Large-screen video content review method and device, and storage medium

Abstract

The application discloses a large-screen video content review method and device and a storage medium, relates to the technical field of Internet security, and comprises the following steps: obtaining a video to be reviewed, performing decoupling processing on the video to be reviewed, obtaining an audio stream and multiple video frames in the video to be reviewed, and converting the audio stream into an audio text; performing a text compliance detection operation on the audio text to obtain a first text review result corresponding to the audio text; performing an image compliance detection operation on the video frames to obtain an image review result corresponding to the video frames; if any one of the first text review result and the image review result is non-compliant, determining that a video review result of the video to be reviewed is non-compliant, and replacing the video to be reviewed with a preset compliant video to enable the preset compliant video to be played on a large-screen display device, thereby solving the technical problem of low recognition accuracy of single-mode detection in a complex scene and improving the accuracy and reliability of video content review.

Description

Technical Field

[0001] This application relates to the field of internet security technology, and in particular to a method, device and storage medium for reviewing large-screen video content. Background Technology

[0002] With the increasing prevalence of large-screen displays in public places, the security of video content has become a growing concern. Current mainstream review methods employ a single-modal independent analysis mechanism. Image review uses deep learning models to detect content deemed unsuitable for workplace viewing (Not Safe For Work, NSFW), while audio review uses keyword matching to perform basic auditory review. These two processes are isolated from each other, making this single-modal detection mode insufficient to comprehensively capture multi-dimensional violations in videos. This vulnerability is easily exploited due to missing information or misjudgments. Therefore, existing large-screen video content review technologies suffer from fragmented multi-modal information processing, making it difficult for review systems to comprehensively and accurately identify sensitive cross-modal content. This deficiency can directly impact the security and reliability of large-screen content distribution, especially in highly sensitive locations such as transportation hubs and commercial centers. If violating information spreads through one modality, it could trigger significant social repercussions. Summary of the Invention

[0003] The main purpose of this application is to provide a method, device and storage medium for reviewing large-screen video content, which aims to solve the technical problem of insufficient recognition accuracy caused by single-modal detection in the prior art.

[0004] To achieve the above objectives, this application proposes a method for reviewing large-screen video content, the method comprising:

[0005] The video to be reviewed is obtained, the video to be reviewed is decoupled to obtain the audio stream and multiple video frames in the video to be reviewed, and the audio stream is converted into audio text;

[0006] Perform a text compliance check on the audio text to obtain the first text review result corresponding to the audio text;

[0007] Perform an image compliance detection operation on the video frame to obtain the image review result corresponding to the video frame;

[0008] If either the first text review result or the image review result is non-compliant, the video review result of the video to be reviewed is determined to be non-compliant, and the video to be reviewed is replaced with a preset compliant video so that the preset compliant video can be played on the large screen display device.

[0009] In one embodiment, the step of performing a text compliance detection operation on the audio text to obtain a first text review result corresponding to the audio text includes:

[0010] Calculate the overall similarity between the audio text and the preset illegal text, and determine the first component result of the first text review result corresponding to the audio text based on the magnitude of the overall similarity.

[0011] The audio text is input into a text classification model, and the audio text is classified by the text classification model to obtain a first classification result corresponding to the audio text. The first classification result is used to indicate whether the category of the audio text is a preset violation category.

[0012] Based on the first classification result, determine the second component result of the first text review result corresponding to the audio text;

[0013] If either the first component result or the second component result of the first text review result is non-compliant, the first text review result is determined to be non-compliant.

[0014] If both the first component result and the second component result of the first text review are compliant, the first text review result is determined to be compliant.

[0015] In one embodiment, the step of calculating the overall similarity between the audio text and the preset illegal text includes:

[0016] The text classification model is used to extract features from the audio text and the preset illegal text respectively, to obtain the audio text feature vector and the preset illegal text feature vector, and the cosine similarity between the audio text feature vector and the preset illegal text feature vector is calculated.

[0017] The keyword matching degree between the audio text and the preset illegal text is calculated using the Best Match 25 (BM25) algorithm.

[0018] The comprehensive similarity between the audio text and the preset illegal text is obtained by weighted summation based on the cosine similarity and the keyword matching degree.

[0019] In one embodiment, before the step of inputting the audio text into the text classification model, the method further includes:

[0020] Collect illegal text data, which includes multiple illegal texts;

[0021] The violation text and multiple preset violation categories are input into a large language model, and the violation text is matched to obtain the text category of the violation text.

[0022] Based on each of the aforementioned violating texts and their respective text categories, a text classification training set is constructed, wherein the text category corresponding to each violating text is the label of the violating text;

[0023] The text classification model is obtained by fine-tuning the training model based on the Bidirectional Encoder Representations from Transformers (BERT) model architecture using the text classification training set.

[0024] In one embodiment, the step of performing an image compliance detection operation on the video frame to obtain the image review result corresponding to the video frame includes:

[0025] Extract the image of the object to be reviewed from the video frame, and calculate the similarity between the object to be reviewed and a preset sensitive object based on the image of the object to be reviewed. Based on the value of the similarity, determine the first component result of the image review result corresponding to the video frame.

[0026] The video frame is input into an image classification model, and the video frame is classified by the image classification model to obtain a second classification result corresponding to the video frame. The second classification result is used to indicate whether the category of the video frame is a preset sensitive category.

[0027] Based on the second classification result, the second component result of the image review result corresponding to the video frame is determined;

[0028] The video frame is input into a character recognition model for character extraction to obtain the video frame text.

[0029] Perform a text compliance check on the video frame text to obtain the second text review result corresponding to the video frame text;

[0030] If any one of the first component result of the image review result, the second component result of the image review result, and the second text review result is non-compliant, then the image review result is determined to be non-compliant.

[0031] If the first component result of the image review result, the second component result of the image review result, and the second text review result are all compliant, then the image review result is determined to be compliant.

[0032] In one embodiment, the step of calculating the similarity between the object to be reviewed in the image of the object to be reviewed and a preset sensitive object based on the image of the object to be reviewed includes:

[0033] The facial features of the object to be reviewed are extracted by the face recognition model to obtain the facial feature vector of the object to be reviewed corresponding to the image of the object to be reviewed.

[0034] Obtain a preset sensitive object facial feature vector, wherein the preset sensitive object facial feature vector is obtained by extracting facial features from a preset sensitive object image containing the preset sensitive object through the face recognition model;

[0035] Based on the facial feature vector of the object to be reviewed and the facial feature vector of the preset sensitive object, the similarity between the object to be reviewed and the preset sensitive object is determined.

[0036] In one embodiment, before the step of inputting the video frame into the image classification model, the method further includes:

[0037] Collect sensitive image data, which includes multiple sensitive images;

[0038] Multiple preset sensitive categories and the sensitive images are input into the visual language big model, and category matching is performed on the sensitive images to obtain the image category of the sensitive images;

[0039] Based on each of the sensitive images and the image categories of each of the sensitive images, an image classification training set and an image classification test set are generated;

[0040] Multiple image classification models to be trained are trained using the image classification training set to obtain multiple trained candidate image classification models.

[0041] The classification accuracy of each candidate image classification model is tested using the image classification test set, and the candidate image classification model with the highest classification accuracy is selected from the multiple candidate image classification models to be determined as the image classification model.

[0042] In one embodiment, after the step of obtaining the image review result corresponding to the video frame, the method further includes:

[0043] If both the first text review result and the image review result are compliant, then the video review result is determined to be compliant.

[0044] The video to be reviewed is sent to the large-screen display device through a preset protocol so that the video to be reviewed can be played on the large-screen display device. The preset protocol is either the High-Definition Multimedia Interface (HDMI) protocol or the Video Graphics Array (VGA) protocol.

[0045] Furthermore, to achieve the above objectives, this application also proposes a large-screen video content review device, which includes:

[0046] The image text extraction module is used to acquire the video to be reviewed, decouple the video to be reviewed to obtain the audio stream and multiple video frames in the video to be reviewed, and convert the audio stream into audio text.

[0047] The text review module performs a text compliance check on the audio text to obtain the first text review result corresponding to the audio text.

[0048] The image review module performs image compliance detection on the video frame to obtain the image review result corresponding to the video frame;

[0049] The determination module is used to determine that the video review result of the video to be reviewed is non-compliant if either the first text review result or the image review result is non-compliant, and to replace the video to be reviewed with a preset compliant video so that the preset compliant video can be played on the large screen display device.

[0050] In addition, to achieve the above objectives, this application also proposes a large-screen video content review device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the large-screen video content review method described above.

[0051] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the large-screen video content review method described above.

[0052] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the large-screen video content review method described above.

[0053] The one or more technical solutions proposed in this application have at least the following technical effects: by acquiring the video to be reviewed, decoupling the video to be reviewed to obtain the audio stream and multiple video frames in the video to be reviewed, and converting the audio stream into audio text; performing text compliance detection on the audio text, performing semantic compliance analysis, and obtaining the first text review result corresponding to the audio text; performing image compliance detection on the video frames, performing visual compliance analysis, and obtaining the image review result corresponding to the video frames; adopting a fusion decision mechanism that intercepts the entire video if any modal violation occurs, if either the first text review result or the image review result is non-compliant, the video review result of the video to be reviewed is determined to be non-compliant, and a preset compliant video is used to replace the video to be reviewed, ensuring the security of the content played on the large screen, solving the technical problem of insufficient recognition accuracy caused by single-modal detection in the prior art, achieving comprehensive coverage of multimodal risks, avoiding missed judgments caused by a single detection path, thereby improving the robustness and reliability of large-screen video content review and ensuring the security of publicly displayed content. This application achieves audio-video separation processing by decoupling the video to be reviewed into an audio stream and multiple video frames, avoiding the information coverage blind spots of single-modal detection. After the audio stream is converted to text, text compliance detection is performed to accurately capture non-compliant content in the speech, compensating for any speech violations that might be missed by image recognition. Image compliance detection is performed on the video frames to identify sensitive elements in the image. When either text or image compliance detection detects non-compliant content, a preset compliant video replacement is triggered, forming a closed loop of audio-image dual verification in the video to be reviewed. This effectively solves the technical problems of low recognition rate and high false positive / false negative rate of single-modal detection in complex scenarios, significantly improving the accuracy and reliability of video content review. Attached Figure Description

[0054] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0055] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating an embodiment of the large-screen video content review method of this application.

[0057] Figure 2 This application provides an architecture diagram for accessing a large-screen video content review device;

[0058] Figure 3A flowchart illustrating a text review module provided in this application;

[0059] Figure 4 A flowchart illustrating an image review module provided in this application;

[0060] Figure 5 A flowchart illustrating a text similarity technique provided in this application;

[0061] Figure 6 A flowchart illustrating a text classification model provided in this application;

[0062] Figure 7 This application provides a flowchart illustrating a face similarity technology.

[0063] Figure 8 A flowchart illustrating the training process of an image classification model provided in this application;

[0064] Figure 9 This is a schematic diagram of the module structure of the large-screen video content review device according to an embodiment of this application;

[0065] Figure 10 This is a schematic diagram of the hardware operating environment involved in the large-screen video content review method in this application embodiment.

[0066] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0067] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0068] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0069] The main solution of this application embodiment is as follows: acquire the video to be reviewed, decouple the video to be reviewed to obtain the audio stream and multiple video frames in the video to be reviewed, and convert the audio stream into audio text; perform a text compliance detection operation on the audio text to obtain the first text review result corresponding to the audio text; perform an image compliance detection operation on the video frames to obtain the image review result corresponding to the video frames; if either the first text review result or the image review result is non-compliant, determine that the video review result of the video to be reviewed is non-compliant, and replace the video to be reviewed with a preset compliant video so that the preset compliant video can be played on the large screen display device.

[0070] In this embodiment, for ease of description, the following description will focus on the large-screen video content review system as the implementing entity.

[0071] Because current mainstream review methods employ a single-modal independent analysis mechanism—image review using deep learning models to detect NSFW content and audio review using keyword matching to achieve basic auditory review—these two processes are isolated from each other. This single-modal detection mode struggles to comprehensively capture multi-dimensional violations in videos, easily leading to review loopholes due to missing information or misjudgments. Therefore, existing large-screen video content review technologies suffer from fragmented multi-modal information processing, making it difficult for review systems to comprehensively and accurately identify sensitive cross-modal content. This deficiency can directly impact the security and reliability of large-screen content distribution, especially in highly sensitive locations such as transportation hubs and commercial centers, where the spread of violations through a single modality could trigger significant social impact.

[0072] This application provides a solution that decouples the video to be reviewed into an audio stream and multiple video frames, achieving audio-video separation processing. This avoids the information coverage blind spots of single-modal detection. After the audio stream is converted into text, text compliance detection is performed to accurately capture non-compliant content in the speech, compensating for any speech violations that image recognition might miss. Image compliance detection is performed on the video frames to identify sensitive elements in the image. When either text compliance detection or image compliance detection detects non-compliant content, a preset compliant video replacement is triggered, forming a closed loop of audio-image dual verification in the video to be reviewed. This effectively solves the technical problems of low recognition rate and high false positive / false negative rate of single-modal detection in complex scenarios, significantly improving the accuracy and reliability of video content review.

[0073] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a large-screen video content review device capable of performing the above functions. The following description uses a large-screen video content review system as an example to illustrate this embodiment and the subsequent embodiments.

[0074] Based on this, the embodiments of this application provide a method for reviewing large-screen video content, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the large-screen video content review method of this application.

[0075] In this embodiment, the large-screen video content review method includes steps 101-104:

[0076] Step 101: Obtain the video to be reviewed, decouple the video to be reviewed to obtain the audio stream and multiple video frames in the video to be reviewed, and convert the audio stream into audio text.

[0077] Specifically, the video to be reviewed is the original video signal that will be played on a public large screen (large-screen display device) and needs to undergo security review. This signal originates from a control device, such as a computer or player. The video format can be HDMI, VGA, or Real-Time Streaming Protocol (RTSP). The audio stream is continuous raw sound data decoupled from the video to be reviewed. It can be encoded in Pulse Code Modulation (PCM), Advanced Audio Coding (AAC), or MPEG-1 Audio Layer III (MP3) formats. The audio text can be converted from speech in the audio stream into readable text strings using Automatic Speech Recognition (ASR) technology for subsequent text compliance checks.

[0078] In some embodiments, reference Figure 2 This application embeds a large-screen video content review system into a large-screen video content review device. The large-screen video content review device is deployed between the control terminal device and the large-screen display device. The large-screen video content review device has video input and output interfaces. The control terminal device is connected to the input port of the large-screen video content review device via HDMI or VGA. The output port of the large-screen video content review device is then connected to the large-screen display device. When the video to be reviewed passes through the large-screen video content review device, it is subjected to real-time compliance detection to obtain the video review result of the video to be reviewed. Based on whether the video review result is compliant, it is decided whether to release the video to be reviewed or to replace the video to be reviewed.

[0079] Optionally, the video to be reviewed can be acquired from the control device via an HDMI capture card. This video is a composite stream containing both audio and video data. Subsequently, a multimedia processing engine (such as FFmpeg) can decapsulate the video, performing audio-video separation to obtain independent audio and video streams. For the video stream, the system extracts multiple static images according to a preset frame extraction strategy, forming a frame sequence for subsequent image review. The preset frame extraction strategy can be to extract 1 frame per second, extract keyframes, or use a fixed sampling rate (such as 5 frames per second). For the audio stream, automatic speech recognition is performed. For example, a pre-trained speech recognition model can be used to convert continuous speech signals (audio stream) into corresponding text content, i.e., audio text. The pre-trained speech recognition model can be Whisper, DeepSpeech, or the iFlytek engine, etc. The audio text retains the semantic information in the audio stream and can be used for text compliance checks such as illegal word detection and semantic analysis. By decoupling the video to be reviewed and converting it into speech, the system achieves structured analysis of multimodal content (the video to be reviewed). This decomposes complex video data into separately processable image and text formats, providing standardized input for subsequent parallel image and text compliance checks. This forms the foundation for multimodal fusion review. By separating the audio and video of the video to be reviewed, the audio and text provide input for text compliance checks, while the video frames provide the basis for image compliance checks (such as identifying inappropriate scenes). This forms the underlying data support for a text-image dual-modal review, effectively compensating for the information coverage blind spots of single-modal detection and improving the ability to capture sensitive content and the reliability of the review process in complex scenarios.

[0080] Step 102: Perform a text compliance check on the audio text to obtain the first text review result corresponding to the audio text.

[0081] Specifically, the first text review result is an independent review output for the audio text, which can be compliant or non-compliant, and can include the type of violation (such as inappropriate content, confidential information, spam) and confidence level. The audio text undergoes text compliance detection. Specifically, firstly, the audio text can be preprocessed, including removing redundant symbols, word segmentation, and standardization. Then, a deep learning-based text classification model (such as a fine-tuned BERT model) is used to perform semantic analysis on the text to determine whether the audio text belongs to the category of inappropriate content, confidential information, spam, or normal text, thus obtaining the first classification result for the audio text, determining whether the audio text belongs to the preset violation category such as inappropriate content, confidential information, or spam. Simultaneously, keyword matching algorithms (such as BM25) and vector similarity comparison technology are used to detect whether other custom-defined violation text exists in the audio text. By comprehensively analyzing multi-dimensional results (classification results and similarity results), the system generates the first text review result corresponding to the audio text, outputs a judgment conclusion of compliance or non-compliance, and can attach the type of violation and confidence level. This can effectively identify potential risky content transmitted through voice in the video to be reviewed, provide an independent judgment basis for audio modality for subsequent fusion decisions, and ensure the comprehensiveness and accuracy of the review.

[0082] In some embodiments, the step of performing a text compliance check on the audio text to obtain a first text review result corresponding to the audio text includes:

[0083] Calculate the overall similarity between the audio text and the preset illegal text, and determine the first component result of the first text review result corresponding to the audio text based on the magnitude of the overall similarity.

[0084] The audio text is input into the text classification model, and the audio text is classified by the text classification model to obtain the first classification result corresponding to the audio text. The first classification result is used to indicate whether the category of the audio text is a preset violation category.

[0085] Based on the first classification result, determine the second component result of the first text review result corresponding to the audio text;

[0086] If either the first component result or the second component result of the first text review is non-compliant, the first text review result is determined to be non-compliant.

[0087] If both the first component result and the second component result of the first text review are compliant, the first text review result is determined to be compliant.

[0088] Specifically, the preset violation texts are either built into the system or user-defined violation text corpora, containing various high-risk statements and their variations. Users can also upload industry-specific violation text libraries as preset violation texts, and the preset text library is updated periodically in the cloud. The comprehensive similarity score is a weighted score calculated by combining semantic vector similarity and keyword matching, used to measure the closeness between the audio text and the preset violation texts. The first component of the first text review result is a sub-result based on the comprehensive similarity judgment, reflecting whether the audio text is similar in form or meaning to known violation expressions (i.e., preset violation texts). The text classification model is a finely tuned deep learning model (such as BERT) used to classify the input text into preset categories such as inappropriate misleading statements, confidential information, spam, and normal text. The first classification result is the category label and its confidence score output by the text classification model, indicating whether the input text belongs to the preset violation category, i.e., whether the input text belongs to the preset violation category such as inappropriate misleading statements, confidential information, or spam. The second component of the first text review result is a sub-result based on the output of the text classification model.

[0089] As an example, see reference Figure 3The system acquires the audio text obtained from the audio stream through speech recognition and preprocesses it, including removing punctuation, standardizing capitalization, word segmentation, and stop word filtering, to improve the accuracy of subsequent analysis. Then, the system uses a dual-channel fusion strategy to generate the first text review result. The first channel is a comprehensive similarity detection: the system compares the preprocessed audio text with various preset illegal texts in a preset illegal text library stored locally or in the cloud at the semantic and keyword levels, generating a comprehensive similarity value. If the comprehensive similarity value exceeds a preset threshold (e.g., 0.8), the audio text is determined to be highly similar to known illegal content (i.e., preset illegal text) in semantics or expression, and the first component result of the first text review result corresponding to the audio text is determined to be non-compliant; otherwise, it is compliant. The second channel is model classification detection: The system inputs the audio text into a text classification model fine-tuned with a large amount of labeled data. This model is trained to identify preset categories such as "normal text," "inappropriate misleading statements," "confidential information," and "spam," and outputs the probability distribution of the audio text's category, obtaining the first classification result. If the category with the highest probability corresponds to any preset violation category and the confidence level is higher than a set threshold (e.g., 0.75), the audio text is determined to belong to the category of violation semantics. Based on this first classification result, the second component result of the first text review is determined to be non-compliant; otherwise, it is compliant. Finally, the two component results (the first component result and the second component result of the first text review) are logically fused. If either the first or second component result of the first text review is non-compliant, the first text review result is ultimately determined to be non-compliant, indicating that the audio text poses a risk of dissemination; only when both component results are compliant is the first text review result determined to be compliant. By employing a dual-channel fusion mechanism to perform in-depth compliance testing on audio text, and combining semantic similarity technology with model classification, more comprehensive and robust review results are generated. This avoids omissions by a single method and helps improve the ability to identify hidden and sensitive content such as variations, metaphors, and homophones.

[0090] Step 103: Perform image compliance detection on the video frame to obtain the image review result corresponding to the video frame.

[0091] Specifically, the image review result is a structured result output after image compliance detection, indicating whether the corresponding video frame image is compliant, whether it contains violation type, confidence level, violation location (such as intra-frame coordinates), etc.

[0092] Optionally, the decoupled multi-frame video data can be subjected to image compliance detection. First, an image classification model can be used to determine whether each video frame belongs to a preset sensitive category such as inappropriate guidance or spam. Simultaneously, Optical Character Recognition (OCR) technology can be used to identify the text content appearing in each video frame, and text compliance detection can be performed on the text content to detect whether it contains prohibited words or illegal semantics. Furthermore, a face detection model or face feature extraction model, such as the deep learning-based FaceNet face recognition network, can be used to identify faces in each video frame and compare their similarity with preset sensitive individuals. The system integrates the above multi-dimensional analysis results. If any detection path determines that the video frame contains prohibited content, the corresponding image review result is output as non-compliant; otherwise, it is compliant. By identifying whether there is visually prohibited content such as inappropriate guidance, sensitive individuals, or illegal slogans in the static images obtained from video frames, and generating clear compliance judgment results, harmful images can be prevented from being played on public screens.

[0093] In some embodiments, the step of performing image compliance detection on video frames to obtain the image review result corresponding to the video frames includes:

[0094] Extract the image of the object to be reviewed from the video frame, and calculate the similarity between the object to be reviewed and the preset sensitive objects based on the image of the object to be reviewed. Based on the magnitude of the similarity, determine the first component result of the image review result corresponding to the video frame.

[0095] The video frame is input into the image classification model, and the video frame is classified by the image classification model to obtain the second classification result corresponding to the video frame. The second classification result is used to indicate whether the category of the video frame is a preset sensitive category.

[0096] Based on the second classification result, determine the second component result of the image review result corresponding to the video frame;

[0097] The video frame is input into the character recognition model for character extraction to obtain the video frame text.

[0098] Perform a text compliance check on the video frame text to obtain the second text review result corresponding to the video frame text;

[0099] If any one of the first component of the image review result, the second component of the image review result, or the second text review result is non-compliant, the image review result is determined to be non-compliant.

[0100] If the first component of the image review result, the second component of the image review result, and the second text review result are all compliant, the image review result is determined to be compliant.

[0101] Specifically, the image to be reviewed is the detected face or specific person region image in the video frame, used to compare with various preset sensitive objects in the preset sensitive object library. The preset sensitive objects are images of people (objects) to be intercepted, either pre-set by the system or uploaded by the user. The feature vectors corresponding to the preset sensitive objects are stored in the database, which supports regular updates to the preset sensitive object library in the cloud. Person similarity is calculated by mapping faces to vectors using face recognition models such as FaceNet, used to measure the degree of matching between the object to be reviewed and the preset sensitive objects. The first component of the image review result is a sub-result based on person similarity judgment, reflecting whether a preset sensitive object exists in the video frame. If a preset sensitive object exists in the video frame, the first component of the image review result is non-compliant; if no preset sensitive object exists in the video frame, the first component of the image review result is also non-compliant. The image classification model is a trained or fine-tuned deep learning model such as Residual Network (ResNet) and Visual Geometry Group Network (VGG), used to determine whether the video frame belongs to preset sensitive categories such as inappropriate guidance or spam advertising. The second component of the image review result is a sub-result based on the image classification model. The character recognition model can be a tool like PaddleOctane (PaddleOCR), used to extract visible text from video frame images.

[0102] As an example, see reference Figure 4 , Figure 4This is a flowchart illustrating an image review module provided in this application. The system performs multi-dimensional visual content security analysis frame by frame from multiple video frames obtained after decoupling processing. For person risk detection, a face detection model and a face recognition model are invoked to locate the face region in each video frame, extracting the image of the object to be reviewed containing the face. The face detection model can be a multi-task cascaded convolutional network (MTCNN) or a high-precision single-stage dense face localization model (RetinaFace), etc. Subsequently, the image of the object to be reviewed is input into a pre-trained face recognition model (such as FaceNet) to generate a feature vector corresponding to the image of the object to be reviewed. The similarity between the feature vector corresponding to the image of the object to be reviewed and all entries in a preset sensitive object feature vector database maintained locally or in the cloud is calculated to obtain the person similarity between the object to be reviewed and the preset sensitive objects. If the similarity of the person is greater than a preset threshold (e.g., 0.95), the object to be reviewed is determined to be a preset sensitive object. The video frame containing the object to be reviewed contains a sensitive person requiring control (preset sensitive object), and the first component of the image review result is determined to be non-compliant; otherwise, it is compliant. For overall content risk, the video frame is input into a trained image classification model, such as a deep neural network based on the ResNet50 architecture. This model is fine-tuned to identify preset sensitive categories such as inappropriate guidance, explosive content, and spam advertising. The image classification model outputs the probability distribution of each category, obtaining the second classification result corresponding to the video frame. If the category with the highest probability corresponds to any preset sensitive category and the confidence level is higher than a set threshold (e.g., 0.75), the video frame is determined to have an overall visual violation risk. Based on this second classification result, the second component of the image review result is determined to be non-compliant; otherwise, it is compliant. For text information hidden in the video frame, the video frame is input into a character recognition model, such as PaddleOCR, to detect and recognize text regions in the video frame and extract the text, which may include scrolling subtitles, static slogans, watermark text, etc. Subsequently, a text compliance check is performed on the extracted video frame text, specifically including text similarity technology and text classification technology based on a text classification model. If illegal content is detected in the video frame text, the second text review result is output as non-compliant; otherwise, it is compliant.Finally, the results of the three independent detection paths (the first component of the image review result, the second component of the image review result, and the second text review result) are fused for decision-making. If any one of the first component of the image review result (person recognition), the second component of the image review result (image classification), or the second text review result (OCR text detection) is non-compliant, the image review result of the video frame is determined to be non-compliant, indicating that the video frame poses a risk of dissemination. Only when all three results (the first component of the image review result, the second component of the image review result, and the second text review result) are compliant is the video frame's image review deemed to have passed. Through the above multimodal, multi-path image compliance detection operation, multi-dimensional, fine-grained image compliance detection of video frames can effectively improve the ability to identify complex and hidden visual violations, ensuring accurate interception in any situation such as the appearance of sensitive persons, violations of image content, or illegal text information, thus guaranteeing the security and compliance of content played on large screens.

[0103] Optionally, the step of performing text compliance checks on the video frame text to obtain the second text review result corresponding to the video frame text includes:

[0104] Calculate the second comprehensive similarity between the video frame text and the preset violation text, and determine the first component result of the second text review result corresponding to the video frame text based on the value of the second comprehensive similarity.

[0105] The video frame text is input into the text classification model, and the text classification model classifies the video frame text to obtain the third classification result corresponding to the video frame text. The third classification result is used to indicate whether the category of the video frame text is a preset violation category.

[0106] Based on the third classification result, the second component result of the second text review result corresponding to the video frame text is determined;

[0107] If either the first component result of the second text review result or the second component result of the second text review result is non-compliant, the second text review result is determined to be non-compliant.

[0108] If both the first component result of the second text review and the second component result of the second text review are compliant, the second text review result is determined to be compliant.

[0109] Step 104: If either the first text review result or the image review result is non-compliant, determine that the video review result of the video to be reviewed is non-compliant, and replace the video to be reviewed with a preset compliant video so that the preset compliant video can be played on the large screen display device.

[0110] If both the first text review result and the image review result are compliant, the video review result is determined to be compliant.

[0111] The video to be reviewed is sent to a large-screen display device via a preset protocol so that it can be played on the large-screen display device. The preset protocol is either HDMI or VGA.

[0112] Specifically, the preset compliant videos are built-in security video clips, such as blue skies and white clouds, public service advertisements, and static logos, used to replace the original video when the video to be reviewed fails the review. Large-screen display devices are display terminals such as LED screens and LCD splicing screens in public places, used to play video content.

[0113] In some embodiments, after completing independent compliance checks on audio text and video frames, it is determined whether both the first text review result and the image review result are compliant. If both are compliant, the video review result of the video to be reviewed is determined to be compliant, the channel between the control device and the large-screen display device is kept open, and the video to be reviewed is sent to the large-screen display device through a preset protocol. If either the first text review result or the image review result is non-compliant, the overall video review result is determined to be non-compliant. Subsequently, corresponding operations can be performed through the hardware video switching module. When the video review result of the video to be reviewed is compliant, the video to be reviewed is transmitted to the large-screen display device through a preset protocol (such as HDMI or VGA) to ensure normal playback of the content. When the video review result of the video to be reviewed is non-compliant, the original video output is cut off, and the system switches to a preset compliant video source (such as a public service advertisement, a static blue sky image, or a safety prompt animation stored locally), and the preset compliant video is pushed to the large-screen device in real time through the same protocol to achieve seamless replacement and avoid black screen or signal interruption. By constructing a content security barrier and reducing the false negative rate through dual-dimensional cross-verification, we can ensure a rapid response and block the spread of risks when illegal content is found in any modality of the video to be reviewed. At the same time, we can maintain a continuous, stable, and secure visual presentation on the large screen, meeting the dual requirements of high security and high availability for content broadcasting in public places.

[0114] Based on the large-screen video content review method disclosed in this application, the method obtains the video to be reviewed, decouples the video to be reviewed to obtain the audio stream and multiple video frames in the video to be reviewed, and converts the audio stream into audio text; performs text compliance detection on the audio text and performs semantic compliance analysis to obtain the first text review result corresponding to the audio text; performs image compliance detection on the video frames and performs visual compliance analysis to obtain the image review result corresponding to the video frames; and adopts a fusion decision mechanism that intercepts the entire video if any modal violation occurs. If either the first text review result or the image review result is non-compliant, the video review result of the video to be reviewed is determined to be non-compliant, and a preset compliant video is used to replace the video to be reviewed, ensuring the security of the content played on the large screen. This method solves the technical problem of insufficient recognition accuracy caused by single-modal detection in the prior art, achieves comprehensive coverage of multimodal risks, avoids missed judgments caused by a single detection path, and thus improves the robustness and reliability of large-screen video content review, ensuring the security of publicly displayed content. This application achieves audio-video separation processing by decoupling the video to be reviewed into an audio stream and multiple video frames, avoiding the information coverage blind spots of single-modal detection. After the audio stream is converted to text, text compliance detection is performed to accurately capture non-compliant content in the speech, compensating for any speech violations that might be missed by image recognition. Image compliance detection is performed on the video frames to identify sensitive elements in the image. When either text or image compliance detection detects non-compliant content, a preset compliant video replacement is triggered, forming a closed loop of audio-image dual verification in the video to be reviewed. This effectively solves the technical problems of low recognition rate and high false positive / false negative rate of single-modal detection in complex scenarios, significantly improving the accuracy and reliability of video content review.

[0115] In some embodiments, the step of calculating the overall similarity between audio text and preset illegal text includes:

[0116] The audio text and the preset illegal text are extracted using a text classification model to obtain feature vectors for the audio text and the preset illegal text, and the cosine similarity between the feature vectors of the audio text and the preset illegal text is calculated.

[0117] The keyword matching degree between the audio text and the preset illegal text is calculated using the best matching 25BM25 algorithm;

[0118] The comprehensive similarity between the audio text and the preset illegal text is obtained by weighted summation based on cosine similarity and keyword matching degree.

[0119] Specifically, the text classification model can be a fine-tuned BERT pre-trained language model, which can be used to extract high-dimensional semantic feature vectors of text.

[0120] As an example, see reference Figure 5 The audio text to be detected and each sensitive sentence (preset illegal text) in the preset illegal text library are input into a finely tuned text classification model. The text classification model is used to extract deep semantic features and output the corresponding high-dimensional semantic feature vectors, thus obtaining the audio text feature vector and the preset illegal text feature vector. Then, the cosine similarity between the audio text feature vector and each preset illegal text feature vector is calculated. The cosine similarity can be used to measure the closeness of the two in terms of semantic intent. Next, the BM25 algorithm is used to perform keyword-level matching between the audio text and the preset illegal text. The keyword matching degree is calculated using parameters such as word frequency and inverse document frequency to identify the occurrence of explicit illegal words or high-frequency illegal combinations. It is especially suitable for detecting variant expressions such as homophones, abbreviations, and word splitting, thus obtaining the keyword matching degree between the audio text and the preset illegal text. The scores of the above two dimensions are weighted and fused. A linear weighted method can be used to calculate the comprehensive similarity. The formula is: Comprehensive similarity = α × cosine similarity + (1-α) × keyword matching degree, where the weight coefficient α can be set to 0.7 according to the actual scenario to highlight the importance of semantic analysis. The obtained comprehensive similarity value serves as an important basis for subsequent determination of whether audio content violates regulations. It effectively combines semantic understanding ability with keyword matching accuracy, improves the system's ability to identify covert and adversarial sensitive expressions, and enhances the robustness and accuracy of text compliance detection.

[0121] In some embodiments, reference Figure 6 Before the step of inputting the audio text into the text classification model, the following steps are also included:

[0122] Step 601: Collect illegal text data, which includes multiple illegal texts;

[0123] Step 602: Input the violation text and multiple preset violation categories into the large language model, perform category matching on the violation text, and obtain the text category of the violation text;

[0124] Step 603: Based on each violating text and its text category, construct a text classification training set, where the text category corresponding to the violating text is the label of the violating text;

[0125] Step 604: Fine-tune the BERT-based model to be trained using the text classification training set to obtain the text classification model.

[0126] Specifically, the model to be trained is based on the BERT model, with the addition of a fully connected linear layer and a softmax activation function, used to transform the model into a multi-classifier. The violation text data consists of real violation statements collected from the internet, historical logs, and report information, used for model training. The preset violation categories are predefined classification systems for violation content, such as inappropriate misleading statements, confidential information, and spam. The large language model can be a powerful semantic understanding model such as Qwen (Question and Answer with English / Chinese), ChatGenerative Pre-trained Transformer (ChatGPT), or DeepSeek-v3, which can be used to automatically determine the category of the text. Category matching involves semantically comparing each violation text with the preset categories to determine its most likely violation type. The text classification training set contains a dataset of violation text-label pairs for supervised learning, where the labels represent the text categories.

[0127] As an example, a large amount of real-world illegal text data is collected through legal channels, including historical review records from public online platforms, sensitive statements from public safety databases, and user reports, forming an initial corpus. This illegal text data covers various risky expressions. Subsequently, each illegal text in the illegal text data, along with multiple preset violation categories (such as inappropriate content, confidential information, and spam), is input into a large language model. Reasonable prompts are designed to guide the large language model in semantic analysis and category matching. For example, the input might be: "Please determine which category the following text belongs to: 'Gather in the square tomorrow.' Possible categories: inappropriate content, confidential information, normal text, spam. Output only the category name." Based on contextual understanding and knowledge reasoning, the large language model outputs the most likely text category to which the illegal text belongs. Each illegal text is paired with its corresponding text category to construct a structured text classification training set, where the illegal text is the input sample and the category serves as the supervision label. This text classification training set effectively solves the problems of high cost and inconsistent standards of manual annotation, while ensuring the semantic accuracy and consistency of the annotation results. Next, supervised fine-tuning of the BERT-based model was performed using a text classification training set. During training, the model received inappropriate text input, extracted features through the encoder, output a classification probability distribution, and calculated the cross-entropy loss with the ground truth labels. Model parameters were then optimized through backpropagation. After multiple rounds of iterative training, the model gradually learned the semantic patterns of various sensitive content types. Ultimately, a text classification model with high-precision recognition capabilities was obtained, accurately determining whether newly input audio text belongs to a preset inappropriate category, providing reliable technical support for subsequent compliance detection of audio content. Deployed in the review system, this model can generate primary classification results, significantly improving the ability to identify hidden and variant-sensitive semantics.

[0128] In some embodiments, the step of calculating the similarity between the object to be reviewed in the image of the object to be reviewed and a preset sensitive object includes:

[0129] The facial features of the image of the object to be reviewed are extracted by a facial recognition model to obtain the facial feature vector of the object to be reviewed.

[0130] Obtain the facial feature vector of the preset sensitive object. The preset sensitive object facial feature vector is obtained by extracting facial features from the image of the preset sensitive object containing the preset sensitive object through a face recognition model.

[0131] Based on the facial feature vector of the object to be reviewed and the facial feature vector of the preset sensitive object, the similarity between the object to be reviewed and the preset sensitive object is determined.

[0132] Specifically, the facial recognition model can be a deep neural network model such as FaceNet or ArcFace, which maps facial images into high-dimensional feature vectors to achieve identity recognition. Preset sensitive targets can be individuals that the system pre-configures for monitoring. Users can also customize preset sensitive targets and upload their images. The facial feature vector of a preset sensitive target is a feature vector extracted from the image of the preset sensitive target using the same facial recognition model, and is stored in a local or cloud database.

[0133] As an example, see reference Figure 7 The system extracts images of the subjects to be reviewed, containing faces, from video frames and inputs them into a face recognition model, such as FaceNet (a unified embedding model for face recognition and clustering) or ArcFace (an arc loss face recognition model). The face recognition model performs deep feature extraction on the image of the subject to be reviewed, outputting a fixed-dimensional numerical vector, i.e., the facial feature vector of the subject to be reviewed, used to characterize the unique biometric features of the subject's face. The system pre-processes clear, frontal facial images (such as ID photos or authorized images) of multiple preset sensitive subjects (such as specific personnel requiring control) using the same face recognition model, extracting the corresponding facial feature vectors of the preset sensitive subjects and storing them in a local security database, forming a preset sensitive subject facial feature vector library. During the comparison phase, the facial feature vector of the subject to be reviewed is compared with the facial feature vectors of all preset sensitive subjects in the preset sensitive subject facial feature vector library for similarity calculation. A cosine similarity algorithm can be used to measure the angle between the directions of the two vectors, with a value range of [-1, 1]. The closer the value is to 1, the more similar the faces are. The highest similarity value is taken as the final person similarity result. This mechanism can trigger timely review alerts when high-risk individuals are detected, providing crucial information for subsequent interception and replacement of video content, thereby enhancing the security and controllability of content played on large screens.

[0134] In some embodiments, reference Figure 8 Before the step of inputting video frames into the image classification model, the following steps are also included:

[0135] Step 801: Collect sensitive image data, which includes multiple sensitive images;

[0136] Step 802: Input multiple preset sensitive categories and sensitive images into the visual language large model, perform category matching on the sensitive images, and obtain the image category of the sensitive images;

[0137] Step 803: Based on each sensitive image and its image category, generate an image classification training set and an image classification test set;

[0138] Step 804: Train multiple image classification models to be trained using the image classification training set to obtain multiple trained candidate image classification models.

[0139] Step 805: Test the classification accuracy of each candidate image classification model using an image classification test set, and select the candidate image classification model with the highest classification accuracy from multiple candidate image classification models to determine it as the image classification model.

[0140] Specifically, the image classification model to be trained can be VGG, ResNet, or the efficient neural network EfficientNet, and has not yet been trained for a specific task. The large visual language model can be a large model with image-text joint understanding capabilities, such as DeepSeek-v3 or internvl3, which can map images to a semantic space and perform classification. Sensitive image data refers to collections of images containing sensitive content such as inappropriate guidance or spam (e.g., flags, weapons, indecent images), and can originate from publicly available secure databases, web crawlers (after anonymization), or manually generated images. Predefined sensitive categories can be a predefined risk classification system, such as inappropriate guidance, spam, or normal content.

[0141] As an example, a large amount of sensitive image data, including images of inappropriate content, terrorist flags, and sensitive slogans, was collected through legal and compliant channels to form an initial image corpus. Each image was initially unlabeled and required automated labeling. Subsequently, each sensitive image, along with several preset sensitive categories (such as inappropriate content, normal content, etc.), was input into a visual language model. This model, possessing image-text joint understanding capabilities, could automatically determine the most likely category based on the alignment between image content and category semantics. By designing appropriate prompts, such as "Please determine which category the following image belongs to: [image]. Optional categories: inappropriate content, normal content. Output only the category name.", the visual language model outputs the corresponding image category, achieving efficient and accurate automatic labeling. Each sensitive image and its labeling result were combined into a "(image, label)" sample to form a structured image dataset, which was then proportionally divided into an image classification training set (e.g., 70%) and an image classification test set (e.g., 30%) to ensure no overlap between training and evaluation data. Next, various deep learning architectures were selected as training models for image classification, such as ResNet, VGG, EfficientNet, and MobileNet (a lightweight neural network for mobile devices). Each model was then fine-tuned using a supervised training set. Inputting an image, the model outputted a predicted category, and the parameters were optimized using the cross-entropy loss function until convergence. After fine-tuning, several well-trained candidate image classification models were obtained. Subsequently, the performance of each candidate model was evaluated using an image classification test set, calculating metrics such as classification accuracy and F1-score. The candidate model with the highest classification accuracy was selected and chosen as the image classification model for deployment in the large-screen video content review system. This model is used for real-time risk identification of video frames, ensuring high accuracy and robustness in the image review process.

[0142] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the large-screen video content review method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0143] This application also provides a large-screen video content review device; please refer to [reference needed]. Figure 9 The large-screen video content review device includes:

[0144] The image text extraction module 901 is used to acquire the video to be reviewed, decouple the video to be reviewed, obtain the audio stream and multiple video frames in the video to be reviewed, and convert the audio stream into audio text.

[0145] The text review module 902 performs a text compliance check on the audio text and obtains the first text review result corresponding to the audio text.

[0146] The image review module 903 performs image compliance detection on video frames and obtains the image review results corresponding to the video frames.

[0147] The determination module 904 is used to determine that the video review result of the video to be reviewed is non-compliant if either the first text review result or the image review result is non-compliant, and to replace the video to be reviewed with a preset compliant video so that the preset compliant video can be played on the large screen display device.

[0148] The large-screen video content review device provided in this application, employing the large-screen video content review method described in the above embodiments, can solve the technical problem of insufficient recognition accuracy caused by single-modal detection in the prior art. Compared with the prior art, the beneficial effects of the large-screen video content review device provided in this application are the same as those of the large-screen video content review method provided in the above embodiments, and other technical features in the large-screen video content review device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0149] This application provides a large-screen video content review device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the large-screen video content review method in the above embodiment 1.

[0150] The following is for reference. Figure 10 The diagram illustrates a structural schematic suitable for implementing a large-screen video content review device according to embodiments of this application. The large-screen video content review device in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, tablets (Portable Application Description, PADs), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 10 The large-screen video content review device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0151] like Figure 10As shown, the large-screen video content review device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The random access memory 1004 also stores various programs and data required for the operation of the large-screen video content review device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the large-screen video content review device to communicate wirelessly or wiredly with other devices to exchange data. Although a large-screen video content review device with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems may be implemented alternatively.

[0152] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0153] The large-screen video content review device provided in this application, employing the large-screen video content review method described in the above embodiments, can solve the technical problem of insufficient recognition accuracy caused by single-modal detection in the prior art. Compared with the prior art, the beneficial effects of the large-screen video content review device provided in this application are the same as those of the large-screen video content review method provided in the above embodiments, and other technical features in this large-screen video content review device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0154] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0155] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0156] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the large-screen video content review method described in the above embodiments.

[0157] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0158] The aforementioned computer-readable storage medium may be included in the large-screen video content review device; or it may exist independently and not be assembled into the large-screen video content review device.

[0159] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the large-screen video content review device, the large-screen video content review device: acquires the video to be reviewed; decouples the video to be reviewed to obtain the audio stream and multiple video frames in the video to be reviewed; and converts the audio stream into audio text; performs a text compliance detection operation on the audio text to obtain a first text review result corresponding to the audio text; performs an image compliance detection operation on the video frames to obtain an image review result corresponding to the video frames; if either the first text review result or the image review result is non-compliant, the video review result of the video to be reviewed is determined to be non-compliant, and a preset compliant video is used to replace the video to be reviewed, so that the preset compliant video is played on the large-screen display device.

[0160] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0161] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0162] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0163] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the above-described large-screen video content review method, which can solve the technical problem of insufficient recognition accuracy caused by single-modal detection in the prior art. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the large-screen video content review method provided in the above embodiments, and will not be repeated here.

[0164] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the large-screen video content review method described above.

[0165] The computer program product provided in this application can solve the technical problem of insufficient recognition accuracy caused by single-modal detection in the prior art. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the large-screen video content review method provided in the above embodiments, and will not be repeated here.

[0166] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for screening large-screen video content, characterized in that, The large-screen video content review method comprises: obtaining a video to be reviewed, decoupling the video to be reviewed to obtain an audio stream and multiple video frames in the video to be reviewed, and converting the audio stream into an audio text; performing a text compliance detection operation on the audio text to obtain a first text review result corresponding to the audio text; performing an image compliance detection operation on the video frames to obtain an image review result corresponding to the video frames; if any one of the first text review result and the image review result is non-compliant, determining that a video review result of the video to be reviewed is non-compliant, and replacing the video to be reviewed with a preset compliant video to enable the preset compliant video to be played on a large-screen display device; wherein the step of performing a text compliance detection operation on the audio text to obtain a first text review result corresponding to the audio text comprises: calculating a comprehensive similarity between the audio text and a preset illegal text, determining a first component result of the first text review result corresponding to the audio text based on a numerical value of the comprehensive similarity; inputting the audio text into a text classification model, classifying the audio text through the text classification model to obtain a first classification result corresponding to the audio text, the first classification result being used to indicate whether a category of the audio text is a preset illegal category; determining a second component result of the first text review result corresponding to the audio text based on the first classification result; if any one of the first component result of the first text review result and the second component result of the first text review result is non-compliant, determining that the first text review result is non-compliant; if both the first component result of the first text review result and the second component result of the first text review result are compliant, determining that the first text review result is compliant; wherein the step of performing an image compliance detection operation on the video frames to obtain an image review result corresponding to the video frames comprises: extracting an object image to be reviewed in the video frames, calculating a person similarity between an object to be reviewed in the object image to be reviewed and a preset sensitive object based on the object image to be reviewed, and determining a first component result of the image review result corresponding to the video frames based on a numerical value of the person similarity; inputting the video frames into an image classification model, classifying the video frames through the image classification model to obtain a second classification result corresponding to the video frames, the second classification result being used to indicate whether a category of the video frames is a preset sensitive category; determining a second component result of the image review result corresponding to the video frames based on the second classification result; inputting the video frames into a character recognition model to extract characters to obtain a video frame text in the video frames; performing a text compliance detection operation on the video frame text to obtain a second text review result corresponding to the video frame text; If any one of the first component result of the image review result, the second component result of the image review result, and the second text review result is the non-compliance, it is determined that the image review result is the non-compliance. If all of the first component result of the image review result, the second component result of the image review result, and the second text review result are the compliance, it is determined that the image review result is the compliance. 2.The method of claim 1, wherein, The step of calculating the comprehensive similarity between the audio text and the preset rule violation text comprises: extracting features of the audio text and the preset rule violation text through the text classification model to obtain an audio text feature vector and a preset rule violation text feature vector, and calculating a cosine similarity between the audio text feature vector and the preset rule violation text feature vector; calculating a keyword matching degree between the audio text and the preset rule violation text through a best matching BM25 algorithm; performing weighted summation based on the cosine similarity and the keyword matching degree to obtain the comprehensive similarity between the audio text and the preset rule violation text. 3.The method of claim 1, wherein, Before the step of inputting the audio text into the text classification model, the method further comprises: collecting rule violation text data, the rule violation text data comprising a plurality of rule violation texts; inputting the rule violation texts and a plurality of preset rule violation categories into a large language model to perform category matching on the rule violation texts to obtain text categories of the rule violation texts; based on each of the rule violation texts and the text category of each of the rule violation texts, constructing a text classification training set, wherein the text category corresponding to the rule violation text is a label of the rule violation text; fine-tuning a to-be-trained model based on a BERT model architecture through the text classification training set to obtain the text classification model. 4.The method of claim 1, wherein, The step of calculating the person similarity between the to-be-reviewed object and the preset sensitive object based on the to-be-reviewed object image comprises: extracting face features of the to-be-reviewed object image through a face recognition model to obtain a to-be-reviewed object face feature vector corresponding to the to-be-reviewed object image; obtaining a preset sensitive object face feature vector, the preset sensitive object face feature vector being obtained by extracting face features of a preset sensitive object image containing the preset sensitive object through the face recognition model; based on the to-be-reviewed object face feature vector and the preset sensitive object face feature vector, determining the person similarity between the to-be-reviewed object and the preset sensitive object. 5.The method of claim 1, wherein, Before the step of inputting the video frame into the image classification model, the method further comprises: collecting sensitive image data, the sensitive image data comprising a plurality of sensitive images; inputting a plurality of preset sensitive categories and the sensitive images into a visual language large model to perform category matching on the sensitive images to obtain image categories of the sensitive images; based on each of the sensitive images and the image category of each of the sensitive images, generating an image classification training set and an image classification test set; training a plurality of to-be-trained image classification models through the image classification training set to obtain a plurality of trained candidate image classification models; The classification accuracy of each candidate image classification model is tested by the image classification test set respectively, and a candidate image classification model with the highest classification accuracy is selected from the plurality of candidate image classification models to determine the image classification model. 6.The method of claim 1, wherein, After the step of obtaining the image review result corresponding to the video frame, the method further comprises: If the first text review result and the image review result are both compliant, the video review result is determined as the compliant; The to-be-reviewed video is sent to the large-screen display device through a preset protocol, so that the to-be-reviewed video is played on the large-screen display device, and the preset protocol is a high-definition multimedia interface (HDMI) protocol or a video graphics array (VGA) protocol.

7. A large-screen video content review device, characterized by, The device comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, and the computer program is configured to implement the steps of the large-screen video content review method according to any one of claims 1 to 6.

8. A storage medium, characterized by The storage medium is a computer readable storage medium, and the storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the large-screen video content review method according to any one of claims 1 to 6.

Images