Granted Patent B2

US 12,513,365 · App. 17/772,392 · Granted Dec 30, 2025

Inventors: Feng Shi (Guangzhou, CN); Zhenqiang Liu (Guangzhou, CN)

Assignee: BIGO TECHNOLOGY PTE. LTD.

H04N21/466G06T3/40G06V10/22G06V10/774G06V10/82G06V20/41G06V20/46G06V20/49H04N21/4542

Abstract

Claims (62)

1 . A method for training a content moderation model, comprising:

extracting part of image data of a sample video file as sample image data;

positioning a time point of the sample image data on a timeline of the sample video file in response to a determination that the sample image data contains offensive content;

extracting salient image region data from the image data around the time point; and

training the content moderation model based on the image region data and the sample image data;

wherein extracting the salient image region data from the image data around the time point comprises:

determining a time range containing the time point;

looking up a salient region detection model configured to identify a salient image region in the image data; and

identifying the salient image region data in the image data by inputting the image data within the time range into the salient region detection model.

2 . The method according to claim 1 , wherein extracting part of image data of the sample video file as the sample image data comprises:

partitioning the sample video file into at least two sample video segments; and

extracting part of the image data from each sample video segment as the sample image data.

3 . The method according to claim 2 , wherein extracting part of image data of the sample video file as the sample image data comprises at least one of:

chronologically ranking the sample image data; and

scaling the sample image data to a preset size.

4 . The method according to claim 1 , wherein positioning the time point of the sample image data on the timeline of the sample video file in response to the determination that the sample image data contains the offensive content comprises:

looking up an offense discrimination model configured to identify an image offense score of the content in the image data;

identifying the image offense score of content in the sample image data by inputting the sample image data into the offense discrimination model;

selecting the sample image data with the image offense score meeting a preset offense condition; and

determining the time point of the sample image data meeting the preset offense condition in the sample video file.

5 . The method according to claim 4 , wherein looking up the offense discrimination model comprises:

determining an offense category marked on the sample video file and representing the offensive content; and

looking up the offense discrimination model corresponding to the offense category, wherein the offense discrimination model is configured to identify the image offense score of the content that belongs to the offense category in the image data.

6 . The method according to claim 4 , wherein selecting the sample image data with the image offense score meeting the preset offense condition comprises:

determining, whether the sample image data comprises the image offense score that is greater than a preset image score threshold;

determining that the image offense score meets the preset offense condition in the case that the sample image data comprises the image offense score that is greater than the preset image score threshold; and

determining that the image offense score of a maximum value meets a preset violation condition in the case that the sample image data does not comprise the image offense score that is greater than the preset image score threshold.

7 . The method according to claim 1 , wherein training the content moderation model based on the image region data and the sample image data comprises:

determining an offense category marked on the sample video file and representing the offensive content;

acquiring a deep neural network and a pre-trained model;

initializing the deep neural network by the pre-trained model;

training, by backpropagation, the deep neural network as the content moderation model based on the image region data, the sample image data, and the offense category.

8 . A method for moderating video content, comprising:

extracting part of image data of a target video file as target image data;

positioning a time point of the target image data on a timeline of the target video file in response to a determination that the target image data contains offensive content;

extracting salient image region data from the image data around the time point; and

moderating content of the target video file by inputting the image region data and the target image data into a preset content moderation model;

wherein extracting the salient image region data from the image data around the time point comprises:

determining a time range containing the time point;

looking up a salient region detection model configured to identify a salient image region in the image data; and

identifying the salient image region data in the image data by inputting the image data within the time range into the salient region detection model.

9 . The method according to claim 8 , wherein inputting the image region data and the target image data into the preset content moderation model to moderate the content of the target video file comprises:

determining a file offense score of the content of the target video file by inputting the image region data and the target image data into the preset content moderation model in the case that the content of the target video file belongs to a preset offense category;

determining a file score threshold;

determining the content of the target video file to be legal in the case that the file offense score is less than or equal to the file score threshold.

10 . The method according to claim 9 , wherein moderating the content of the target video file by inputting the image region data and the target image data into the preset content moderation model further comprises:

distributing the target video file to a designated client in the case that the file offense score is greater than the file score threshold;

determining the content of the target video file to be legal in the case that first moderation information is received from the client; and

determining the content of the target video file to be offensive in the case that second moderation information is received from the client.

11 . The method according to claim 9 , wherein determining the file score threshold comprises:

determining a total quantity of target video files with a previous time period, wherein the file offense score of the target video file has been determined;

generating the file score threshold, such that a ratio of a moderation quantity to the total quantity matches a preset push ratio, wherein the moderation quantity is a quantity of the target video files of which file offense scores are greater than the file score threshold.

12 . A computer device for training content moderation model, comprising:

one or more processors;

a memory configured to store one or more programs;

wherein the one or more processors, when running the one or more programs, is caused to perform the method for training the content moderation model as defined in claim 1 .

13 . A non-volatile computer readable storage medium, storing a computer program, wherein the computer program, when run by a processor of a computer device, causes the computer device to perform the method for training the content moderation model as defined in claim 1 .

14 . A computer device for moderating video content, comprising:

one or more processors;

a memory configured to store one or more programs;

wherein the one or more processors, when running the one or more programs, is caused to perform the method for moderating the video content as defined in claim 8 .

15 . A non-volatile computer readable storage medium, storing a computer program, wherein the computer program, when run by a processor of a computer device, causes the computer device to perform the method for moderating the video content as defined in claim 8 .

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (43)