Live broadcast room detection method and device, electronic equipment and computer readable storage medium

Abstract

The embodiment of the invention provides a live broadcast room detection method and device, electronic equipment and a computer readable storage medium, and relates to the technical field of large models, and the method comprises the steps: carrying out the analysis of a live broadcast data stream, obtaining a to-be-detected target video and a to-be-detected target voice, and storing the to-be-detected target video and the to-be-detected target voice; detecting the target video and the target voice based on the trained multi-modal large model and at least one preset detection strategy to obtain a detection result corresponding to each detection strategy; and if at least one of the detection results satisfies an alarm condition, generating alarm information. Through the above mode, unified video analysis of the live broadcasting room is realized through the multi-mode large model, a more intelligent and accurate violation detection method is provided, and powerful guarantee is provided for market supervision. Moreover, different image, voice and video acquisition modes are designed according to different timeliness and detection modes of illegal scene detection, and shunting detection is carried out, so that the detection efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of large models, and in particular to a live broadcast room detection method, a live broadcast room detection device, an electronic device, and a computer-readable storage medium. Background Art

[0002] With the rapid development of the network live broadcast industry, problems such as lack of main body responsibility, poor content ecology, and uneven quality of live broadcast hosts have emerged. These problems have severely restricted the healthy development of the network live broadcast industry and brought challenges to ideological security and social public interests. There are problems such as false and exaggerated publicity, data fraud, and fake and shoddy goods in the live broadcast industry.

[0003] At the same time, the national policy requirements for the live broadcast industry emphasize the importance of standardized development and promoting the healthy and orderly development of the industry. In order to enhance the comprehensive governance ability, the state has also proposed various policies aimed at protecting the rights and interests of consumers and minors, promoting the healthy development of the industry, while enhancing the leading role of mainstream values, and ensuring the positive impact and social responsibility of the network live broadcast industry.

[0004] Therefore, it is very necessary for market supervision departments to strengthen the supervision of network live broadcasts, which helps to build a good industrial ecosystem and create a positive, healthy, content-rich, and full-of-positive-energy network live broadcast space for the general public, especially teenagers.

[0005] Traditional market supervision departments' methods for network live broadcast supervision mainly include the following:

[0006] 1. Manual review: Arrange special review personnel to watch the live broadcast content in real time and judge whether there are any violations based on experience and relevant regulations.

[0007] 2. Automatic review: Utilize traditional deep learning technologies, use image recognition algorithms to detect whether there are any illegal image contents such as violence, porn, and blood in the live broadcast screen, analyze the behavior patterns of the live broadcast host, such as whether there are abnormal behaviors such as inducing rewards and malicious speculation. Through speech recognition, convert the host's words into text, set a series of sensitive keyword and illegal vocabulary lists, analyze whether they contain illegal terms or inappropriate remarks, and give an early warning when these words appear in the live broadcast.

[0008] However, the above methods have the following problems:

[0009] 1. Manual review requires a large amount of manpower and financial resources to allocate corresponding review personnel, requires training before taking up the post, and manual review has strong subjectivity and non-standardization.

[0010] 2. Traditional deep learning methods also have certain limitations in the supervision of live videos and audio, especially in image recognition and speech recognition, where misjudgments and incorrect judgments occur, and moreover, there is a lack of the ability to understand the overall video content. Summary of the Invention

[0011] In view of the above problems, embodiments of the present invention are proposed to provide a live broadcast room detection method, a live broadcast room detection device, an electronic device, and a computer-readable storage medium that overcome the above problems or at least partially solve the above problems.

[0012] To solve the above problems, embodiments of the present invention disclose a live broadcast room detection method, and the method includes:

[0013] Parse the live data stream to obtain the target video and target voice to be detected;

[0014] Detect the target video and the target voice based on at least one preset detection strategy to obtain a detection result corresponding to each detection strategy;

[0015] If at least one of the detection results in each detection result satisfies the alarm condition, an alarm message is generated.

[0016] In one or more embodiments, it further includes:

[0017] Count the number of alarm times for each detection strategy;

[0018] Perform a violation assessment based on each alarm count to obtain a violation assessment result;

[0019] If the violation assessment result meets the preset violation penalty condition, a violation penalty is imposed on the live broadcast room corresponding to the live data stream.

[0020] In one or more embodiments, the detecting the target video and the target voice based on at least one preset detection strategy to obtain a detection result corresponding to each detection strategy includes:

[0021] Perform a bullet screen detection on the target video based on a bullet screen detection strategy to obtain a bullet screen detection result;

[0022] Perform a blacklist detection of the host on the target video based on a host detection strategy to obtain a host detection result;

[0023] Perform a commodity identification on the target video and the target voice based on a commodity identification strategy to obtain a commodity identification result;

[0024] Perform a violation behavior detection on the target video and the target voice based on a violation behavior detection strategy to obtain a violation behavior detection result.

[0025] In one or more embodiments, performing bullet screen detection on the target video based on the bullet screen detection strategy to obtain a bullet screen detection result, including:

[0026] Performing frame extraction on the target video based on a preset first frame extraction strategy to obtain a plurality of first video frames;

[0027] Detecting the bullet screen in each of the first video frames to obtain a bullet screen detection result.

[0028] In one or more embodiments, performing host blacklist detection on the target video based on the host detection strategy to obtain a host detection result, including:

[0029] Performing face recognition on the target video to obtain face data;

[0030] Matching the face data with a blacklist face database to obtain a host detection result.

[0031] In one or more embodiments, performing commodity identification on the target video and the target voice based on the commodity identification strategy to obtain a commodity identification result, including:

[0032] Performing frame extraction on the target video based on a preset second frame extraction strategy to obtain a plurality of second video frames;

[0033] Obtaining first voice segments corresponding to the plurality of second video frames from the target voice;

[0034] Performing commodity identification using the plurality of second video frames and the first voice segments to obtain a commodity identification result.

[0035] In one or more embodiments, performing violation detection on the target video and the target voice based on the violation detection strategy to obtain a violation detection result, including:

[0036] Performing frame extraction on the target video based on a preset third frame extraction strategy to obtain a plurality of third video frames;

[0037] Obtaining second voice segments corresponding to the plurality of third video frames from the target voice;

[0038] Performing violation detection using the plurality of third video frames and the second voice segments to obtain a violation detection result.

[0039] Correspondingly, an embodiment of the present invention discloses a live broadcast room detection device, and the device includes:

[0040] A parsing module for parsing the live data stream to obtain a target video and a target voice to be detected;

[0041] A detection module for detecting the target video and the target voice based on at least one preset detection strategy to obtain a detection result corresponding to each detection strategy;

[0042] A generation module for generating an alarm message if at least one of the detection results satisfies an alarm condition.

[0043] In one or more embodiments, it further includes:

[0044] A statistics module for counting the number of alarm times for each detection strategy;

[0045] An evaluation module for performing a violation evaluation based on each of the alarm times to obtain a violation evaluation result;

[0046] A punishment module for performing a violation punishment on the live broadcast room corresponding to the live data stream if the violation evaluation result satisfies a preset violation punishment condition.

[0047] In one or more embodiments, the detection module includes:

[0048] A bullet screen detection sub-module for performing bullet screen detection on the target video based on a bullet screen detection strategy to obtain a bullet screen detection result;

[0049] A host blacklist detection sub-module for performing host blacklist detection on the target video based on a host detection strategy to obtain a host detection result;

[0050] A commodity identification sub-module for performing commodity identification on the target video and the target voice based on a commodity identification strategy to obtain a commodity identification result;

[0051] A violation behavior detection sub-module for performing violation behavior detection on the target video and the target voice based on a violation behavior detection strategy to obtain a violation behavior detection result.

[0052] In one or more embodiments, the bullet screen detection sub-module is specifically configured to:

[0053] Perform frame extraction on the target video based on a preset first frame extraction strategy to obtain a plurality of first video frames;

[0054] Detect the bullet screen in each of the first video frames to obtain a bullet screen detection result.

[0055] In one or more embodiments, the host blacklist detection sub-module is specifically configured to:

[0056] Perform face recognition on the target video to obtain face data;

[0057] Match the face data with the blacklist face database to obtain the host detection result.

[0058] In one or more embodiments, the commodity identification sub-module is specifically configured to:

[0059] Extract frames from the target video based on a preset second frame extraction strategy to obtain a plurality of second video frames;

[0060] Obtain first voice segments corresponding to the plurality of second video frames from the target voice;

[0061] Perform commodity identification using the plurality of second video frames and the first voice segments to obtain a commodity identification result.

[0062] In one or more embodiments, the violation detection sub-module is specifically configured to:

[0063] Extract frames from the target video based on a preset third frame extraction strategy to obtain a plurality of third video frames;

[0064] Obtain second voice segments corresponding to the plurality of third video frames from the target voice;

[0065] Perform violation detection using the plurality of third video frames and the second voice segments to obtain a violation detection result.

[0066] Correspondingly, an embodiment of the present invention discloses an electronic device, including: a processor, a memory, and a computer program stored on the memory and capable of running on the processor. When the computer program is executed by the processor, each step of the above-mentioned live broadcast room detection method embodiment is implemented.

[0067] Correspondingly, an embodiment of the present invention discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, each step of the above-mentioned live broadcast room detection method embodiment is implemented.

[0068] Embodiments of the present invention have the following advantages:

[0069] The detection system can parse the live data stream corresponding to the live broadcast room to obtain the target video and target voice to be detected, and then detect the target video and the target voice based on the trained multi-modal large model and at least one preset detection strategy to obtain the detection result corresponding to each detection strategy; if at least one of the detection results in each detection result meets the alarm condition, an alarm message is generated. Through the above method, not only the unified video parsing of the live broadcast room is realized through the multi-modal large model, providing a more intelligent and accurate illegal detection method, which provides a strong guarantee for market supervision. Moreover, in view of the different timeliness and detection methods of illegal scene detection, different image, voice, and video acquisition methods are designed for split detection, improving the detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 is a flowchart of the steps of Embodiment 1 of a live broadcast room detection method of the present invention;

[0071] Figure 2 is a flowchart of the steps of Embodiment 2 of a live broadcast room detection method of the present invention;

[0072] Figure 3 is a logical flowchart of Embodiment 2 of a live broadcast room detection method of the present invention;

[0073] Figure 4 is a structural block diagram of an embodiment of a live broadcast room detection device of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0074] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0075] One of the core concepts of the embodiments of the present invention is that not only the unified video parsing of the live broadcast room is realized through the multi-modal large model, providing a more intelligent and accurate illegal detection method, which provides a strong guarantee for market supervision. Moreover, in view of the different timeliness and detection methods of illegal scene detection, different image, voice, and video acquisition methods are designed for split detection, improving the detection efficiency.

[0076] Refer to Figure 1 , which shows a flowchart of the steps of Embodiment 1 of a live broadcast room detection method of the present invention. The embodiments of the present invention can be applied to a detection system, which can detect each live broadcast room in the live broadcast system to determine whether there are any violations in each live broadcast room. Among them, the detection system and the live broadcast system operate independently of each other. Moreover, the detection system and the live broadcast system can be deployed on the same platform or on different platforms. In actual applications, the deployment method of the detection system and the live broadcast system can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0077] Further, the administrator can input the address of the live broadcast room to be detected in the detection system, so that the detection system can detect the live broadcast room according to the address. Of course, in practical applications, the detection system can also detect the live broadcast room in other ways, and the embodiments of the present invention do not limit this.

[0078] Moreover, the detection system can detect a single live broadcast room or multiple live broadcast rooms. In practical applications, the number of live broadcast rooms to be detected can be set according to actual needs, and the embodiments of the present invention do not limit this. In addition to detecting live broadcast rooms, it is also possible to detect the historical live data stream of the live broadcast room (such as live broadcast replay). In practical applications, the detection object can also be set according to actual needs, and the embodiments of the present invention do not limit this either.

[0079] In the embodiments of the present invention, taking the detection of the live data stream of a single live broadcast room as an example for illustration, the following steps can be specifically included:

[0080] Step 101, parse the live data stream to obtain the target video and target voice to be detected.

[0081] After the detection system obtains the live data stream of the live broadcast room to be detected, it can parse the live data stream to obtain the video to be detected (denoted as "target video") and voice (denoted as "target voice").

[0082] Step 102, detect the target video and the target voice based on at least one preset detection strategy to obtain the detection result corresponding to each detection strategy.

[0083] After obtaining the target video and target voice, the trained multi-modal large model can be used to detect the target video and target voice. The trained multi-modal large model can include a language multi-modal service and a graphic and text multi-modal service.

[0084] Among them, the language multi-modal service can be built based on qwen2-audio and is used for directly performing semantic parsing and speech recognition on audio segments; the graphic and text multi-modal service can be built by fine-tuning the base large model of qwen2-vl-7b and is used for in-depth parsing of videos and pictures, and can not only analyze dynamic video content, but also extract single-frame picture content.

[0085] It should be noted that the language multi-modal service can be built based on qwen2-audio, or it can be built based on other models. The graphic and text multi-modal service can be built by qwen2-vl-7b, or it can be built based on other models. In practical applications, the specific models for building the language multi-modal service and the graphic and text multi-modal service can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0086] Furthermore, during detection, at least one preset detection strategy can be adopted to detect the target video and the target voice, so as to obtain the detection result corresponding to each detection strategy, that is, to detect whether the target video and the target voice meet each detection strategy.

[0087] In the embodiments of the present invention, the detection of the target video and the target voice based on at least one preset detection strategy to obtain the detection result corresponding to each detection strategy includes:

[0088] Performing bullet screen detection on the target video based on the bullet screen detection strategy to obtain a bullet screen detection result;

[0089] Performing blacklist detection of the host on the target video based on the host detection strategy to obtain a host detection result;

[0090] Performing commodity identification on the target video and the target voice based on the commodity identification strategy to obtain a commodity identification result;

[0091] Performing detection of illegal behaviors on the target video and the target voice based on the illegal behavior detection strategy to obtain an illegal behavior detection result.

[0092] Specifically, the detection strategies can include a bullet screen detection strategy, a host detection strategy, a commodity identification strategy, and an illegal behavior detection strategy.

[0093] Among them, the bullet screen detection strategy is used to automatically analyze the content of the bullet screen area for display and detect inappropriate remarks in the live broadcast room bullet screen. Inappropriate remarks include but are not limited to pornographic, violent, political, terrorist, and abusive remarks.

[0094] The host detection strategy is used to identify and mark the hosts prohibited from live broadcasting.

[0095] The commodity identification strategy is used to evaluate whether the commodities displayed in the live broadcast comply with regulations and whether there is false promotion, etc.

[0096] The illegal behavior detection strategy is used to detect inappropriate behaviors and remarks in the live broadcast. Inappropriate behaviors include but are not limited to borderline behaviors, violent behaviors, etc.

[0097] Based on this, a bullet chat detection strategy can be adopted to detect the bullet chat content in the target video, thereby obtaining a bullet chat detection result; a host detection strategy can be adopted to detect the host in the target video for the host blacklist, thereby obtaining a host detection result; a commodity identification strategy can be adopted to identify the commodities in the target video and the target voice, thereby obtaining a commodity identification result; a violation detection strategy can be adopted to detect the actions and speeches in the target video and the target voice for violations, thereby obtaining a violation detection result.

[0098] In this way, through the bullet chat detection strategy, real-time and accurate identification and display of the content in the bullet chat area can be achieved, ensuring that users can obtain bullet chat information in a timely manner. At the same time, real-time monitoring of the illegal terms in the bullet chat is carried out. Once illegal content is found, the alarm system is immediately triggered to fully maintain the good order of the platform.

[0099] Through the commodity identification strategy, voice and video analysis technologies can be fully utilized to deeply analyze the products introduced by the host in the video and carefully compare them with the actual products. Through this comparative analysis process, accurately judge whether the product introduction in the video matches the actual product, so as to quickly detect and properly handle false propaganda phenomena, and effectively guarantee the authenticity and reliability of market information.

[0100] Through the violation detection strategy, the behaviors and speeches of the host that do not conform to the norms can be detected. Once illegal content is found, the alarm system is immediately triggered to fully maintain the good order of the platform.

[0101] Through the host detection strategy, real-time monitoring and accurate identification of the host in the video can be carried out. Once a prohibited host is identified, an alarm is immediately issued and even the live broadcast is prohibited.

[0102] It should be noted that in practical applications, the preset detection strategies may include, in addition to the bullet chat detection strategy, the host detection strategy, the commodity identification strategy, and the violation detection strategy, other detection strategies. The specific detection strategies can be adjusted according to actual needs, and the embodiments of the present invention do not limit this.

[0103] Moreover, each detection strategy can be executed simultaneously or sequentially. In practical applications, the execution order of each detection strategy can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0104] In the embodiments of the present invention, the bullet chat detection of the target video based on the bullet chat detection strategy to obtain a bullet chat detection result includes:

[0105] Performing frame extraction on the target video based on a preset first frame extraction strategy to obtain a plurality of first video frames;

[0106] Detect the bullet screens in each of the said first video frames to obtain bullet screen detection results.

[0107] Specifically, when performing bullet screen detection on a target video, a preset frame extraction strategy (denoted as the "first frame extraction strategy") can be used to extract frames from the target video, thereby obtaining multiple video frames (denoted as "first video frames"). For example, a video frame can be extracted from the target video every 10 seconds, thereby obtaining multiple first video frames.

[0108] After obtaining multiple first video frames, a text and image multi-modal service can be used to detect the bullet screens in each first video frame. If there are any illegal bullet screens, the illegal bullet screens can be recorded in the detection system and displayed through the detection system.

[0109] It should be noted that, in addition to the above frame extraction method, the first frame extraction strategy can also be other frame extraction methods. In practical applications, the specific frame extraction method can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0110] In the embodiments of the present invention, performing the host blacklist detection on the target video based on the host detection strategy to obtain host detection results includes:

[0111] Perform face recognition on the target video to obtain face data;

[0112] Match the face data with the blacklist face database to obtain host detection results.

[0113] Specifically, before performing host blacklist detection, a blacklist face database can be built and the face data of the banned hosts can be stored in the blacklist face database. When performing host blacklist detection, face recognition can be performed on the host in the target video to obtain the host's face data, and then the host's face data can be matched with the blacklist face database. If the match is successful, it means that there is a banned host in the target video.

[0114] Among them, when performing host blacklist detection, face recognition and matching can be performed through the magface model. Of course, other models can also be used for detection. In practical applications,

[0115] Furthermore, before matching, a similarity threshold can be set, such as 0.67. When matching, when the similarity between the host's face data and the face data in the blacklist face database exceeds the similarity threshold, it means that the match is successful.

[0116] Moreover, before performing the host blacklist detection, a threshold for the number of video frames can be set. When performing the host blacklist detection, multiple consecutive video frames can be detected. If the number of video frames of the banned host exceeds the threshold for the number of video frames, it indicates that there is a banned host in the live room. For example, when detecting 30 consecutive frames in the target video, if the number of video frames of the banned host detected in the 30 frames is not less than 10 frames, it can be determined that there is a banned host in the live room.

[0117] It should be noted that, in addition to setting the threshold for the number of video frames, the host blacklist detection can also be performed by other methods. In practical applications, the specific detection method can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0118] In the embodiments of the present invention, the commodity identification of the target video and the target voice based on the commodity identification strategy to obtain a commodity identification result includes:

[0119] Extracting frames from the target video based on a preset second frame extraction strategy to obtain multiple second video frames;

[0120] Obtaining a first voice segment corresponding to the multiple second video frames from the target voice;

[0121] Performing commodity identification using the multiple second video frames and the first voice segment to obtain a commodity identification result.

[0122] Specifically, when performing commodity identification, a preset frame extraction strategy (denoted as "second frame extraction strategy") can be used to extract frames from the target video, thereby obtaining multiple video frames (denoted as "second video frames"). For example, one video frame can be extracted from the target video every 5 seconds, and a total of 10 extractions are performed, thereby obtaining 10 second video frames.

[0123] And, obtaining a voice segment corresponding to the second video frame from the target voice (denoted as "first voice segment"). For example, continuing with the above example, the video duration corresponding to the 10 second video frames is 50 seconds. Then, the voice of the 50 - second video can be used as the second voice segment, that is, the duration of the second voice segment is also 50 seconds.

[0124] Then, a graphic - text multi - modal service can be used to detect each second video frame, including but not limited to whether there are exaggerated publicity words such as "number one in the whole network", "cheapest", "top - level", etc. in the video frame, whether there are inappropriate words such as "specially supplied", "exclusive agent", etc., and at the same time, identifying the product type, product information, etc. of the product in the video frame.

[0125] Moreover, a language multi-modal service can be used to detect the first voice segment, including but not limited to whether the voice segment includes exaggerated publicity words, inappropriate words, product types, and product information.

[0126] Compare the product type in the video frame with the product type in the voice segment, and compare the product information in the video frame with the product information in the voice segment. If there are differences in at least one of the product type and product information, it indicates that there are situations such as false publicity and counterfeit and shoddy goods in the live broadcast room.

[0127] Furthermore, if there are any illegal words in either the words in the video frame or the words in the voice segment, it indicates that there are situations of exaggerated publicity and inappropriate publicity in the live broadcast room.

[0128] It should be noted that in addition to the above frame extraction methods, the second frame extraction strategy can also be other frame extraction methods. In actual applications, the specific frame extraction method can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0129] In the embodiments of the present invention, detecting illegal behaviors of the target video and the target voice based on the illegal behavior detection strategy to obtain an illegal behavior detection result includes:

[0130] Performing frame extraction on the target video based on a preset third frame extraction strategy to obtain a plurality of third video frames;

[0131] Obtaining a second voice segment corresponding to the plurality of third video frames from the target voice;

[0132] Performing illegal behavior detection on the plurality of third video frames and the second voice segment to obtain an illegal behavior detection result.

[0133] Specifically, when performing illegal behavior detection, a preset frame extraction strategy (denoted as "third frame extraction strategy") can be used to perform frame extraction on the target video, thereby obtaining a plurality of video frames (denoted as "third video frames"). For example, one video frame can be extracted from the target video every 10 seconds, and a total of 20 extractions are performed, thereby obtaining 20 third video frames.

[0134] Moreover, a voice segment corresponding to the third video frame (denoted as "second voice segment") is obtained from the target voice. For example, continuing with the above example, the video duration corresponding to the 20 third video frames is 200 seconds. Then, the voice of the 200-second video can be used as the second voice segment, that is, the duration of the second voice segment is also 200 seconds.

[0135] Then, a graphic and text multi-modal service can be used to detect each third video frame, including but not limited to whether there are predefined violations in the video frame. Additionally, a language multi-modal service can be used to detect the second voice segment, including but not limited to whether the voice segment contains pornographic, violent, political, terrorist, or abusive remarks.

[0136] If there is at least one of the violations and inappropriate remarks, it means that there are violations in the live broadcast room.

[0137] It should be noted that in addition to the above frame extraction methods, the third frame extraction strategy can also be other frame extraction methods. In actual applications, the specific frame extraction method can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0138] Step 103, if at least one of the detection results satisfies the alarm condition, an alarm message is generated.

[0139] Specifically, after obtaining each detection result, if at least one of the detection results satisfies the alarm condition, an alarm message can be generated and the alarm message can be stored for subsequent viewing.

[0140] Furthermore, after generating the alarm message, the alarm message can be output so that the management personnel can manage the corresponding live broadcast room in a timely manner. Of course, in actual applications, whether to output the alarm message can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0141] In the embodiments of the present invention, the detection system can parse the live data stream corresponding to the live broadcast room to obtain the target video and target voice to be detected, and then detect the target video and the target voice based on the trained multi-modal large model and at least one preset detection strategy to obtain the detection result corresponding to each detection strategy; if at least one of the detection results satisfies the alarm condition, an alarm message is generated. Through the above method, not only the unified video parsing of the live broadcast room is realized through the multi-modal large model, providing a more intelligent and accurate violation detection method, which provides a strong guarantee for market supervision. Moreover, aiming at the different timeliness and detection methods of violation scene detection, different image, voice, and video acquisition methods are designed for split detection, improving the detection efficiency.

[0142] Refer to Figure 2 , which shows the step flowchart of the second embodiment of a live broadcast room detection method of the present invention, and specifically may include the following steps:

[0143] Step 201, parse the live data stream to obtain the target video and target voice to be detected.

[0144] After obtaining the live data stream of the live broadcast room to be detected, the detection system can parse the live data stream to obtain the target video and target voice to be detected.

[0145] Step 202, detect the target video and the target voice based on at least one preset detection strategy to obtain the detection result corresponding to each detection strategy.

[0146] After obtaining the target video and target voice, the trained multi-modal large model can be used to detect the target video and target voice. The trained multi-modal large model can include a language multi-modal service and a graphic and text multi-modal service.

[0147] Among them, the language multi-modal service can be built based on qwen2-audio and is used to directly perform semantic parsing and speech recognition on audio segments; the graphic and text multi-modal service can be built by fine-tuning the base large model of qwen2-vl-7b and is used to deeply analyze videos and pictures. It can not only analyze dynamic video content but also extract single-frame picture content.

[0148] It should be noted that the language multi-modal service can be built based on qwen2-audio or other models, and the graphic and text multi-modal service can be built by qwen2-vl-7b or other models. In practical applications, the specific models for building the language multi-modal service and the graphic and text multi-modal service can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0149] Furthermore, during detection, at least one preset detection strategy can be used to detect the target video and the target voice, so as to obtain the detection result corresponding to each detection strategy, that is, to detect whether the target video and the target voice meet each detection strategy.

[0150] In the embodiments of the present invention, the detecting the target video and the target voice based on at least one preset detection strategy to obtain the detection result corresponding to each detection strategy includes:

[0151] Perform bullet screen detection on the target video based on the bullet screen detection strategy to obtain a bullet screen detection result;

[0152] Perform blacklist detection of the host on the target video based on the host detection strategy to obtain a host detection result;

[0153] Perform commodity identification on the target video and the target voice based on the commodity identification strategy to obtain a commodity identification result;

[0154] Perform violation detection on the target video and the target voice based on the violation detection strategy to obtain the violation detection result.

[0155] Specifically, the detection strategy may include a bullet screen detection strategy, a host detection strategy, a product identification strategy, and a violation detection strategy.

[0156] Among them, the bullet screen detection strategy is used to automatically parse the content in the bullet screen area for display and detect inappropriate remarks in the live room bullet screen. Inappropriate remarks include, but are not limited to, pornographic, violent, political, terrorist, and abusive remarks.

[0157] The host detection strategy is used to identify and mark the hosts prohibited from live streaming.

[0158] The product identification strategy is used to evaluate whether the products displayed in the live stream comply with the regulations and whether there is false promotion, etc.

[0159] The violation detection strategy is used to detect inappropriate behaviors and remarks in the live stream. Inappropriate behaviors include, but are not limited to, borderline behaviors, violent behaviors, etc.

[0160] Based on this, the bullet screen detection strategy can be used to perform bullet screen detection on the bullet screen content in the target video to obtain the bullet screen detection result; the host detection strategy can be used to perform host blacklist detection on the host in the target video to obtain the host detection result; the product identification strategy can be used to identify the products in the target video and the target voice to obtain the product identification result; the violation detection strategy can be used to detect the actions and remarks in the target video and the target voice for violation detection to obtain the violation detection result.

[0161] In this way, through the bullet screen detection strategy, real-time and accurate identification and display of the content in the bullet screen area can be achieved, ensuring that users can obtain bullet screen information in a timely manner. At the same time, real-time monitoring of the illegal terms in the bullet screen is carried out. Once illegal content is found, the alarm system is immediately triggered to fully maintain the good order of the platform.

[0162] Through the product identification strategy, voice and video analysis technologies can be fully utilized to deeply analyze the products introduced by the host in the video and carefully compare them with the actual products. Through this comparative analysis process, accurately judge whether the product introduction in the video matches the actual product, so as to quickly detect and properly handle false promotion phenomena, and effectively guarantee the authenticity and reliability of market information.

[0163] Through the violation detection strategy, the behaviors and remarks of the host that do not meet the specifications can be detected. Once illegal content is found, the alarm system is immediately triggered to fully maintain the good order of the platform.

[0164] Through the host detection strategy, the host in the video can be monitored in real time and accurately identified. Once a prohibited host is identified, an alarm will be issued immediately, and even the live broadcast will be prohibited.

[0165] It should be noted that in practical applications, the preset detection strategies can include not only the bullet screen detection strategy, the host detection strategy, the commodity identification strategy, and the violation behavior detection strategy, but also other detection strategies. The specific detection strategies can be adjusted according to actual needs, and the embodiments of the present invention do not limit this.

[0166] In the embodiment of the present invention, the bullet screen detection of the target video based on the bullet screen detection strategy to obtain a bullet screen detection result includes:

[0167] Performing frame extraction on the target video based on a preset first frame extraction strategy to obtain a plurality of first video frames;

[0168] Detecting the bullet screen in each of the first video frames to obtain a bullet screen detection result.

[0169] Specifically, when performing bullet screen detection on the target video, a preset first frame extraction strategy can be used to extract frames from the target video, thereby obtaining a plurality of first video frames.

[0170] After obtaining a plurality of first video frames, the bullet screen in each first video frame can be detected using a graphic and text multi-modal service. If there are illegal bullet screens, the illegal bullet screens can be recorded in the detection system and displayed through the detection system.

[0171] It should be noted that the first frame extraction strategy can also be other frame extraction methods in addition to the above frame extraction method. In practical applications, the specific frame extraction method can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0172] In the embodiment of the present invention, the host blacklist detection of the target video based on the host detection strategy to obtain a host detection result includes:

[0173] Performing face recognition on the target video to obtain face data;

[0174] Matching the face data with the blacklist face database to obtain a host detection result.

[0175] Specifically, before performing host blacklist detection, a blacklist face database can be built and the face data of the prohibited hosts can be stored in the blacklist face database. When performing host blacklist detection, face recognition can be performed on the host in the target video to obtain the host's face data, and then the host's face data can be matched with the blacklist face database. If the match is successful, it means that there is a prohibited host in the target video.

[0176] Among them, when detecting the host blacklist, face recognition and matching can be performed through the magface model. Of course, other models can also be used for detection. In actual applications,

[0177] Furthermore, before matching, a similarity threshold can be set. When matching, if the similarity between the face data of the host and the face data in the blacklist face database exceeds the similarity threshold, it means that the matching is successful.

[0178] Moreover, before detecting the host blacklist, a video frame number threshold can be set. When detecting the host blacklist, multiple consecutive video frames can be detected. If the number of video frames of the banned host exceeds the video frame number threshold, it means that there is a banned host in the live broadcast room.

[0179] It should be noted that, in addition to setting the video frame number threshold, the host blacklist detection can also be performed in other ways. In actual applications, the specific detection method can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0180] In the embodiments of the present invention, performing commodity identification on the target video and the target voice based on a commodity identification strategy to obtain a commodity identification result includes:

[0181] Extracting frames from the target video based on a preset second frame extraction strategy to obtain multiple second video frames;

[0182] Obtaining first voice segments corresponding to the multiple second video frames from the target voice;

[0183] Performing commodity identification using the multiple second video frames and the first voice segments to obtain a commodity identification result.

[0184] Specifically, when performing commodity identification, a preset second frame extraction strategy can be used to extract frames from the target video, thereby obtaining multiple second video frames.

[0185] And, obtaining first voice segments corresponding to the second video frames from the target voice. Then, a graphic and text multi-modal service can be used to detect each second video frame, including but not limited to whether there are exaggerated publicity words such as "number one in the whole network", "cheapest", "top level", etc. in the video frame, whether there are inappropriate words such as "specially provided", "exclusive agent", etc., and at the same time, identifying the product type, product information, etc. of the product in the video frame.

[0186] In addition, a language multi-modal service can be used to detect the first voice segment, including but not limited to whether the voice segment contains exaggerated publicity words, inappropriate words, product types, and product information.

[0187] Compare the product type in the video frame with the product type in the voice segment, and, compare the product information in the video frame with the product information in the voice segment. If there is a difference in at least one of the product type and product information, it indicates that there are situations such as false publicity and fake and shoddy goods in the live broadcast room.

[0188] Moreover, if there is any violation word in the words in the video frame or the words in the voice segment, it indicates that there are situations of exaggerated publicity and inappropriate publicity in the live broadcast room.

[0189] It should be noted that in addition to the above frame extraction methods, the second frame extraction strategy can also be other frame extraction methods. In actual applications, the specific frame extraction method can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0190] In the embodiments of the present invention, the detecting of the target video and the target voice for violation behaviors based on the violation behavior detection strategy to obtain a violation behavior detection result includes:

[0191] Extract frames from the target video based on a preset third frame extraction strategy to obtain a plurality of third video frames;

[0192] Obtain a second voice segment corresponding to the plurality of third video frames from the target voice;

[0193] Perform violation behavior detection using the plurality of third video frames and the second voice segment to obtain a violation behavior detection result.

[0194] Specifically, when performing violation behavior detection, a preset third frame extraction strategy can be used to extract frames from the target video, thereby obtaining a plurality of third video frames.

[0195] In addition, obtain a second voice segment corresponding to the third video frame from the target voice. Then, a graphic and text multi-modal service can be used to detect each third video frame, including but not limited to whether there are predefined violation behaviors in the video frame. And, a language multi-modal service can be used to detect the second voice segment, including but not limited to whether the voice segment contains pornographic, violent, political, terrorist, or abusive remarks.

[0196] If there is at least one of violation behaviors and inappropriate remarks, it indicates that there is a violation behavior in the live broadcast room.

[0197] It should be noted that, in addition to the above-mentioned frame extraction methods, the third frame extraction strategy can also be other frame extraction methods. In practical applications, the specific frame extraction method can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0198] Step 203, if at least one of the detection results satisfies the alarm condition, an alarm message is generated.

[0199] Specifically, after obtaining each detection result, if at least one of the detection results satisfies the alarm condition, then an alarm message can be generated and the alarm message can be stored for subsequent browsing.

[0200] Furthermore, after generating the alarm message, the alarm message can be output so that the management personnel can manage the corresponding live broadcast room in a timely manner. Of course, in practical applications, whether to output the alarm message can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0201] Step 204, count the number of alarm times for each of the detection strategies.

[0202] Step 205, perform a violation assessment based on each of the alarm times to obtain a violation assessment result.

[0203] Step 206, if the violation assessment result meets the preset violation penalty condition, execute a violation penalty on the live broadcast room corresponding to the live data stream.

[0204] Specifically, after generating the alarm message, the number of alarm times corresponding to each detection strategy can be counted, and a violation assessment can be performed based on the number of alarm times corresponding to each detection strategy. For example, a violation score can be calculated based on each alarm time to obtain the violation assessment result. If the violation assessment result meets the preset violation penalty condition, for example, if the violation score exceeds the violation score threshold (such as 60), then it can be determined that the violation penalty condition is met, and thus a violation penalty is executed on the live broadcast room corresponding to the live data stream; if the violation assessment result does not meet the violation penalty condition, then it can jump to Step 201 to continue detecting the live broadcast room.

[0205] Among them, when performing the violation assessment, the following formula can be used to calculate the violation score:

[0206] score = cumulative number of bullet screen alarms * 1 + cumulative number of banned anchor alarms * 4

[0207] + cumulative number of violation behavior alarms * 2 + cumulative number of product violation alarms * 3

[0208] It should be noted that the "1", "2", "3", and "4" in the above formula are the coefficients corresponding to the respective alarm times. In actual applications, the specific values of each coefficient can be set according to actual needs, and the embodiments of the present invention do not limit this.

[0209] Moreover, in addition to using the method of calculating the violation score for violation assessment, other methods can also be used for violation assessment. In actual applications, the method of violation assessment for the tooth root can also be set according to actual needs, and the embodiments of the present invention do not limit this either.

[0210] For ease of understanding, refer to Figure 3 , which shows the logic flow chart of the embodiments of the present invention. Specifically, after obtaining the live data data stream of the live broadcast room, the live data stream is parsed to obtain the target video and the target voice.

[0211] Perform a host blacklist detection on the target video to obtain a host detection result; use a preset first frame extraction strategy to extract frames from the target video to obtain a plurality of first video frames, perform a bullet screen detection on the plurality of first video frames to obtain a bullet screen detection result; use a preset second frame extraction strategy to extract frames from the target video to obtain a plurality of second video frames, obtain a first voice segment corresponding to the plurality of second video frames from the target voice, and perform a commodity identification on the plurality of second video frames and the first voice segment to obtain a commodity identification result; use a preset third frame extraction strategy to extract frames from the target video to obtain a plurality of third video frames, obtain a second voice segment corresponding to the plurality of third video frames from the target voice, and perform a violation behavior detection on the plurality of third video frames and the second voice segment to obtain a violation behavior detection result.

[0212] If there is a violation behavior, then generate an alarm (not shown in the figure), and increment the violation behavior alarm count by 1; if there is a banned host, then generate an alarm (not shown in the figure), and increment the banned host alarm count by 1; if there is a bullet screen violation, then generate an alarm (not shown in the figure), and increment the bullet screen violation alarm count by 1; if there is a commodity violation, then generate an alarm (not shown in the figure), and increment the commodity violation alarm count by 1.

[0213] Calculate the violation score based on each alarm count. If the violation score exceeds the violation score threshold, then impose a violation penalty on the live broadcast room; if the violation assessment result does not meet the violation penalty conditions, then the live broadcast room can continue to be detected.

[0214] In an embodiment of the present invention, the detection system can parse the live data stream corresponding to the live broadcast room to obtain the target video and target voice to be detected, and then detect the target video and the target voice based on the trained multi-modal large model and at least one preset detection strategy to obtain the detection result corresponding to each detection strategy; if at least one of the detection results in each detection result meets the alarm condition, an alarm message is generated. In the above manner, not only is the unified video parsing of the live broadcast room achieved through the multi-modal large model, providing a more intelligent and accurate method for detecting violations, which provides a strong guarantee for market supervision. Moreover, in view of the different timeliness and detection methods for detecting violation scenarios, different image, voice, and video acquisition methods are designed for diversion detection, improving the detection efficiency.

[0215] Furthermore, a more reasonable evaluation mechanism is designed to evaluate each detection strategy separately and as a whole, so as to make the violation punishment for the live broadcast room more reasonable. This not only improves the transparency of market supervision, enables the punished live broadcast room to easily understand the specific reasons for the violation, and thus effectively avoids future violations, but also realizes the intelligence and rationality of the supervision process, providing a strong guarantee for the healthy development of the network live broadcast environment.

[0216] It should be noted that for the method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the embodiments of the present invention are not limited by the described action sequence, because according to the embodiments of the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential for the embodiments of the present invention.

[0217] Refer to Figure 4 , which shows the structural block diagram of an embodiment of a live broadcast room detection device of the present invention, and specifically may include the following modules:

[0218] The parsing module 401 is used to parse the live data stream to obtain the target video and target voice to be detected;

[0219] The detection module 402 is used to detect the target video and the target voice based on at least one preset detection strategy to obtain the detection result corresponding to each detection strategy;

[0220] The generation module 403 is used to generate an alarm message if at least one of the detection results meets the alarm condition.

[0221] In an embodiment of the present invention, it further includes:

[0222] The statistics module is used to count the number of alarm times for each detection strategy;

[0223] An evaluation module, configured to perform a violation evaluation based on each of the alarm counts to obtain a violation evaluation result;

[0224] A penalty module, configured to perform a violation penalty on the live broadcast room corresponding to the live data stream if the violation evaluation result meets a preset violation penalty condition.

[0225] In an embodiment of the present invention, the detection module includes:

[0226] A bullet screen detection sub-module, configured to perform bullet screen detection on the target video based on a bullet screen detection strategy to obtain a bullet screen detection result;

[0227] A host blacklist detection sub-module, configured to perform host blacklist detection on the target video based on a host detection strategy to obtain a host detection result;

[0228] A commodity identification sub-module, configured to perform commodity identification on the target video and the target voice based on a commodity identification strategy to obtain a commodity identification result;

[0229] A violation behavior detection sub-module, configured to perform violation behavior detection on the target video and the target voice based on a violation behavior detection strategy to obtain a violation behavior detection result.

[0230] In an embodiment of the present invention, the bullet screen detection sub-module is specifically configured to:

[0231] Perform frame extraction on the target video based on a preset first frame extraction strategy to obtain a plurality of first video frames;

[0232] Detect the bullet screens in each of the first video frames to obtain a bullet screen detection result.

[0233] In an embodiment of the present invention, the host blacklist detection sub-module is specifically configured to:

[0234] Perform face recognition on the target video to obtain face data;

[0235] Match the face data with a blacklist face database to obtain a host detection result.

[0236] In an embodiment of the present invention, the commodity identification sub-module is specifically configured to:

[0237] Perform frame extraction on the target video based on a preset second frame extraction strategy to obtain a plurality of second video frames;

[0238] Obtain a first voice segment corresponding to the plurality of second video frames from the target voice;

[0239] Use multiple of the second video frames and the first voice segment for product authentication to obtain a product authentication result.

[0240] In an embodiment of the present invention, the illegal behavior detection sub-module is specifically configured to:

[0241] Perform frame extraction on the target video based on a preset third frame extraction strategy to obtain multiple third video frames;

[0242] Obtain a second voice segment corresponding to multiple of the third video frames from the target voice;

[0243] Use multiple of the third video frames and the second voice segment for illegal behavior detection to obtain an illegal behavior detection result.

[0244] For the apparatus embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For related parts, refer to the partial description of the method embodiment.

[0245] An embodiment of the present invention further provides an electronic device, including:

[0246] Including a processor, a memory, and a computer program stored on the memory and capable of running on the processor. When the computer program is executed by the processor, it implements each process of the above-mentioned live broadcast room detection method embodiment and can achieve the same technical effect. To avoid repetition, it will not be elaborated here.

[0247] An embodiment of the present invention further provides a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program is executed by the processor, it implements each process of the above-mentioned live broadcast room detection method embodiment and can achieve the same technical effect. To avoid repetition, it will not be elaborated here.

[0248] Each embodiment in this specification is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. For the same or similar parts among the embodiments, refer to each other.

[0249] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, an apparatus, or a computer program product. Therefore, the embodiments of the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0250] Embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal devices generate means for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or means for implementing the functions specified in multiple blocks.

[0251] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implement the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or means for implementing the functions specified in multiple blocks.

[0252] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, so that a series of operation steps are executed on the computer or other programmable terminal device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable terminal device provide steps for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or means for implementing the functions specified in multiple blocks.

[0253] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications to these embodiments once they know the basic creative concepts. Therefore, the appended claims are intended to be construed to include the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

[0254] Finally, it should also be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or terminal device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or terminal device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or terminal device comprising the said element.

[0255] The above has introduced in detail a live broadcast room detection method and a live broadcast room detection device provided by the present invention. Specific examples are used in this text to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A live broadcast room detection method, characterized in that: The method comprises: Parse the live data stream to obtain the target video and target voice to be detected; Detecting the target video and the target voice based on at least one preset detection strategy to obtain a detection result corresponding to each detection strategy; If at least one of the detection results satisfies the alarm condition, an alarm message is generated.

2. The live broadcast room detection method according to claim 1, characterized in that: Also includes: Counting the number of alarms for each detection strategy; Performing violation assessment based on each of the alarm times to obtain a violation assessment result; If the violation assessment result meets the preset violation penalty conditions, the violation penalty will be executed on the live broadcast room corresponding to the live broadcast data stream.

3. The live broadcast room detection method according to claim 1, characterized in that: The detecting the target video and the target voice based on at least one preset detection strategy to obtain a detection result corresponding to each detection strategy includes: Performing bullet screen detection on the target video based on the bullet screen detection strategy to obtain a bullet screen detection result; Perform anchor blacklist detection on the target video based on the anchor detection strategy to obtain an anchor detection result; Performing commodity identification on the target video and the target voice based on a commodity identification strategy to obtain a commodity identification result; Based on the illegal behavior detection strategy, illegal behavior detection is performed on the target video and the target voice to obtain an illegal behavior detection result.

4. The live broadcast room detection method according to claim 3, characterized in that: The step of performing bullet screen detection on the target video based on the bullet screen detection strategy to obtain a bullet screen detection result includes: Extract frames from the target video based on a preset first frame extraction strategy to obtain a plurality of first video frames; The bullet comments in each of the first video frames are detected to obtain a bullet comment detection result.

5. The live broadcast room detection method according to claim 3, characterized in that: The performing anchor blacklist detection on the target video based on the anchor detection strategy to obtain an anchor detection result includes: Performing face recognition on the target video to obtain face data; The facial data is matched with the blacklist facial database to obtain the anchor detection result.

6. The live broadcast room detection method according to claim 3, characterized in that: The performing commodity identification on the target video and the target voice based on the commodity identification strategy to obtain a commodity identification result includes: Extract frames from the target video based on a preset second frame extraction strategy to obtain a plurality of second video frames; Acquire a first speech segment corresponding to a plurality of the second video frames from the target speech; The plurality of second video frames and the first voice segment are used to perform commodity identification to obtain a commodity identification result.

7. The live broadcast room detection method according to claim 3, characterized in that: The step of performing violation detection on the target video and the target voice based on the violation detection strategy to obtain a violation detection result includes: Extract frames from the target video based on a preset third frame extraction strategy to obtain a plurality of third video frames; Acquire a second speech segment corresponding to the plurality of third video frames from the target speech; A plurality of the third video frames and the second voice segment are used to perform violation detection to obtain a violation detection result.

8. A live broadcast room detection device, characterized in that: The device comprises: The parsing module is used to parse the live data stream to obtain the target video and target voice to be detected; A detection module, used to detect the target video and the target voice based on at least one preset detection strategy, and obtain a detection result corresponding to each detection strategy; The generating module is used to generate alarm information if at least one of the detection results meets the alarm condition.

9. An electronic device, characterized in that: include: A processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein when the computer program is executed by the processor, the steps of the live broadcast room detection method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the live broadcast room detection method as described in any one of claims 1 to 7 are implemented.

Images