval, if so, perform incremental counting at the corresponding position of the counting key1, if not, reset the count and count again; Step 533: if the adopted counting rule is rule accumulation, a corresponding counting key1 is generated, wherein the counting key1 includes the counting mode as rule, the risk identification rule ID and the video data ID, and it is determined whether the current target frame matches the counting key1, and the number of the counting key1 is 1; it is determined whether the current target frame is the first target frame, and if so, the corresponding position of the counting key1 is counted as 1; if not, an incremental count is performed at the corresponding position of the counting key1; Step 534: if the adopted counting rule is label continuity, then generate a corresponding counting key2 according to different label contents in the risk label in the risk identification rule, the counting key2 includes a counting method of label, a risk identification rule ID, a video data ID and a label content, and the number of the counting key2 is consistent with the number of label contents preset in the risk label in the risk identification rule; match a corresponding counting key2 according to the risk label in the current target frame, and judge whether the current target frame is the first target frame on the paired counting key2, if yes, then count as 1 at the corresponding position of the paired counting key2; if not, judge whether the current target frame is continuous with the previous target frame in the paired counting key2 according to the frame cut time and the set time interval, if yes, then perform incremental counting at the corresponding position of the paired counting key2, if not, reset the count and count again; Step 535: if the adopted counting rule is label accumulation, then generate corresponding counting key2 according to different label contents in the risk label in the risk identification rule, the counting key2 includes a counting method of label, a risk identification rule ID, a video data ID and a label content, and the number of the counting key2 is consistent with the number of label contents preset in the risk label in the risk identification rule; match a corresponding counting key2 according to the risk label in the current target frame, and determine whether the current target frame is the first target frame on the paired counting key2, if yes, then count the corresponding position of the paired counting key2 as 1; if not, perform incremental counting at the corresponding position of the paired counting key2; Step 54, select the next key frame as the current key frame, repeat steps 52-53, and complete the counting of all target frames; Step 6: determine whether the violation is a hit according to the violation counting result, and perform corresponding treatment on the hit video data; and review the video data through the comment data; Step 7: Label and grade the video data based on the processing and review results.

2. A multi-level video risk mining method as claimed in claim 1, characterized in that: The step 1 specifically includes: Step 11: Use RTC and HTTP protocols to establish real-time communication between the acquisition end and multiple live broadcast ends. The acquisition end receives the video data address and collects corresponding video data from different live broadcast ends according to the video data address. The video data is a video stream or a video file. Step 12: Obtain environmental data through the client, where the environmental data includes the user's location information; Step 13: Collect the comment content corresponding to the video data and push it into the message list.

3. A multi-level video risk mining method as claimed in claim 2, characterized in that: The step 2 specifically includes: Step 21, converting video data in different formats into video data in a video coding format; Step 22: inputting the video data after format conversion into a risk detection model; Step 23, the risk detection model divides the video data into multiple segments according to a set time interval based on a computer vision algorithm, and the set time interval ranges from 1 to 5 seconds, and determines whether the image difference between the first video frame of the current segment and the first video frame of the next segment is within the allowable difference range. If so, the first video frame of the current segment is used as a key frame, if not, each video frame of the current segment is used as a key frame, and for the last segment, each video frame in the last segment is used as a key frame; each key frame includes a picture and / or corresponding audio; Step 24: the risk detection model identifies multiple groups of texts from multiple key frames based on speech recognition technology and OCR text recognition software, where one key frame corresponds to one group of texts; Step 25: The risk detection model matches each group of texts with pre-stored sensitive words, detects whether there is a risk through the matching results, and returns the detection information corresponding to each key frame; the detection information includes the frame capture time, risk result, risk label, identified text, violation status and content type.

4. A multi-level video risk mining method as claimed in claim 3, characterized in that: The step 3 specifically includes: Step 31: Obtain user characteristics from a user database, wherein the user characteristics include gender, user identity, and charm value; Step 32, setting a standardized format to sequentially concatenate user features, user location information, and detection information; Step 33, according to the set standardized format, the user features, user location information and specific contents of the detection information corresponding to each key frame are sequentially spliced ​​to obtain each key frame with a standardized format. The standardized format of each key frame is specifically: gender-user identity-charm value-user location information-frame capture time-risk result-risk label-identified text-violation status-content type.

5. A multi-level video risk mining method as claimed in claim 4, characterized in that: The step 4 specifically includes: Step 41: pre-configure risk identification rules, including rule content, user data rules, and hit results; wherein: The rule content includes: content type, risk label, risk result, counting method and counting type; the content type includes framed picture, framed voice and empty voice, the risk label includes at least one preset label content, and the risk result includes rejection and suspected; the counting method includes label or rule, and the counting type includes continuous or cumulative and threshold N; The user data rules include gender, user identity, charm value, user location information and violation status; the user data rules are to limit or not limit the specific content of gender, user identity, charm value, user location information and violation status according to the needs of the user; The hit result includes a treatment type, a hit tag, and a reset count; the treatment type includes observation or treatment, the hit tag includes warning or advertising ticket, and the reset count includes clearing the count after the hit or not clearing the count after the hit; Step 42: Upload the risk identification rules as configuration information to a configuration center; Step 43: When the configuration information changes, the configuration center publishes the new configuration information, and multiple servers subscribe to and update the new configuration information according to the published information.

6. A multi-level video risk mining method as claimed in claim 1, characterized in that: The step 6 specifically includes: Step 61: determine whether the final counting result is greater than the threshold value N. If yes, proceed to step 62; if not, do not process; Step 62: Process the video and / or user according to the treatment type and hit tag in the risk identification rule: If the disposal type is observation, the video data is pushed to relevant personnel using websocket communication, and the priority of the video data is increased for priority processing, and the processing results are recorded; If the disposal type is disposal, the video data will be deleted and the user will be punished; If the hit tag is warning, the user is warned; If the hit tag is an advertising ticket, an advertising ticket will be issued to the user; Step 63: After executing a disposal, warning or advertising ticket, record the penalty time and store it in the penalty queue; when the next hit occurs, record the current hit time and store it in the penalty queue; in the penalty queue, determine whether the current hit time and the previous penalty time are within the set time period, if not, execute the corresponding penalty; if yes, do not repeat the penalty, and remove the current hit time from the penalty queue; Step 64: Check whether there is user feedback and reporting content based on the comment content under the video data. If yes, review and process the published video data; if not, do not process it.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, a multi-level video risk mining method as described in any one of claims 1 to 6 is implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, a multi-level video risk mining method as described in any one of claims 1 to 6 is implemented.

Images