WO2022143541A1 - 直播审核的方法、装置、服务器和存储介质 - Google Patents

直播审核的方法、装置、服务器和存储介质 Download PDF

Info

Publication number
WO2022143541A1
WO2022143541A1 PCT/CN2021/141704 CN2021141704W WO2022143541A1 WO 2022143541 A1 WO2022143541 A1 WO 2022143541A1 CN 2021141704 W CN2021141704 W CN 2021141704W WO 2022143541 A1 WO2022143541 A1 WO 2022143541A1
Authority
WO
WIPO (PCT)
Prior art keywords
violation
live
live broadcast
frame
review
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2021/141704
Other languages
English (en)
French (fr)
Inventor
李益永
孙准
黄秋实
井雪
项伟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Bigo Technology Pte Ltd
Original Assignee
Bigo Technology Pte Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Bigo Technology Pte Ltd filed Critical Bigo Technology Pte Ltd
Priority to US18/259,646 priority Critical patent/US12010358B2/en
Priority to EP21914277.5A priority patent/EP4274236A4/en
Priority to JP2023540184A priority patent/JP7584054B2/ja
Publication of WO2022143541A1 publication Critical patent/WO2022143541A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
    • H04N21/21—Server components or server architectures
    • H04N21/218—Source of audio or video content, e.g. local disk arrays
    • H04N21/2187—Live feed
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/20—Movements or behaviour, e.g. gesture recognition
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
    • H04N21/23—Processing of content or additional data; Elementary server operations; Server middleware
    • H04N21/234—Processing of video elementary streams, e.g. splicing of video streams or manipulating encoded video stream scene graphs
    • H04N21/23418—Processing of video elementary streams, e.g. splicing of video streams or manipulating encoded video stream scene graphs involving operations for analysing video streams, e.g. detecting features or characteristics
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
    • H04N21/454—Content or additional data filtering, e.g. blocking advertisements
    • H04N21/4542—Blocking scenes or portions of the received content, e.g. censoring scenes

Definitions

  • the present application relates to the field of illegal content review in Internet resources, for example, a method, device, server and storage medium for live broadcast review.
  • a neural network model trained with a single feature is used to analyze the illegal content of the live visual features embodied in multiple live video frames and the live audio features under the live voice during the live broadcast process, and determine whether there are illegal pictures in the multiple live video frames. , or to determine whether there is any illegal audio involving topics such as terror, violence, pornography or political sensitivity in the live broadcast, so as to review the suspected illegal live broadcast room and push it to the manual review platform for manual review; at this time, the live broadcast is extracted during the live broadcast process.
  • the extraction results may be inaccurate due to the instability of the live broadcast network. Therefore, when using the extracted live video frames and live voices to check whether there is any illegal content in the live broadcast room, the live broadcast room cannot be guaranteed.
  • the accuracy and recall rate of violation audits can easily lead to misjudgments or missed judgments.
  • the present application provides a method, device, server and storage medium for live broadcast review, which avoids erroneous review and missed review under live broadcast review, and improves the accuracy and recall rate of live broadcast review.
  • a method of live auditing which includes:
  • the first violation score of the current live frame under the high-accuracy review model, the second violation score under the high-recall review model, and the The multi-dimensional behavioral features in the live broadcast room are reviewed and input into a pre-built behavioral review model to obtain the target violation score of the current live broadcast frame.
  • an apparatus for live auditing comprising:
  • the preliminary audit module is set to conduct preliminary violation audit of the current live frames in the live broadcast room to be audited through the cascaded high-precision audit model and high-recall audit model;
  • the violation score determination module is set to, if the current live frame passes the preliminary violation review, the first violation score of the current live frame under the high-precision review model and the second violation score of the current live frame under the high-recall review model.
  • the violation score and the multi-dimensional behavior characteristics in the live broadcast room to be reviewed are input into a pre-built behavior audit model to obtain the target violation score of the current live broadcast frame.
  • a server that includes:
  • processors one or more processors
  • storage means arranged to store one or more programs
  • the one or more processors implement the method for live auditing described in any embodiment of the present application.
  • a computer-readable storage medium is also provided, storing a computer program, and when the computer program is executed by a processor, the method for live auditing described in any embodiment of the present application is implemented.
  • FIG. 1A is a flowchart of a method for live auditing provided in Embodiment 1 of the present application;
  • FIG. 1B is a schematic diagram of the principle of the live broadcast review process provided in Embodiment 1 of the present application.
  • FIG. 2A is a flowchart of a method for live auditing provided in Embodiment 2 of the present application.
  • FIG. 2B is a schematic diagram of the principle of the live broadcast review process provided in Embodiment 2 of the present application.
  • FIG. 3A is a flowchart of a method for live auditing provided in Embodiment 3 of the present application.
  • FIG. 3B is a schematic diagram of the principle of the live broadcast review process provided in Embodiment 3 of the present application.
  • FIG. 4 is a schematic structural diagram of a device for live auditing provided in Embodiment 4 of the present application.
  • FIG. 5 is a schematic structural diagram of a server according to Embodiment 5 of the present application.
  • FIG. 1A is a flowchart of a method for live broadcast review provided in Embodiment 1 of the present application, and this embodiment can be applied to a situation in which a live broadcast screen in any live broadcast room is subjected to violation review.
  • the method for live auditing provided in this embodiment may be performed by the apparatus for live auditing provided in this embodiment of the present application.
  • the device may be implemented in software and/or hardware, and is integrated in a server that executes the method.
  • the server may be a background server carrying video review capabilities.
  • the method may include the following steps.
  • the live broadcast room is aimed at a wide range of audience types, if the live broadcast content of some live broadcast rooms involves topics such as terror, violence, pornography or political sensitivity, the illegal content will be quickly spread on the Internet. Therefore, in order to prevent the illegal content from affecting users The adverse impact of life needs to be detected in real time during the live broadcast process to see if there is any illegal live broadcast content.
  • the entire review process under the live broadcast review will be divided into two parts: a preliminary violation review and a second violation review.
  • a preliminary violation review the more obvious violation live broadcasts that are easy to distinguish whether the violation is in violation can be quickly reviewed
  • second violation review process a higher review intensity than the initial violation review is adopted.
  • Violation review cannot accurately determine whether the live frames with illegal content are reviewed for violations again, so as to avoid the problems of false review and omission of review under the live review, so as to ensure the comprehensive violation review of the current live frames in the live broadcast room to be reviewed.
  • the high-precision review model has high-accuracy violation review capabilities.
  • the high recall audit model has high recall violations Audit capability, when using the high recall audit model to audit the current live frames in the live broadcast room for violations, it can ensure as much as possible that even if the current live frame contains very little illegal content, it can still be audited, reducing the occurrence of violations and missed audits. happening.
  • each training sample in the training sample set can be divided into four categories: easy positive samples, easy negative samples, difficult positive samples, and difficult negative samples.
  • Easy positive samples and easy negative samples refer to positive samples and negative samples that are easy to be identified by the model for violations
  • difficult positive samples and difficult negative samples refer to positive samples and negative samples that are difficult to be identified by the model for violations.
  • the training samples of the high-precision audit model can be composed of easy-positive samples and easy-negative samples in the training sample set. There are violations, and the model parameters and network structure of the high-precision audit model are continuously optimized; and because the high-recall audit model has high requirements for the recall of violation audits, the training samples of the high-recall audit model can be set from the training sample set. It is composed of hard positive samples and hard negative samples.
  • the high recall review model can also review the illegal live frames that are difficult to review, thereby ensuring high recall.
  • the audit model has high recall capability for non-compliance audits.
  • this embodiment can cascade the high-precision audit model and the high-recall audit model, as shown in FIG.
  • Each current live frame in the room undergoes an initial violation review, and for the current live frame that has not been reviewed by the high-precision review model for violations, the high-recall review model continues to conduct violation review, using a cascaded two-layer review model.
  • Violation audits are performed in sequence to complete the preliminary violation audit process in this embodiment, so as to ensure high accuracy and high recall of the preliminary violation audits.
  • a live broadcast picture may be intercepted from the live broadcast room to be reviewed every preset time length (such as every 2s, etc.), and used as the current live broadcast frame in this embodiment, the The current live frame is input into the high-accuracy audit model in the two-layer cascaded network, and the high-precision audit model determines whether there is any illegal content in the current live frame.
  • the high-accuracy audit model believes that there is no illegal content in the current live frame, and because the high-accuracy audit model has violations and missed audits, it is necessary to input the current live frame into the high-recall audit model in the two-layer cascaded network again. , the high-recall audit model continues to determine whether there is any illegal content in the current live broadcast frame, so as to complete the preliminary illegal audit of the current live broadcast frame in the live broadcast room to be audited.
  • the high-recall audit model fails the illegal audit, indicating that the high-recall audit model can determine that there is illegal content in the current live frame, Directly confirm that the live broadcast room to be reviewed where the current live frame is located is the illegal live broadcast room; 2) Pass the violation audit of the high recall audit model, indicating that the high recall audit model believes that there is no illegal content in the current live broadcast frame, then the follow-up need to use the advance
  • the constructed behavior auditing model continues to conduct violation auditing of the current live frame.
  • the current live frame in the live broadcast room to be reviewed passes both the high-precision review model and the high-recall review model, it means that the current live frame has passed the preliminary violation review and needs to be reviewed again. , to ensure the accuracy and comprehensiveness of the violation audit.
  • the high-precision audit model will output the current The first violation score of the live frame
  • the high recall audit model will output the second violation score of the current live frame, and both the first violation score and the second violation score can represent the visual characteristics of the current live frame.
  • the live broadcast room will be used as a reference for judging whether the current live broadcast frame has any illegal content. Therefore, in this embodiment, after it is determined that the current live broadcast frame has passed the preliminary violation review, when the current live broadcast frame is reviewed for violations again, in addition to referring to the current live broadcast frame.
  • the visual picture features represented by the first violation score of the frame under the high-accuracy review model and the second violation score under the high-recall review model it also analyzes the live broadcast moment of the current live frame in the live broadcast room to be reviewed. Then, the multi-dimensional behavior characteristics of the current live frame in the live broadcast room to be reviewed are determined, as shown in Figure 1B, and the first violation of the current live frame.
  • the visual picture features represented by the score and the second violation score are combined with the multi-dimensional behavioral features determined by multiple live broadcast interactive behaviors in the live broadcast room to be reviewed, and are jointly input into the pre-built behavioral auditing model.
  • the behavioral auditing model can Integrate and analyze the visual picture features represented by the first violation score and the second violation score and the multi-dimensional behavior features determined by multiple live broadcast interactive behaviors in the live broadcast room to be reviewed, and uniformly determine whether there is any illegal content in the current live broadcast frame. , so as to output the target violation score of the current live frame.
  • the target violation score adopts the multi-angle feature fusion analysis combined with the multi-dimensional behavior features of the visual picture features to realize the re-violation review without using additional review equipment. Violation audit costs, to ensure the accuracy and recall of violation audits in the live broadcast room.
  • the technical solution provided by this embodiment through the cascaded high-precision audit model and high-recall audit model, conducts preliminary violation audit on the current live frames in the live broadcast room to be audited, and can quickly audit the illegal live frames that are easy to distinguish and have obvious violations.
  • the first violation score of the current live frame under the high-precision review model and the second violation score under the high-recall review model will be based on the current live frame.
  • the behavior audit model is used to perform a unified feature analysis of the current live frame combining visual picture characteristics and behavior characteristics, so as to obtain the current live frame.
  • the target violation score to achieve a comprehensive violation review of the current live broadcast frame in the live broadcast room to be reviewed, to avoid the problems of erroneous review and missed review under the live broadcast review, and to improve the accuracy and recall rate of live broadcast review on the basis of reducing the cost of violation review.
  • FIG. 2A is a flowchart of a method for live broadcast auditing provided by Embodiment 2 of the present application
  • FIG. 2B is a schematic schematic diagram of the principle of a live broadcast audit process provided by Embodiment 2 of the present application. This embodiment is described on the basis of the above-mentioned embodiment. As shown in FIG. 2A , in this embodiment, the review process of the re-violation review of the current live frame in the live broadcast room to be reviewed and the modeling process of the behavior review model are explained.
  • this embodiment may include the following steps.
  • the current live frame passes the preliminary violation review, it means that both the high-accuracy review model and the high-recall review model believe that there is no illegal content in the current live frame. Therefore, in order to ensure the accuracy of the live broadcast review and avoid the occurrence of violations and missed reviews , the current live frame will continue to be reviewed for violations again by using the visual image features and behavioral characteristics of the current live frame. At this time, the first violation score of the current live frame under the high-precision review model and the high-recall review will be determined.
  • the second violation score under the model takes the first violation score and the second violation score as the visual image features corresponding to the current live frame, and at the same time, according to the live broadcast time of the current live frame, a number of live interactive behaviors in the live broadcast room to be reviewed , determine the multi-dimensional behavior characteristics of the current live frame in the live broadcast room to be reviewed, and combine the first violation score, the second violation score and the multi-dimensional behavior characteristics to obtain the live broadcast behavior characteristics of the current live frame, as the multi-dimensional behavior under the review of the second violation.
  • the angle fusion feature, and the subsequent use of the live broadcast behavior feature to determine whether there is any illegal content in the current live broadcast frame can ensure the comprehensiveness and accuracy of the live broadcast review.
  • the number of live broadcast rewards in the room x8 is the number of historical violations of the host in the live broadcast room to be reviewed in the multi-dimensional behavior feature
  • x9 is the historical violation ratio of the host in the live broadcast room to be reviewed in the multi-dimensional behavior feature
  • x10 is the multi-dimensional behavior feature.
  • the number of historical voice violations of the anchor in the live broadcast room to be reviewed x11 is the proportion of historical voice violations of the anchor in the live broadcast room to be reviewed in the multi-dimensional behavior characteristics.
  • the multi-dimensional behavior feature in this embodiment may be dynamically configured, which is not limited.
  • S230 Input the live broadcast behavior feature into the behavior audit model to obtain the target violation score of the current live broadcast frame.
  • the live broadcast behavior feature can be used as a multi-angle violation fusion feature and input into a pre-built behavior audit model, and the live broadcast behavior feature can be displayed in each dimension through the behavior audit model.
  • the eigenvalues below are used for fusion analysis to determine whether there is illegal content information, and then output the target violation score of the current live frame. Based on the target violation score, it is judged whether there is any illegal content in the current live frame, which improves the accuracy of live broadcast auditing.
  • the behavior auditing model may be modeled by adopting the following steps: extracting corresponding historical live frames from each historical live video in the historical live video set, and determining where each historical live frame is located. The multi-dimensional historical behavior characteristics in the live broadcast room and the violation label of the historical live broadcast frame; for each historical live broadcast frame, the first historical violation score of the historical live broadcast frame under the high-accuracy audit model, and the first history of the live broadcast frame under the high-recall audit model. 2.
  • the historical violation score and the multi-dimensional historical behavior characteristics in the live broadcast room are input into the pre-built behavior audit model to obtain the historical violation score of the historical live broadcast frame; in order to minimize the historical violation score and violation of each historical live broadcast frame
  • the differences between the tags are targeted, and the behavioral audit model is continuously optimized.
  • a historical live video set for modeling is constructed. Since the violation review of the live room is mainly performed on the live frames during the live broadcast, it is necessary to review the history of the live broadcast. Video frame extraction is performed for each historical live video in the live video set, for example, a historical live picture is captured from each historical live video every preset time length as the historical live frame extracted from the historical live video. , all the historical live frames extracted at this time can constitute the modeling samples of the behavior audit model.
  • the first historical violation score, x2 is the second historical violation score of the historical live frame under the high-recall audit model, x3 is the number of live live viewing of the historical live frame in the live broadcast room in the set multi-dimensional historical behavior feature, x4 It is the number of public screens sent to users in the live broadcast room where the historical live frames in the multi-dimensional historical behavior feature are located, x5 is the number of illegal public screens in the live broadcast room where the historical live frames in the multi-dimensional historical behavior feature are located, and x6 is the history in the multi-dimensional historical behavior feature.
  • the proportion of the illegal public screens in the live broadcast room where the live broadcast frame is in all public screens x7 is the number of live broadcast rewards of the historical live broadcast frame in the live broadcast room in the multi-dimensional historical behavior feature, x8 is the historical live broadcast frame in the multi-dimensional historical behavior feature
  • the number of historical violations of the host in the live broadcast room x9 is the historical violation ratio of the historical live frame in the live broadcast room in the multi-dimensional historical behavior feature, and x10
  • the historical live broadcast frame in the multi-dimensional historical behavior feature is the host in the live broadcast room.
  • the number of historical voice violations, x11 is the proportion of historical voice violations of the anchor in the live broadcast room of the historical live frame in the multi-dimensional historical behavior feature.
  • the multi-dimensional historical behavior characteristics of the historical live frames in the multiple modeling features in the live studio can be determined during the historical live broadcast, after extracting the corresponding historical live frames from each historical live video, By judging the historical live interactive behavior of historical live frames in the live room, the multi-dimensional historical behavior characteristics of each historical live frame in the live room can be determined, and the violation label of each historical live frame can also be determined. , as shown in Figure 2B, so that the subsequent modeling optimization can be carried out continuously according to the modeling optimization objective.
  • the historical live broadcast frame can be input into the high-precision audit model and the high-recall audit model in this embodiment respectively, so as to obtain the first historical violation score and the second historical violation score of each historical live broadcast frame , so that the first historical violation score, the second historical violation score and the multi-dimensional historical behavior characteristics in the live broadcast room of each historical live broadcast frame can be combined to form the modeling feature corresponding to each historical live broadcast frame, and
  • the modeling features corresponding to each historical live broadcast frame are continuously input into the behavior audit model, and the historical violation score of each historical live broadcast frame is obtained.
  • the modeling optimization goal of the behavior audit model in this embodiment may be to minimize the difference between the historical violation score and the violation label of each historical live broadcast frame. Therefore, in the modeling optimization process, each The difference between the historical violation scores of the historical live frames and the violation labels of each historical live frame, and the difference is continuously reduced by optimizing the mathematical parameters in the behavior audit model to achieve the historical violation scores and violation labels of the historical live frames. The difference between them can be minimized, so as to obtain the final optimized behavior audit model. At this time, the final optimized behavior audit model has the ability to audit violations with high accuracy.
  • modeling optimization objective of the behavior audit model in this embodiment may be:
  • x i is the modeling feature of the ith historical live frame
  • yi is the violation label of the ith historical live frame
  • V (x 1 , x 2 , L, x m ), representing the modeling samples composed of historical live frames
  • m is the number of samples
  • p( xi ) represents the predicted probability of xi obtained by the behavior audit model.
  • the technical solution provided by this embodiment takes minimizing the difference between the historical violation score and the violation label of each historical live frame as the modeling optimization goal, and models the behavior audit model to ensure the accuracy of the behavior audit model in the violation audit. Afterwards, by combining the first violation score of the current live frame in the live broadcast room to be reviewed under the high-accuracy audit model, the second violation score under the high recall audit model, and the multi-dimensional behavior characteristics of the live broadcast room to be reviewed, the current The live broadcast behavior characteristics of the live broadcast frame, and the behavioral audit model after modeling is used to analyze the violation characteristics of the live broadcast behavior characteristics from multiple angles, so as to improve the accuracy and recall rate of the live broadcast audit.
  • FIG. 3A is a flowchart of a method for live broadcast auditing provided by Embodiment 3 of the present application
  • FIG. 3B is a schematic schematic diagram of a principle of a live broadcast audit process provided by Embodiment 3 of the present application. This embodiment is described on the basis of the above-mentioned embodiment. As shown in FIG. 3A , in this embodiment, the overall violation review process in the live broadcast room to be reviewed is explained.
  • this embodiment may include the following steps.
  • S310 Input the current live frame in the live broadcast room to be reviewed into a pre-built high-accuracy review model to obtain a first violation score of the current live frame.
  • the live broadcast room to be audited when the live broadcast room to be audited is subjected to a violation review, it will be detected in real time during the live broadcast whether there is any illegal content on the live broadcast screen. Therefore, it is necessary to conduct a violation review on the current live broadcast frame during the live broadcast process in real time. Input the current live frame into a pre-built high-accuracy review model, and perform an initial violation review of the live screen in the current live frame through the high-precision review model, thereby outputting the first violation score of the current live frame, and then pass the judgment. Whether the first violation score exceeds the preset precise violation threshold can determine whether there is any violation content in the current live broadcast frame.
  • S320 determine whether the first violation score exceeds the preset precise violation threshold, if yes, go to S380; if not, go to S330.
  • the high-accuracy audit model believes that there is no illegal content in the current live frame.
  • the live frame continues to be input into the pre-built high-recall audit model, as shown in Figure 3B, the high-recall audit model continues to conduct corresponding violation audits on the live pictures in the current live frame, thereby outputting the second image of the current live frame.
  • Violation score and subsequently, by judging whether the second violation score exceeds the preset recall violation threshold, it can be determined again whether there is any illegal content in the current live broadcast frame.
  • S340 determine whether the second violation score exceeds the preset recall violation threshold, if yes, go to S380; if not, go to S350.
  • the high-recall audit model also believes that there is no illegal content in the current live frame, that is, the current live frame has passed both the high-precision audit model and the high-recall audit model. That is, through the preliminary violation review, the behavior review model is used to continue to conduct another violation review of the current live frame.
  • the current live frame has passed the preliminary violation review that the first violation score of the current live frame does not exceed the preset precise violation threshold, and the second violation score of the current live frame does not exceed the preset recall violation threshold.
  • S350 Input the first violation score of the current live frame under the high-precision audit model, the second violation score under the high-recall audit model, and the multi-dimensional behavior characteristics in the live broadcast room to be audited into a pre-built behavior audit In the model, the target violation score of the current live frame is obtained.
  • S360 determine whether the target violation score exceeds the preset behavior violation threshold, if yes, go to S380; if not, go to S370.
  • the behavior audit model After using the behavior audit model to output the target violation score of the current live frame, by judging whether the target violation score exceeds the preset behavior violation threshold, it is possible to finally verify whether there is any illegal content in the current live frame, so as to execute different live broadcasts. Process flow.
  • the target violation score does not exceed the preset behavior violation threshold, it means that the behavior review model also believes that there is no violation content in the current live broadcast frame.
  • the three models in this embodiment all believe that there is no violation in the current live broadcast frame. content, it can be determined that the current live broadcast frame is a non-violating frame, indicating that there is no illegal content in the live broadcast room to be reviewed.
  • the live broadcast duration will be preset at intervals (for example, every 2s) After that, continue to intercept a new live broadcast picture from the current live broadcast moment in the live broadcast room to be reviewed, as the new current live broadcast frame collected, and continue to carry out the new current live broadcast frame according to the steps of S310-S380 of the present embodiment. Violation review, so as to conduct real-time violation review in the live broadcast room to be reviewed, until the live broadcast ends or the illegal live frame is reviewed.
  • this embodiment will additionally set up a manual review platform, so that when the current live broadcast frame contains illegal content, the The live broadcast information of the live broadcast room to be reviewed where the current live broadcast frame is located is pushed to the manual review platform to further manually review the live broadcast content of the live broadcast room to be reviewed, so as to ensure the accuracy of the live broadcast review.
  • the live broadcast information of the live broadcast room to be reviewed may be the live broadcast address of the live broadcast room to be reviewed or the video image being broadcasted, etc., which is not limited.
  • S390 Monitor the live broadcast interaction items in the live broadcast room to be reviewed in real time. If the live broadcast interaction item exceeds the preset interaction threshold, push the live broadcast information of the live broadcast room to be reviewed to the manual review platform, and stop the current live broadcast frame in the live broadcast room to be reviewed. Violation review.
  • this embodiment on the basis of building a model to realize live broadcast violation review, also monitors the live broadcast in the live broadcast room to be reviewed in real time.
  • Interactive items to determine whether it needs to be directly pushed to the manual review platform for manual review, and use the manual review as the minimum audit plan for live review.
  • the live broadcast interaction item can include the number of live broadcast viewings and live broadcast rewards in the live broadcast room to be reviewed, etc., so as to analyze the audience breadth of the live broadcast room to be reviewed. If the live broadcast interaction item exceeds the preset interaction threshold, it means that it is pending review.
  • the live broadcast audience rate of the live broadcast room is very large.
  • the technical solution provided in this embodiment is to directly push the live broadcast information of the live broadcast room to be reviewed to the manual review platform for manual review after it is determined that there is illegal content in the current live broadcast frame, so as to avoid the situation of illegal and erroneous review, and improve the accuracy of live broadcast illegal review. Accuracy; at the same time, use the live broadcast interaction items in the live broadcast room to be reviewed to additionally set up an abnormal minimum guarantee plan for live broadcast audits to ensure the comprehensiveness of live broadcast violation audits.
  • FIG. 4 is a schematic structural diagram of an apparatus for live auditing according to Embodiment 4 of the present application. As shown in Figure 4, the apparatus may include:
  • the preliminary audit module 410 is configured to perform preliminary violation audit on the current live frame in the live broadcast room to be audited through the cascaded high-precision audit model and high-recall audit model;
  • Violation score determination module 420 is set to, if the current live frame passes the preliminary violation review, the first violation score of the current live frame under the high-precision review model, and the first violation score under the high-recall review model of the current live frame.
  • the second violation score and the multi-dimensional behavior characteristics in the live broadcast room to be reviewed are input into a pre-built behavior audit model to obtain the target violation score of the current live broadcast frame.
  • the technical solution provided by this embodiment through the cascaded high-precision audit model and high-recall audit model, conducts preliminary violation audit on the current live frames in the live broadcast room to be audited, and can quickly audit the illegal live frames that are easy to distinguish and have obvious violations.
  • the first violation score of the current live frame under the high-precision review model and the second violation score under the high-recall review model will be based on the current live frame.
  • the behavior audit model is used to perform a unified feature analysis on the current live frame that combines visual picture characteristics and behavior characteristics, so as to obtain the current live frame.
  • the target violation score to achieve a comprehensive violation review of the current live broadcast frame in the live broadcast room to be reviewed, avoid the problems of false review and missed review under the live broadcast review, and improve the accuracy and recall rate of the live broadcast review on the basis of reducing the cost of violation review.
  • the apparatus for live auditing provided in this embodiment can be applied to execute the live auditing method provided in any of the foregoing embodiments, and has corresponding functions and effects.
  • FIG. 5 is a schematic structural diagram of a server according to Embodiment 5 of the present application.
  • the server includes a processor 50, a storage device 51 and a communication device 52; the number of processors 50 in the server may be one or more One processor 50 is taken as an example in FIG. 5 ; the processor 50 , the storage device 51 and the communication device 52 in the server may be connected through a bus or other means, and the connection through a bus is taken as an example in FIG. 5 .
  • the server provided in this embodiment can be used to execute the method for live auditing provided by any of the foregoing embodiments, and has corresponding functions and effects.
  • Embodiment 6 of the present application further provides a computer-readable storage medium on which a computer program is stored.
  • the computer program is executed by a processor, the method for live auditing in any of the foregoing embodiments can be implemented.
  • the method can include:
  • the first violation score of the current live frame under the high-accuracy review model, the second violation score under the high-recall review model, and the The multi-dimensional behavioral features in the live broadcast room are reviewed and input into a pre-built behavioral review model to obtain the target violation score of the current live broadcast frame.
  • a storage medium containing computer-executable instructions provided by an embodiment of the present application the computer-executable instructions of the computer-executable instructions are not limited to the above-mentioned method operations, and can also perform relevant aspects of the live auditing method provided by any embodiment of the present application. operate.
  • the storage medium may be a non-transitory storage medium.
  • the present application can be implemented by means of software and general hardware, and can also be implemented by hardware.
  • the technical solution of the present application can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a floppy disk of a computer, a read-only memory (Read-Only Memory, ROM), a random access memory ( Random Access Memory, RAM), flash memory (FLASH), hard disk or optical disk, etc., including multiple instructions to enable a computer device (which may be a personal computer, server, or network device, etc.) to execute the various embodiments of the present application.
  • a computer-readable storage medium such as a floppy disk of a computer, a read-only memory (Read-Only Memory, ROM), a random access memory ( Random Access Memory, RAM), flash memory (FLASH), hard disk or optical disk, etc.
  • the multiple units and modules included are only divided according to functional logic, but are not limited to the above-mentioned division, as long as the corresponding functions can be realized; in addition, multiple functional units
  • the names are only for the convenience of distinguishing from each other, and are not used to limit the protection scope of this application.

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Signal Processing (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Social Psychology (AREA)
  • Health & Medical Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Psychiatry (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Two-Way Televisions, Distribution Of Moving Picture Or The Like (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Testing, Inspecting, Measuring Of Stereoscopic Televisions And Televisions (AREA)
  • Information Transfer Between Computers (AREA)

Abstract

本申请提供了一种直播审核的方法、装置、服务器和存储介质。其中，该直播审核的方法包括：通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核；如果当前直播帧通过初步违规审核，则将当前直播帧在高精确审核模型下的第一违规得分、在高召回审核模型下的第二违规得分以及在待审核直播间内的多维行为特征，输入到预先构建的行为审核模型中，得到当前直播帧的目标违规得分。

Description

直播审核的方法、装置、服务器和存储介质
本申请要求在2020年12月30日提交中国专利局、申请号为202011613190.6的中国专利申请的优先权，该申请的全部内容通过引用结合在本申请中。
技术领域
本申请涉及互联网资源中违规内容审核领域，例如涉及一种直播审核的方法、装置、服务器和存储介质。
背景技术
随着互联网技术的快速发展，网络上的直播平台也越来越多，使得互联网资源的传播越来越广泛。与此同时，由于一些直播间的直播内容可能会涉及到恐怖、暴力、色情或政治敏感等话题，而使得大量违规的直播内容在互联网上被快速传播出来，因此针对此类直播间，需要实时检测是否存在违规的直播内容，以防止违规直播内容的传播。
通常会利用单特征训练的神经网络模型对直播过程中的多个直播视频帧所体现的直播视觉特征和直播语音下的直播音频特征进行违规内容分析，判断多个直播视频帧中是否存在违规画面，或者判断直播语音中是否存在涉及恐怖、暴力、色情或政治敏感等话题的违规音频，从而审核出可疑的违规直播间，推送给人工审核平台进行人工审核；此时，在直播过程中提取直播视频帧和直播语音时，会存在由于直播网络不稳定而导致提取结果不准确的问题，因此在利用所提取的直播视频帧和直播语音来审核直播间内是否存在违规内容时，无法保证直播间违规审核的准确率和召回率，极易造成误判或漏判的问题。
发明内容
本申请提供了一种直播审核的方法、装置、服务器和存储介质，避免直播审核下存在的误审核和漏审核，提高直播审核的准确率和召回率。
提供了一种直播审核的方法，该方法包括：
通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核；
如果所述当前直播帧通过初步违规审核，则将所述当前直播帧在所述高精确审核模型下的第一违规得分、在所述高召回审核模型下的第二违规得分以及在所述待审核直播间内的多维行为特征，输入到预先构建的行为审核模型中， 得到所述当前直播帧的目标违规得分。
还提供了一种直播审核的装置，该装置包括：
初步审核模块，设置为通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核；
违规得分确定模块，设置为如果所述当前直播帧通过初步违规审核，则将所述当前直播帧在所述高精确审核模型下的第一违规得分、在所述高召回审核模型下的第二违规得分以及在所述待审核直播间内的多维行为特征，输入到预先构建的行为审核模型中，得到所述当前直播帧的目标违规得分。
还提供了一种服务器，该服务器包括：
一个或多个处理器；
存储装置，设置为存储一个或多个程序；
当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现本申请任意实施例所述的直播审核的方法。
还提供了一种计算机可读存储介质，存储有计算机程序，该计算机程序被处理器执行时实现本申请任意实施例所述的直播审核的方法。
附图说明
图1A为本申请实施例一提供的一种直播审核的方法的流程图；
图1B为本申请实施例一提供的直播审核过程的原理示意图；
图2A为本申请实施例二提供的一种直播审核的方法的流程图；
图2B为本申请实施例二提供的直播审核过程的原理示意图；
图3A为本申请实施例三提供的一种直播审核的方法的流程图；
图3B为本申请实施例三提供的直播审核过程的原理示意图；
图4为本申请实施例四提供的一种直播审核的装置的结构示意图；
图5为本申请实施例五提供的一种服务器的结构示意图。
具体实施方式
下面结合附图和实施例对本申请进行说明。此处所描述的实施例仅仅用于 解释本申请，而非对本申请的限定。为了便于描述，附图中仅示出了与本申请相关的部分而非全部结构。此外，在不冲突的情况下，本申请中的实施例及实施例中的特征可以相互组合。
实施例一
图1A为本申请实施例一提供的一种直播审核的方法的流程图，本实施例可适用于对任一个直播间内的直播画面进行违规审核的情况中。本实施例提供的一种直播审核的方法可以由本申请实施例提供的直播审核的装置来执行，该装置可以通过软件和/或硬件的方式来实现，并集成在执行本方法的服务器中，该服务器可以是承载有视频审核能力的后台服务端。
参考图1A，该方法可以包括如下步骤。
S110，通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核。
由于直播间面向的观众类型较为广泛，如果一些直播间的直播内容涉及到恐怖、暴力、色情或政治敏感等话题，那么违规内容就会在互联网中被快速传播起来，因此为了防止违规内容对用户生活产生的不利影响，需要在直播过程中，实时检测是否存在违规的直播内容。
在本实施例中，为了保证直播违规审核的全面性，会将直播审核下的整个审核过程分成初步违规审核和再次违规审核两部分。此时，在初步违规审核过程中，能够快速审核出容易区分是否违规的较为明显的违规直播，而在再次违规审核过程中，采用相比初步违规审核来说更高的审核力度，来对初步违规审核无法准确判断是否存在违规内容的直播帧再次进行违规审核，以避免直播审核下存在误审核和漏审核的问题，从而保证待审核直播间内当前直播帧的全面违规审核。
本实施例在初步违规审核过程中，会预先构建出高精确审核模型和高召回审核模型两个网络模型；其中，高精确审核模型具备高准确性的违规审核能力，在通过高精确审核模型对待审核直播间内的当前直播帧进行违规审核时，能够保证所审核出的违规直播帧极大可能是真实违规的，避免出现违规误审核的问题；同时，高召回审核模型具备高召回性的违规审核能力，在通过高召回审核模型对待审核直播间内的当前直播帧进行违规审核时，能够尽可能保证即使当前直播帧存在极小的违规内容，也能够被审核出来，减少出现违规漏审核的情况。
本实施例在对高精确审核模型和高召回审核模型进行训练时，会预先构建一个训练样本集，该训练样本集中会包括多类直播间内历史直播视频中的历史 直播帧，同时按照每个历史直播帧是否真实违规，以及被准确审核出是否违规的难易程度，可以将训练样本集内的每个训练样本分为易正样本、易负样本、难正样本和难负样本四类，易正样本和易负样本分别是指容易被模型识别出是否违规的正样本和负样本，难正样本和难负样本分别是指难以被模型识别出是否违规的正样本和负样本，此时由于高精确审核模型对于违规审核的精确性具有较高要求，因此该高精确审核模型的训练样本可以由训练样本集中的易正样本和易负样本构成，通过审核易正样本和易负样本是否存在违规内容，不断优化该高精确审核模型的模型参数和网络结构等；而由于高召回审核模型对于违规审核的召回性具有较高要求，因此该高召回审核模型的训练样本可以由训练样本集中的难正样本和难负样本构成，通过采用难以被审核是否违规的训练样本进行审核训练，使得该高召回审核模型面对难以被审核出来的违规直播帧，也能够审核出来，从而保证高召回审核模型对于违规审核的高召回能力。
同时，由于高精确审核模型只针对违规审核精确性进行训练，存在较大的违规漏审核的情况，而高召回审核模型却具备高召回的违规审核能力，能够尽可能的减少违规漏审核的情况，因此为了保证初步违规审核的高准确性和高召回性，本实施例可以将高精确审核模型和高召回审核模型级联起来，如图1B所示，从而先利用高精确审核模型对待审核直播间内的每一个当前直播帧进行初始的违规审核，而对高精确审核模型未审核出是否违规的当前直播帧，则由高召回审核模型继续进行违规审核，以采用级联的两层审核模型依次进行违规审核，完成本实施例中的初步违规审核过程，保证初步违规审核的高准确性和高召回性。
在待审核直播间的视频直播过程中，本实施例可以每隔预设时长(如每隔2s等)从待审核直播间内截取一张直播图片，作为本实施例中的当前直播帧，将该当前直播帧输入到两层级联网络中的高精确审核模型中，由该高精确审核模型判断该当前直播帧内是否存在违规内容，若该当前直播帧通过该高精确审核模型的违规审核，说明该高精确审核模型认为该当前直播帧内不存在违规内容，而由于高精确审核模型存在违规漏审核的情况，因此需要再次将该当前直播帧输入到两层级联网络中的高召回审核模型中，由该高召回审核模型继续判断该当前直播帧内是否存在违规内容，从而完成对待审核直播间内当前直播帧的初步违规审核。
此时，高召回审核模型对于该当前直播帧的违规审核结果存在两种情况：1)未通过高召回审核模型的违规审核，说明该高召回审核模型能够确定该当前直播帧中存在违规内容，直接确认该当前直播帧所在的待审核直播间为违规直播间；2)通过高召回审核模型的违规审核，说明该高召回审核模型认为该当前直播帧内不存在违规内容，则后续需要利用预先构建的行为审核模型继续对该当 前直播帧进行违规审核。
S120，如果当前直播帧通过初步违规审核，则将当前直播帧在高精确审核模型下的第一违规得分、在高召回审核模型下的第二违规得分以及在待审核直播间内的多维行为特征，输入到预先构建的行为审核模型中，得到当前直播帧的目标违规得分。
可选的，如果待审核直播间内的当前直播帧既通过高精确审核模型的违规审核，也通过高召回审核模型的违规审核，则说明当前直播帧通过初步违规审核，需要继续进行再次违规审核，以保证违规审核的准确性和全面性。
通过级联的高精确审核模型和高召回审核模型，对当前直播帧进行初步违规审核时，通常是对当前直播帧的直播视觉画面内的信息进行违规分析，此时高精确审核模型会输出当前直播帧的第一违规得分，而高召回审核模型会输出当前直播帧的第二违规得分，且该第一违规得分和第二违规得分均能够代表当前直播帧的视觉画面特征。
然而，由于直播间内的直播内容具有多变性，且直播间内的主播与观众之间也会针对直播内容进行互动，例如直播打赏、内容公屏评论和主播历史违规情况等，使得直播间内的互动行为也能够作为判断当前直播帧是否存在违规内容的参考依据，因此本实施例在确定当前直播帧通过初步违规审核后，在对当前直播帧进行再次违规审核时，除了可以参考当前直播帧在高精确审核模型下的第一违规得分和在高召回审核模型下的第二违规得分所代表的视觉画面特征外，还会通过分析当前直播帧在待审核直播间内所处的直播时刻下，主播与观众在该待审核直播间内的多项直播互动行为，确定当前直播帧在该待审核直播间内的多维行为特征，如图1B所示，并将当前直播帧的第一违规得分和第二违规得分所代表的视觉画面特征和由待审核直播间内的多项直播互动行为所确定的多维行为特征结合起来，共同输入到预先构建的行为审核模型中，该行为审核模型能够对第一违规得分和第二违规得分所代表的视觉画面特征和由待审核直播间内的多项直播互动行为所确定的多维行为特征进行融合分析，统一判断该当前直播帧内是否存在违规内容，从而输出当前直播帧的目标违规得分，此时该目标违规得分采用由视觉画面特征的多维行为特征相结合的多角度特征融合分析，来实现再次违规审核，无需采用额外的审核设备，降低了违规审核成本，保证直播间违规审核的准确性和召回性。
本实施例提供的技术方案，通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核，能够快速审核出容易区分和违规明显的违规直播帧，而对于通过初步违规审核的不易区分或者违规不明显的当前直播帧，会在该当前直播帧在高精确审核模型下的第一违规得分和在 高召回审核模型下的第二违规得分的基础上，结合该当前直播帧在待审核直播间内的多维行为特征，通过行为审核模型再次对当前直播帧进行视觉画面特征和行为特征相结合的统一特征分析，从而得到该当前直播帧的目标违规得分，实现待审核直播间内当前直播帧的全面违规审核，避免直播审核下存在误审核和漏审核的问题，在降低违规审核成本的基础上，提高直播审核的准确率和召回率。
实施例二
图2A为本申请实施例二提供的一种直播审核的方法的流程图，图2B为本申请实施例二提供的直播审核过程的原理示意图。本实施例是在上述实施例的基础上进行说明。如图2A所示，本实施例中对于待审核直播间内当前直播帧的再次违规审核的审核过程和行为审核模型的建模过程进行解释说明。
可选的，如图2A所示，本实施例中可以包括如下步骤。
S210，通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核。
S220，如果当前直播帧通过初步违规审核，则合并当前直播帧在高精确审核模型下的第一违规得分、在高召回审核模型下的第二违规得分以及在待审核直播间内的多维行为特征，得到当前直播帧的直播行为特征。
可选的，如果当前直播帧通过初步违规审核，说明高精确审核模型和高召回审核模型都认为当前直播帧内不存在违规内容，因此为了保证直播审核的准确性，避免出现违规漏审核的情况，还会利用当前直播帧的视觉画面特征和行为特征共同对该当前直播帧继续进行再次违规审核，此时会确定出当前直播帧在高精确审核模型下的第一违规得分和在高召回审核模型下的第二违规得分，将第一违规得分和第二违规得分作为当前直播帧对应的视觉画面特征，同时根据当前直播帧的直播时刻下，在待审核直播间内的多项直播互动行为，确定出当前直播帧在待审核直播间内的多维行为特征，将第一违规得分、第二违规得分和多维行为特征合并起来，得到当前直播帧的直播行为特征，作为再次违规审核下的多角度融合特征，后续采用该直播行为特征判断当前直播帧内是否存在违规内容，能够保证直播审核的全面性和准确性。
示例性的，本实施例中的直播行为特征可以为X＝(x1，x2，…，x11)，其中x1为当前直播帧在高精确审核模型下的第一违规得分，x2为当前直播帧在高召回审核模型下的第二违规得分，x3为多维行为特征中在待审核直播间内的直播实时观看数目，x4为多维行为特征中在待审核直播间内的用户发送公屏数量， x5为多维行为特征中在待审核直播间内的违规公屏数量，x6为多维行为特征中在待审核直播间内违规公屏在所有公屏中的占比，x7为多维行为特征中在待审核直播间内的直播打赏数量，x8为多维行为特征中在待审核直播间内的主播历史违规次数，x9为多维行为特征中在待审核直播间内的主播历史违规占比，x10为多维行为特征中在待审核直播间内的主播历史语音违规次数，x11为多维行为特征中在待审核直播间内的主播历史语音违规占比。
为了保证直播违规审核的准确性，本实施例中的多维行为特征可以动态配置，对此不作限定。
S230，将直播行为特征输入到行为审核模型中，得到当前直播帧的目标违规得分。
可选的，在得到直播行为特征后，可以将该直播行为特征作为多角度下的违规融合特征，输入到预先构建的行为审核模型中，通过该行为审核模型对该直播行为特征在每一个维度下的特征值进行融合分析，判断是否携带违规内容信息，进而输出当前直播帧的目标违规得分，通过该目标违规得分判断当前直播帧内是否存在违规内容，提高直播审核的准确性。
示例性的，本实施例可以采用如下步骤对行为审核模型进行建模：从历史直播视频集内的每一个历史直播视频中分别提取对应的历史直播帧，并确定每一个历史直播帧在所处直播间内的多维历史行为特征以及该历史直播帧的违规标签；针对每一个历史直播帧，将该历史直播帧在高精确审核模型下的第一历史违规得分、在高召回审核模型下的第二历史违规得分以及在所处直播间内的多维历史行为特征，输入到预先构建的行为审核模型中，得到历史直播帧的历史违规得分；以最小化每一个历史直播帧的历史违规得分和违规标签之间的差异为目标，不断优化行为审核模型。
通过查找多个直播间的历史直播视频，构建出用于建模的历史直播视频集，而由于对直播间进行违规审核时，主要是对直播过程中的直播帧进行违规审核，因此需要对历史直播视频集内的每一个历史直播视频进行视频帧提取，例如从每一个历史直播视频内每隔预设时长便截取一张历史直播图片，作为从该历史直播视频内所提取出的历史直播帧，此时所提取出的全部历史直播帧可以构成该行为审核模型的建模样本。
此时，为了保证行为审核模型的成功建模，需要设定建模过程中所需要参考的多个建模特征，以及该行为审核模型的建模优化目标，以按照该建模优化目标对多个建模特征下的实际建模情况进行不断优化，例如本实施例中的建模特征可以为X＝(x1，x2，…，x11)，其中x1为历史直播帧在高精确审核模型下的第一历史违规得分，x2为历史直播帧在高召回审核模型下的第二历史违规得 分，x3为所设定的多维历史行为特征中历史直播帧在所在直播间内的直播实时观看数目，x4为多维历史行为特征中历史直播帧在所在直播间内的用户发送公屏数量，x5为多维历史行为特征中历史直播帧在所在直播间内的违规公屏数量，x6为多维历史行为特征中历史直播帧在所在直播间内违规公屏在所有公屏中的占比，x7为多维历史行为特征中历史直播帧在所在直播间内的直播打赏数量，x8为多维历史行为特征中历史直播帧在所在直播间内的主播历史违规次数，x9为多维历史行为特征中历史直播帧在所在直播间内的主播历史违规占比，x10为多维历史行为特征中历史直播帧在所在直播间内的主播历史语音违规次数，x11为多维历史行为特征中历史直播帧在所在直播间内的主播历史语音违规占比。
由于多个建模特征中得历史直播帧在所处直播间内的多维历史行为特征在历史直播过程中就可以确定，因此在从每一个历史直播视频中分别提取出对应的历史直播帧后，通过判断历史直播帧在所处直播间内的历史直播互动行为，即可确定出每一个历史直播帧在所处直播间内的多维历史行为特征，同时还能够确定每一个历史直播帧的违规标签，如图2B所示，以便后续按照建模优化目标进行不断建模优化。针对每一个历史直播帧，可以将该历史直播帧分别输入到本实施例的高精度审核模型和高召回审核模型中，从而得到每一个历史直播帧的第一历史违规得分和第二历史违规得分，以此即可将每一个历史直播帧的第一历史违规得分、第二历史违规得分和在所处直播间内的多维历史行为特征，共同组成每一个历史直播帧对应的建模特征，并将每一个历史直播帧对应的建模特征不断输入到该行为审核模型中，得到每一个历史直播帧的历史违规得分。
此时，本实施例中行为审核模型的建模优化目标可以为以最小化每一个历史直播帧的历史违规得分和违规标签之间的差异为目标，因此在建模优化过程中不断分析每一个历史直播帧的历史违规得分与每一个历史直播帧的违规标签之间的差异，并通过优化行为审核模型中的数学参数来不断减小该差异，以达到历史直播帧的历史违规得分和违规标签之间的差异能够最小化的目标，从而得到最终优化后的行为审核模型，此时最终优化后的行为审核模型具备高准确性的违规审核能力。
示例性的，本实施例中行为审核模型的建模优化目标可以为：
Figure PCTCN2021141704-appb-000001
其中，x i为第i个历史直播帧的建模特征，y i为第i个历史直播帧的违规标签，y i＝1表示存在违规内容，y i＝0表示不存在违规内容，
Figure PCTCN2021141704-appb-000002
为通过行为审核模型对第i个历史直播帧确定的历史违规得分， V＝(x 1,x 2,L,x m)，表示历史直播帧组成的建模样本，m为样本数量，p(x i)表示x i通过行为审核模型得到的预测概率。
本实施例提供的技术方案，以最小化每一个历史直播帧的历史违规得分和违规标签之间的差异为建模优化目标，对行为审核模型进行建模，保证行为审核模型进行违规审核的准确性，后续通过合并待审核直播间内当前直播帧在高精确审核模型下的第一违规得分、在高召回审核模型下的第二违规得分以及在待审核直播间内的多维行为特征，得到当前直播帧的直播行为特征，并通过建模后的行为审核模型对该直播行为特征进行多角度下的违规特征分析，提高直播审核的准确率和召回率。
实施例三
图3A为本申请实施例三提供的一种直播审核的方法的流程图，图3B为本申请实施例三提供的直播审核过程的原理示意图。本实施例是在上述实施例的基础上进行说明。如图3A所示，本实施例中对待审核直播间的整体违规审核过程进行解释说明。
可选的，如图3A所示，本实施例中可以包括如下步骤。
S310，将待审核直播间内的当前直播帧输入到预先构建的高精确审核模型中，得到当前直播帧的第一违规得分。
可选的，在对待审核直播间进行违规审核时，会在直播过程中实时检测直播画面是否存在违规内容，因此需要实时对直播过程中的当前直播帧进行违规审核。将当前直播帧输入到预先构建的高精确审核模型中，通过该高精确审核模型对当前直播帧内的直播画面进行初始的违规审核，从而输出该当前直播帧的第一违规得分，后续通过判断该第一违规得分是否超出预设精确违规阈值，能够确定该当前直播帧内是否存在违规内容。
S320，判断第一违规得分是否超出预设精确违规阈值，若是，执行S380；若否，执行S330。
S330，将当前直播帧继续输入到预先构建的高召回审核模型中，得到当前直播帧的第二违规得分。
可选的，如果第一违规得分未超出预设精确违规阈值，说明高精确审核模型认为该当前直播帧内不存在违规内容，然而由于高精确审核模型存在违规漏审核的问题，因此需要将当前直播帧继续输入到预先构建的高召回审核模型中， 如图3B所示，通过该高召回审核模型继续对当前直播帧内的直播画面进行相应的违规审核，从而输出该当前直播帧的第二违规得分，后续通过判断该第二违规得分是否超出预设召回违规阈值，能够再次确定该当前直播帧内是否存在违规内容。
S340，判断第二违规得分是否超出预设召回违规阈值，若是，执行S380；若否，执行S350。
如果第二违规得分未超出预设召回违规阈值，说明高召回审核模型也认为该当前直播帧内不存在违规内容，也就是当前直播帧同时通过高精确审核模型和高召回审核模型的违规审核，也就是通过初步违规审核，利用行为审核模型继续对当前直播帧进行再次违规审核。所述当前直播帧通过所述初步违规审核为所述当前直播帧的第一违规得分未超出所述预设精确违规阈值，且所述当前直播帧的第二违规得分未超出预设召回违规阈值。
S350，将当前直播帧在所述高精确审核模型下的第一违规得分、在高召回审核模型下的第二违规得分以及在待审核直播间内的多维行为特征，输入到预先构建的行为审核模型中，得到当前直播帧的目标违规得分。
S360，判断目标违规得分是否超出预设行为违规阈值，若是，执行S380；若否，执行S370。
可选的，在利用行为审核模型输出当前直播帧的目标违规得分后，通过判断目标违规得分是否超出预设行为违规阈值，可以最终审核出当前直播帧内是否存在违规内容，以便执行不同的直播处理流程。
S370，间隔预设直播时长后，从待审核直播间内采集新的当前直播帧，继续对新的当前直播帧进行违规审核。
可选的，如果目标违规得分未超出预设行为违规阈值，说明行为审核模型也认为当前直播帧内不存在违规内容，此时本实施例中的三个模型均认为当前直播帧内不存在违规内容，则可以确定当前直播帧为非违规帧，说明待审核直播间内当前不存在违规内容，此时为了保证直播间违规审核的实时性，会在间隔预设直播时长(如每隔2s)后，继续从待审核直播间内的当前直播时刻下截取一张新的直播图片，作为所采集的新的当前直播帧，按照本实施例S310-S380的步骤继续对该新的当前直播帧进行违规审核，从而对待审核直播间进行实时违规审核，直至直播结束或者审核出违规的直播帧。
S380，将待审核直播间的直播信息推送给人工审核平台。
可选的，在第一违规得分超出预设精确违规阈值，或者第二违规得分超出预设召回违规阈值，或者目标违规得分超出预设行为违规阈值时，都说明当前 审核使用的模型认为该当前直播帧内存在违规内容，但是为了避免出现违规误审核的情况，如图3B所示，本实施例还会额外设置一个人工审核平台，以在审核出当前直播帧内存在违规内容时，将该当前直播帧所在的待审核直播间的直播信息推送给人工审核平台，以进一步对该待审核直播间的直播内容进行人工审核，从而保证直播审核的准确性。
本实施例中待审核直播间的直播信息可以为待审核直播间的直播地址或正在直播的视频画面等，对此不作限定。
S390，实时监控待审核直播间内的直播互动项，如果直播互动项超出预设互动阈值，则将待审核直播间的直播信息推送给人工审核平台，并停止对待审核直播间内当前直播帧的违规审核。
可选的，为了避免出现违规审核的异常情况，而保证直播审核的高召回率，本实施例在通过构建模型来实现直播违规审核的基础上，还会通过实时监控待审核直播间内的直播互动项，来判断是否需要直接推送给人工审核平台进行人工审核，将人工审核作为直播审核的审核异常保底方案。此时，该直播互动项可以包括待审核直播间内的直播观看数量和直播打赏数量等，以此分析待审核直播间的受众广泛程度，如果直播互动项超出预设互动阈值，说明待审核直播间的直播受众率很大，一旦出现违规内容，则会快速传播出去，因此需要直接将待审核直播间的直播信息推送给人工审核平台进行人工审核，以确保直播违规审核的高效性，同时停止利用所构建的多个模型对待审核直播间内当前直播帧的违规审核。
本实施例提供的技术方案，在确定当前直播帧内存在违规内容后，直接将待审核直播间的直播信息推送给人工审核平台进行人工审核，避免出现违规误审核的情况，提高直播违规审核的准确性；同时，利用待审核直播间内的直播互动项额外设置直播审核的异常保底方案，保证直播违规审核的全面性。
实施例四
图4为本申请实施例四提供的一种直播审核的装置的结构示意图。如图4所示，该装置可以包括：
初步审核模块410，设置为通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核；
违规得分确定模块420，设置为如果所述当前直播帧通过初步违规审核，则将所述当前直播帧在所述高精确审核模型下的第一违规得分、在所述高召回审核模型下的第二违规得分以及在所述待审核直播间内的多维行为特征，输入到 预先构建的行为审核模型中，得到所述当前直播帧的目标违规得分。
本实施例提供的技术方案，通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核，能够快速审核出容易区分和违规明显的违规直播帧，而对于通过初步违规审核的不易区分或者违规不明显的当前直播帧，会在该当前直播帧在高精确审核模型下的第一违规得分和在高召回审核模型下的第二违规得分的基础上，结合该当前直播帧在待审核直播间内的多维行为特征，通过行为审核模型再次对当前直播帧进行视觉画面特征和行为特征相结合的统一特征分析，从而得到该当前直播帧的目标违规得分，实现待审核直播间内当前直播帧的全面违规审核，避免直播审核下存在误审核和漏审核的问题，在降低违规审核成本的基础上，提高直播审核的准确率和召回率。
本实施例提供的直播审核的装置可适用于执行上述任意实施例提供的直播审核的方法，具备相应的功能和效果。
实施例五
图5为本申请实施例五提供的一种服务器的结构示意图，如图5所示，该服务器包括处理器50、存储装置51和通信装置52；服务器中处理器50的数量可以是一个或多个，图5中以一个处理器50为例；服务器中的处理器50、存储装置51和通信装置52可以通过总线或其他方式连接，图5中以通过总线连接为例。
本实施例提供的一种服务器可用于执行上述任意实施例提供的直播审核的方法，具备相应的功能和效果。
实施例六
本申请实施例六还提供了一种计算机可读存储介质，其上存储有计算机程序，该计算机程序被处理器执行时可实现上述任意实施例中的直播审核的方法。该方法可以包括：
通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核；
如果所述当前直播帧通过初步违规审核，则将所述当前直播帧在所述高精确审核模型下的第一违规得分、在所述高召回审核模型下的第二违规得分以及在所述待审核直播间内的多维行为特征，输入到预先构建的行为审核模型中， 得到所述当前直播帧的目标违规得分。
本申请实施例所提供的一种包含计算机可执行指令的存储介质，其计算机可执行指令不限于如上所述的方法操作，还可以执行本申请任意实施例所提供的直播审核的方法中的相关操作。存储介质可以是非暂态(non-transitory)存储介质。
本申请可借助软件及通用硬件来实现，也可以通过硬件实现。本申请的技术方案可以以软件产品的形式体现出来，该计算机软件产品可以存储在计算机可读存储介质中，如计算机的软盘、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、闪存(FLASH)、硬盘或光盘等，包括多个指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本申请多个实施例所述的方法。
上述直播审核的装置的实施例中，所包括的多个单元和模块只是按照功能逻辑进行划分的，但并不局限于上述的划分，只要能够实现相应的功能即可；另外，多个功能单元的名称也只是为了便于相互区分，并不用于限制本申请的保护范围。

Claims (12)

  1. 一种直播审核的方法，包括：
    通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核；
    在所述当前直播帧通过所述初步违规审核的情况下，将所述当前直播帧在所述高精确审核模型下的第一违规得分、在所述高召回审核模型下的第二违规得分以及在所述待审核直播间内的多维行为特征，输入到预先构建的行为审核模型中，得到所述当前直播帧的目标违规得分。
  2. 根据权利要求1所述的方法，其中，所述将所述当前直播帧在所述高精确审核模型下的第一违规得分、在所述高召回审核模型下的第二违规得分以及在所述待审核直播间内的多维行为特征，输入到预先构建的行为审核模型中，得到所述当前直播帧的目标违规得分，包括：
    合并所述当前直播帧在所述高精确审核模型下的第一违规得分、在所述高召回审核模型下的第二违规得分以及在所述待审核直播间内的多维行为特征，得到所述当前直播帧的直播行为特征；
    将所述直播行为特征输入到所述行为审核模型中，得到所述当前直播帧的目标违规得分。
  3. 根据权利要求1所述的方法，其中，所述行为审核模型通过执行如下步骤建模：
    从历史直播视频集内的每一个历史直播视频中提取对应的历史直播帧，并确定每一个历史直播帧在所处直播间内的多维历史行为特征以及所述历史直播帧的违规标签；
    针对每一个历史直播帧，将所述历史直播帧在所述高精确审核模型下的第一历史违规得分、在所述高召回审核模型下的第二历史违规得分以及在所处直播间内的多维历史行为特征，输入到预先构建的行为审核模型中，得到所述历史直播帧的历史违规得分；
    以最小化每一个历史直播帧的历史违规得分和违规标签之间的差异为目标，不断优化所述行为审核模型。
  4. 根据权利要求1所述的方法，其中，在所述将所述当前直播帧在所述高精确审核模型下的第一违规得分、在所述高召回审核模型下的第二违规得分以及在所述待审核直播间内的多维行为特征，输入到预先构建的行为审核模型中，得到所述当前直播帧的目标违规得分之后，还包括：
    在所述目标违规得分超出预设行为违规阈值的情况下，将所述待审核直播 间的直播信息推送给人工审核平台；
    在所述目标违规得分未超出所述预设行为违规阈值的情况下，间隔预设直播时长后，从所述待审核直播间内采集新的当前直播帧，继续对所述新的当前直播帧进行违规审核。
  5. 根据权利要求4所述的方法，还包括：
    实时监控所述待审核直播间内的直播互动项，在所述直播互动项超出预设互动阈值的情况下，将所述待审核直播间的直播信息推送给所述人工审核平台，并停止对所述待审核直播间内当前直播帧的违规审核。
  6. 根据权利要求1所述的方法，其中，所述通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核，包括：
    将所述待审核直播间内的当前直播帧输入到预先构建的高精确审核模型中，得到所述当前直播帧的第一违规得分；
    在所述第一违规得分未超出预设精确违规阈值的情况下，将所述当前直播帧继续输入到预先构建的高召回审核模型中，得到所述当前直播帧的第二违规得分。
  7. 根据权利要求6所述的方法，其中，所述当前直播帧通过所述初步违规审核为所述当前直播帧的第一违规得分未超出所述预设精确违规阈值，且所述当前直播帧的第二违规得分未超出预设召回违规阈值。
  8. 根据权利要求6所述的方法，还包括：
    在所述第一违规得分超出所述预设精确违规阈值，或者所述第二违规得分超出预设召回违规阈值的情况下，将所述待审核直播间的直播信息推送给人工审核平台。
  9. 根据权利要求1-8中任一项所述的方法，其中，所述高精确审核模型的训练样本由训练样本集中的易正样本和易负样本构成，所述高召回审核模型的训练样本由训练样本集中的难正样本和难负样本构成。
  10. 一种直播审核的装置，包括：
    初步审核模块，设置为通过级联的高精确审核模型和高召回审核模型，对待审核直播间内的当前直播帧进行初步违规审核；
    违规得分确定模块，设置为在所述当前直播帧通过所述初步违规审核的情况下，将所述当前直播帧在所述高精确审核模型下的第一违规得分、在所述高召回审核模型下的第二违规得分以及在所述待审核直播间内的多维行为特征，输入到预先构建的行为审核模型中，得到所述当前直播帧的目标违规得分。
  11. 一种服务器，包括：
    一个或多个处理器；
    存储装置，设置为存储一个或多个程序；
    当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现如权利要求1-9中任一项所述的直播审核的方法。
  12. 一种计算机可读存储介质，存储有计算机程序，其中，所述计算机程序被处理器执行时实现如权利要求1-9中任一项所述的直播审核的方法。
PCT/CN2021/141704 2020-12-30 2021-12-27 直播审核的方法、装置、服务器和存储介质 Ceased WO2022143541A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
US18/259,646 US12010358B2 (en) 2020-12-30 2021-12-27 Live streaming moderating method and apparatus, server, and storage medium
EP21914277.5A EP4274236A4 (en) 2020-12-30 2021-12-27 Live streaming auditing method and apparatus, server, and storage medium
JP2023540184A JP7584054B2 (ja) 2020-12-30 2021-12-27 ライブ配信審査の方法、ライブ配信審査の装置、サーバーおよび記憶媒体

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011613190.6 2020-12-30
CN202011613190.6A CN112788356B (zh) 2020-12-30 2020-12-30 一种直播审核的方法、装置、服务器和存储介质

Publications (1)

Publication Number Publication Date
WO2022143541A1 true WO2022143541A1 (zh) 2022-07-07

Family

ID=75754059

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/141704 Ceased WO2022143541A1 (zh) 2020-12-30 2021-12-27 直播审核的方法、装置、服务器和存储介质

Country Status (5)

Country Link
US (1) US12010358B2 (zh)
EP (1) EP4274236A4 (zh)
JP (1) JP7584054B2 (zh)
CN (1) CN112788356B (zh)
WO (1) WO2022143541A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119402678A (zh) * 2025-01-03 2025-02-07 杭州高顾智能科技有限公司 一种金融行业直播智能实时质检方法与系统

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112788356B (zh) * 2020-12-30 2023-04-28 百果园技术(新加坡)有限公司 一种直播审核的方法、装置、服务器和存储介质
CN114049084A (zh) * 2021-11-10 2022-02-15 数贸科技（北京）有限公司 风控审核处理方法、装置及电子设备
JP7391174B1 (ja) * 2022-12-02 2023-12-04 １７Ｌｉｖｅ株式会社 サーバ及びコンピュータプログラム

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107197331A (zh) * 2017-05-03 2017-09-22 北京奇艺世纪科技有限公司 一种实时监测直播内容的方法及装置
CN109862394A (zh) * 2019-03-27 2019-06-07 北京周同科技有限公司 视频内容的审核方法、装置、设备和存储介质
US20190253744A1 (en) * 2018-02-13 2019-08-15 Ernest Huang Systems and methods for content management of live or streaming broadcasts and video publishing systems
CN111079816A (zh) * 2019-12-11 2020-04-28 北京金山云网络技术有限公司 图像的审核方法、装置和服务器
CN111385602A (zh) * 2018-12-29 2020-07-07 广州市百果园信息技术有限公司 基于多层级多模型的视频审核方法、介质及计算机设备
CN111382623A (zh) * 2018-12-28 2020-07-07 广州市百果园信息技术有限公司 一种直播审核的方法、装置、服务器和存储介质
CN111506708A (zh) * 2020-04-22 2020-08-07 上海极链网络科技有限公司 一种文本审核方法、装置、设备和介质
CN112788356A (zh) * 2020-12-30 2021-05-11 百果园技术(新加坡)有限公司 一种直播审核的方法、装置、服务器和存储介质

Family Cites Families (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8788442B1 (en) 2010-12-30 2014-07-22 Google Inc. Compliance model training to classify landing page content that violates content item distribution guidelines
BR102016007265B1 (pt) * 2016-04-01 2022-11-16 Samsung Eletrônica da Amazônia Ltda. Método multimodal e em tempo real para filtragem de conteúdo sensível
CN106131595A (zh) * 2016-05-26 2016-11-16 武汉斗鱼网络科技有限公司 一种用于视频直播的标题敏感词控制方法及装置
CN107959863B (zh) * 2016-10-14 2020-04-03 北京酷我科技有限公司 一种视频审核方法及系统
US10349126B2 (en) * 2016-12-19 2019-07-09 Samsung Electronics Co., Ltd. Method and apparatus for filtering video
CN109766807A (zh) 2018-12-28 2019-05-17 广州华多网络科技有限公司 机器审核处理方法、装置、电子设备及存储介质
CN111382605B (zh) * 2018-12-28 2023-08-18 广州市百果园信息技术有限公司 视频内容审核方法、装置、存储介质和计算机设备
CN109831698B (zh) * 2018-12-28 2021-07-23 广州华多网络科技有限公司 信息审核方法、装置、电子设备及计算机可读取存储介质
JP7042768B2 (ja) 2019-03-20 2022-03-28 ヤフー株式会社 付与装置、付与方法および付与プログラム
CN111090776B (zh) * 2019-12-20 2023-06-30 广州市百果园信息技术有限公司 一种视频审核的方法、装置、审核服务器和存储介质
CN111225234B (zh) * 2019-12-23 2022-07-26 广州市百果园信息技术有限公司 视频审核方法、视频审核装置、设备和存储介质
CN111147876A (zh) * 2019-12-26 2020-05-12 山东爱城市网信息技术有限公司 基于区块链的直播内容监测方法、设备、存储介质及平台
CN112116613B (zh) * 2020-09-25 2021-10-15 贝壳找房(北京)科技有限公司 图像矢量化方法及系统

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107197331A (zh) * 2017-05-03 2017-09-22 北京奇艺世纪科技有限公司 一种实时监测直播内容的方法及装置
US20190253744A1 (en) * 2018-02-13 2019-08-15 Ernest Huang Systems and methods for content management of live or streaming broadcasts and video publishing systems
CN111382623A (zh) * 2018-12-28 2020-07-07 广州市百果园信息技术有限公司 一种直播审核的方法、装置、服务器和存储介质
CN111385602A (zh) * 2018-12-29 2020-07-07 广州市百果园信息技术有限公司 基于多层级多模型的视频审核方法、介质及计算机设备
CN109862394A (zh) * 2019-03-27 2019-06-07 北京周同科技有限公司 视频内容的审核方法、装置、设备和存储介质
CN111079816A (zh) * 2019-12-11 2020-04-28 北京金山云网络技术有限公司 图像的审核方法、装置和服务器
CN111506708A (zh) * 2020-04-22 2020-08-07 上海极链网络科技有限公司 一种文本审核方法、装置、设备和介质
CN112788356A (zh) * 2020-12-30 2021-05-11 百果园技术(新加坡)有限公司 一种直播审核的方法、装置、服务器和存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4274236A4

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119402678A (zh) * 2025-01-03 2025-02-07 杭州高顾智能科技有限公司 一种金融行业直播智能实时质检方法与系统

Also Published As

Publication number Publication date
US12010358B2 (en) 2024-06-11
CN112788356B (zh) 2023-04-28
EP4274236A1 (en) 2023-11-08
EP4274236A4 (en) 2024-05-08
JP7584054B2 (ja) 2024-11-15
CN112788356A (zh) 2021-05-11
JP2024504576A (ja) 2024-02-01
US20230396816A1 (en) 2023-12-07

Similar Documents

Publication Publication Date Title
JP7584054B2 (ja) ライブ配信審査の方法、ライブ配信審査の装置、サーバーおよび記憶媒体
CN112492343A (zh) 一种视频直播监控方法及相关装置
CN112818888B (zh) 视频审核模型训练方法、视频审核方法及相关装置
CN110784759A (zh) 弹幕信息处理方法、装置、电子设备及存储介质
TWI470562B (zh) 用於使用選定的區域分類之臉部的追蹤及辨識之方法，設備及電腦可讀取非暫態儲存媒體
CN110263680B (zh) 图像处理方法、装置和系统及存储介质
WO2021082589A1 (zh) 内容审核模型的训练方法和装置、视频内容的审核方法和装置、计算机设备以及存储介质
CN112183166A (zh) 确定训练样本的方法、装置和电子设备
CN110569804B (zh) 基于文本数据的失效场景判断方法及系统、服务器及介质
CN111586432B (zh) 空播直播间确定方法、装置、服务器及存储介质
CN107784083A (zh) 一种网络舆情信息有效性的自动识别处理方法
US20230179627A1 (en) Learning apparatus, detecting apparatus, learning method, detecting method, learning program, and detecting program
WO2022111688A1 (zh) 人脸活体检测方法、装置及存储介质
CN110490079A (zh) 巡检数据处理方法、装置、计算机设备和存储介质
CN116527879A (zh) 机顶盒测试方法、装置、介质和计算机装置
CN119832673B (zh) 一种基于物联网的校园安防系统及其服务终端
Lee et al. Activating self-attention for multi-scene absolute pose regression
CN112911323B (zh) 直播互动评估方法、装置、电子设备和可读存储介质
CN112257768B (zh) 一种非法金融图片的识别方法、装置、计算机存储介质
CN111382626A (zh) 视频中违规图像的检测方法、装置、设备及存储介质
CN113656314A (zh) 压力测试处理方法及装置
Phan et al. A hybrid approach for multimedia use verification
CN112733809B (zh) 一种用于自然保护区监控系统的智能图像识别方法及系统
CN116012785A (zh) 一种火势等级的确定方法、装置、设备及介质
CN112116460B (zh) 基于区块链的专利质押业务处理方法、系统及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21914277

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2023540184

Country of ref document: JP

WWE Wipo information: entry into national phase

Ref document number: 2023119048

Country of ref document: RU

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2021914277

Country of ref document: EP

Effective date: 20230731

WWE Wipo information: entry into national phase

Ref document number: 11202305030Y

Country of ref document: SG

WWG Wipo information: grant in national office

Ref document number: 11202305030Y

Country of ref document: SG

WWP Wipo information: published in national office

Ref document number: 11202305030Y

Country of ref document: SG