A short video release information detection method, system, device and medium

Abstract

The present invention discloses a short video release information detection method, system, device and medium, which relate to the field of artificial intelligence and social network technology. The method comprises: collecting short video information data, extracting multimodal features of the short video data; fusing the multimodal features to obtain a unified representation of the short video features; extracting short video key frames; using a large visual-language model to judge the authenticity of the short video from the perspectives of video style, video content and information matching, and generating corresponding reasoning basis and confidence level; interactively enhancing the reasoning basis and the short video features; evaluating the importance of different reasoning basis according to the confidence level; and obtaining features for final classification based on the short video features and reasoning basis after the importance aggregation enhancement. The method captures the mutual relationship and supplementary information between different modes, strengthens the perception of the video information detector to the auxiliary information, and improves the detection accuracy.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and social network technology, and in particular to a short video release information detection method, system, device and medium. Background Art

[0002] Short video platforms have rapidly risen around the world. While promoting cultural dissemination and economic development, they have also brought the risk of widespread dissemination of false information. Therefore, designing a false information detection method suitable for short video platforms to mitigate the negative impact is an urgent problem to be solved.

[0003] At present, it is usually started from the perspective of multimodal content matching, such as entity enhancement and semantic matching. In recent years, with the popularity of TikTok around the world, false information in the form of short videos has begun to show strong harm. To this end, researchers use cross-modal attention mechanisms to fuse video multimodal features to detect false information videos. Researchers also proposed knowledge-based prompt learning, integrating external knowledge into prompt representation as auxiliary information, but the process of collecting external knowledge is often very cumbersome, and it is difficult to ensure the quality of the collection results. In recent years, large visual language models (LLMs) have begun to be applied to work related to information detection due to their powerful natural language processing capabilities.

[0004] Although the above methods have achieved certain results, existing methods that only use large language models often lack fine-tuning, making it difficult to perceive the fine-grained clues of the multimodal content of short video information, and thus generate inaccurate or misleading information (i.e., the hallucination generation phenomenon). Summary of the invention

[0005] In view of the deficiency that the prior art has difficulty in perceiving the fine-grained clues of the multimodal content of short video information, resulting in the generation of inaccurate or misleading information, the present invention proposes a short video release information detection method, system, device and storage medium. Through prompt word engineering, the LVLM is enabled to intelligently understand the semantics of short video information and generate reasoning basis and confidence level. Through key information selection and learning modules, the negative impact of the large model hallucination generation problem is reduced, and an effective connection between the LVLM and a small model in a specific field is established, thereby solving the problems existing in the prior art.

[0006] A short video release information detection method comprises the following steps:

[0007] Collecting information data released by short videos and extracting multimodal features of the information data released by short videos; the multimodal features include text features, audio features and visual features; fusing the text features, audio features and visual features of the information data released by short videos to obtain a unified representation of short video features;

[0008] Extract key frames from short videos, combine key frames with prompt word engineering including role-playing, context learning and thought chain, and construct multimodal prompt words; use multimodal prompt words to drive the large visual language model LVLM to generate reasoning basis and confidence level for judging the authenticity of short videos;

[0009] Extract features from the inference basis, and interactively enhance the extracted features with the short video features after unified representation; evaluate the importance of different inference bases according to the confidence level;

[0010] Based on the importance, the enhanced short video features and reasoning basis features are aggregated; by inputting the aggregated features into a multi-layer perceptron, the information released by the short video is detected.

[0011] Furthermore, BERT is used to extract text features. , extract audio features using the pre-trained VGGish model , using the deep neural network VGG19 to extract static visual frame features from visual features , using C3D model to extract dynamic visual video clip features from visual features .

[0012] Furthermore, the text features, audio features and visual features of the information data released by the short video are integrated to obtain a unified representation of the short video features: :

[0013] ,

[0014] in, They are the text features, audio features, static visual frame features, and dynamic visual video clip features of the short video release information data enhanced by the co-attention mechanism. Trm represents the Transformer encoder layer, AP Represents average pooling.

[0015] Furthermore, the key frames of the short video are extracted by sampling the key frames at time intervals, and the sampling process is expressed as follows:

[0016] ,

[0017] in, is a collection of key frames; Indicates the first j frame; s is the sampling interval; t is the index of the key frame, ranging from 0 to , N Indicates the total number of frames in a short video.

[0018] Furthermore, the reasoning basis and the short video features are interactively enhanced, and the enhancement process is respectively expressed as follows:

[0019] ,

[0020] in represents the co-attention mechanism, Respectively represent i The initial features of the reasoning basis and the features after interactive enhancement, Indicates that you have learned i The video features obtained after the inference basis are Indicates enhanced video features.

[0021] Furthermore, the method of aggregating the enhanced short video features and the reasoning basis based on the importance to obtain the features for the final classification specifically includes the following steps:

[0022] Importance It is expressed as:

[0023] ,

[0024] in σ Indicates the control quantitative trend, b Indicates the control quantization range, θ represents the hyperparameters for which inference is based on the average level, Indicates the confidence level corresponding to the inference basis R.

[0025] After quantifying the importance of the reasoning basis, it is expressed as :

[0026] ,

[0027] in, They are three reasoning evidence sets outputted from the perspectives of video style, video content, and information matching using LVLM;

[0028] Multimodal features for video enhancement Aggregation to obtain the final aggregate features for classification :

[0029] .

[0030] Furthermore, the information released by the short video is detected by inputting the aggregated features into a multi-layer perceptron, which specifically includes the following steps:

[0031] Aggregate features Input to the multi-layer perceptron , predict the probability that the short video release information is false :

[0032] ,

[0033] The cross entropy loss function is used to measure the difference between the probability distribution of the predicted label and the actual label. The loss function is expressed as:

[0034] ,

[0035] in, y The actual tags for short video posting information, is the predicted probability that the information released in the short video is false, 1 represents a false news video and 0 represents a true news video.

[0036] The present invention also includes a short video release information detection system, including:

[0037] A collection module is used to collect information data released by short videos and extract multimodal features of the information data released by short videos; the multimodal features include text features, audio features and visual features; the text features, audio features and visual features of the short video information data are integrated to obtain a unified representation of the short video features;

[0038] The generation module is used to extract key frames of short videos, combine the key frames with the prompt word engineering including role-playing, context learning and thought chain, and construct multimodal prompt words; use the multimodal prompt words to drive the large visual language model LVLM to generate reasoning basis and confidence level for judging the authenticity of short videos;

[0039] The enhancement module is used to extract features of the reasoning basis and interactively enhance the extracted features with the short video features after unified representation; the importance of different reasoning bases is evaluated according to the confidence level;

[0040] The detection module is used to aggregate the enhanced short video features and reasoning basis features based on importance; the information released by the short video is detected by inputting the aggregated features into a multi-layer perceptron.

[0041] The present invention also includes a short video release information detection computer device, including: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the short video release information detection method when executing the computer program.

[0042] The present invention also includes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the short video release information detection method.

[0043] The present invention provides a short video release information detection method, system, device and medium, which have the following beneficial effects:

[0044] The present invention realizes the effective fusion of multimodal features of short video information data to capture the mutual relationship and supplementary information between different modes, so as to fully understand the content of short videos; constructs multimodal prompt words through prompt word engineering, uses LVLM to understand the multimodal information of videos based on time series, performs auxiliary reasoning on short video information from the perspectives of video style, video content and information matching, obtains reasoning basis and confidence level, and evaluates the importance of generating auxiliary information based on the confidence level, thereby reducing the negative impact of the hallucination generation problem; at the same time, interactively enhances the features of the reasoning basis and the short video features after unified representation, strengthens the perception of the video information detector to the auxiliary information, better understands the reasoning basis, improves the accuracy of detection, and solves the problem that traditional methods are prone to generate inaccurate or misleading information. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a short video release information detection method according to an embodiment of the present invention;

[0046] Figure 2 A framework diagram of false news short video detection in an embodiment of the present invention;

[0047] Figure 3 A schematic diagram of the number of samples distributed in different confidence level intervals and prediction accuracy in an embodiment of the present invention;

[0048] Figure 4 Schematic diagram of the effect of changes in the hyperparameter θ on the performance of the present invention in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0050] The present invention proposes a short video release information detection method, such as Figure 1 As shown, the specific steps include:

[0051] S1. Collect information data released by short videos and extract multimodal features of the short video information data; multimodal features include text features, audio features and visual features.

[0052] S2. Fuse the text features, audio features, and visual features of the short video information data to obtain a unified representation of the short video features.

[0053] S3. Extract key frames of short videos, combine the key frames with prompt word engineering including role-playing, context learning and thinking chain, and construct multimodal prompt words.

[0054] S4. Use multimodal prompt words to drive the large visual language model LVLM to generate reasoning basis and confidence level for judging the authenticity of short videos.

[0055] S5. Extract the features of the reasoning basis and interactively enhance them with the short video features after unified representation; evaluate the importance of different reasoning bases according to the confidence level.

[0056] S6. Based on importance, aggregate the enhanced short video features and reasoning basis features.

[0057] S7. The information released by the short video is detected by inputting the aggregated features into a multi-layer perceptron.

[0058] Based on the above short video release information detection method, the present invention proposes an embodiment for detecting false news released by short videos, which specifically includes the following steps:

[0059] S1: Input news video, news video dataset Every news video instance in Format as , Represents text content, audio, keyframes, and video clips respectively.

[0060] S2: Extract the features of each modality of the short video and use BERT to encode the text into Dimensional Features , using the pre-trained VGGish model to represent the audio features as , using VGG19 to extract static visual features , using C3D to extract dynamic visual features , the number of video clips corresponds to the number of key frames.

[0061] S3: Multimodal feature fusion, the features of text, audio, frame and video clips are uniformly vectorized and represented as :

[0062] (1)

[0063] in, They are the text features, audio features, static visual frame features, and dynamic visual video clip features of the short video release information data enhanced by the co-attention mechanism. Trm represents the Transformer encoder layer, AP Represents average pooling.

[0064] S4: Multimodal prompt words are constructed through prompt word engineering including role-playing, context learning and thought chaining, and key frames are sampled at time intervals to represent the approximate complete semantic information of the video, which can be formally expressed as follows:

[0065] (2).

[0066] is a collection of key frames, Indicates the first j frame, s is the sampling interval, t is the index of the key frame, ranging from 0 to , N Indicates the total number of frames in the video.

[0067] S5: Based on S3, LVLM is used to output three reasoning evidence sets from the perspectives of video style, video content, and information matching, which are respectively denoted as , and LVLM generates a confidence level for the reasoning.

[0068] S6: Interactively enhance the reasoning basis and video features. The process is defined as follows:

[0069] (3);

[0070] (4);

[0071] in represents the co-attention mechanism, Respectively represent i The initial features of the reasoning basis and the features after interactive enhancement, Indicates that you have learned i The video features obtained after the inference basis are Represents the enhanced video features. After all the interactions, the enhanced video features and the representation of each reasoning basis are finally obtained. .

[0072] S7: Evaluate the importance of different inference bases according to the confidence level. For the inference base R, the corresponding confidence level is recorded as , its importance The quantization equation is as follows:

[0073] (5).

[0074] in σ Control quantitative trends, b Controls the quantization range, which are learnable parameters, θ is a hyperparameter that represents the average level of inference.

[0075] S8: Based on S6 and S7, the reasoning basis and video features are aggregated to obtain the features used for final classification. The multi-dimensional reasoning basis after quantifying the importance is expressed as:

[0076] (6).

[0077] Multimodal features for video enhancement and Aggregation to obtain the final aggregate features for classification :

[0078] (7).

[0079] S9: Based on S8, aggregate features Input into the multi-layer perceptron to get the probability of predicting that the news video is false :

[0080] (8).

[0081] The cross entropy loss function is used to measure the difference between the probability distribution of the predicted label and the actual label. The formula of the loss function is:

[0082] (9).

[0083] in, y is the actual label of the news video (1 for fake news video, 0 for real news video), is the predicted probability that a news short video is false. The identification of short video false news is based on the following three criteria: (1) Short videos with a production style that shows a deliberate tendency to falsify are false news; (2) Short video news information with factual conflicts are false news; (3) Short video content information in different modes is inconsistent is false news.

[0084] S10: Output the false news short video classification results.

[0085] Figure 2It is a framework diagram for fake news short video detection, which is divided into the following three stages: (a) Video feature representation: Obtain the vectorized representation of the news short video, where the text-audio interactor, text-image interactor and text-video interactor are modules that use the common attention mechanism to interactively enhance text features and audio features, static visual frame features and dynamic visual video clip features, respectively. The video multimodal feature fusion module is the execution module of the above S3 step; (b) Auxiliary information generation: The auxiliary information of the news short video is generated by LVLM through prompt engineering, including reasoning basis and confidence level, where the multimodal prompt word generator is the execution module of the above S4 step. Block, the reasoning basis includes three perspectives: video style, video content and information matching; (c) LVLM-assisted reasoning: LVLM-assisted small model predicts the authenticity of news short videos, where the video-style basis interactor, video-content basis interactor and video-matching basis interactor are modules for interactive enhancement of the unified feature representation of the video with the video style reasoning basis, video content reasoning basis and information matching reasoning basis, i.e., the execution module of the above step S6, the key information selector is the execution module of the above step S7, the multimodal feature aggregator is the execution module of the above step S8, and the classifier is the execution module of the above step S9. The matrix represents the existence form of the features in the process of model training and prediction, the loss is the loss function that calculates the difference between the model prediction result and the true label, freezing means not updating the parameters of the module during the model training process, and fine-tuning means updating the parameters of the module during the model training process.

[0086] Table 1 compares the experimental results of the present invention with other methods on a real-world fake news short video dataset. Three experimental groups are set up, namely, the method using only a small model, the method using only a large model, and the method combining a large model and a small model. The experiment uses five-fold cross validation and uses four indicators: accuracy (Acc), precision (Prec), recall (recall), and harmonic mean (F1). The results show that the indicators of the present invention on the FakeSV dataset are better than those of existing methods.

[0087] Table 1 Performance comparison between the proposed method and the baseline model (%)

[0088]

[0089] Table 2 is the ablation experiment results of the present invention. Module VII is a video-information interactor, including a video-style basis interactor, a video-content basis interactor, and a video-matching basis interactor. Module KIS is a key information selector. The reasoning basis includes video style basis, video content basis, and information matching basis.

[0090] The results show that both the VII module and the KIS module can improve the performance of the present invention, among which the contribution of VII is more significant, because the VII module helps the small model understand the LVLM knowledge from the perspective of semantic content, providing greater improvement potential. The KIS module also brings enhancements, which helps the small model evaluate the quality of LVLM knowledge and selectively learn, thereby effectively alleviating the negative impact of the LVLM hallucination generation problem. The reasoning basis from three angles can also improve the performance of the present invention, among which the reasoning basis based on video content contributes more significantly to the improvement of model performance. The results show that the VII module and the KIS module play an indispensable role in establishing a bridge between the small model and the LVLM. Compared with the above two, the reasoning basis directly injects the extensive knowledge of the LVLM into the small model, which has the greatest impact on the performance of the model.

[0091] Table 2 Ablation study of each component of the present invention

[0092]

[0093] Figure 3 is the number of samples and prediction accuracy distributed in different confidence level intervals. The results show that the prediction accuracy of LVLM for news short videos is positively correlated with the confidence level, which also illustrates the importance of the generated auxiliary information. For a binary classification problem, the confidence level given by LVLM is generally in the range of [0.5,1.0], and about 47% of the confidence levels are distributed in the interval [0.70,0.80). The bar graph part represents the number of samples, in units of number, with the left vertical axis as the reference, and the line graph part represents the prediction accuracy, in units of percentage, with the right vertical axis as the reference, and the horizontal axis is the confidence level interval.

[0094] Figure 4 is the impact of changes in the hyperparameter θ on the performance of the present invention. θ represents the general level of the reasoning basis. When θ is low, the small model will overestimate the overall quality of the reasoning basis, which will increase the degree of learning of the reasoning basis, and thus increase the negative impact of the LVLM hallucination generation problem. On the contrary, a higher θ will reduce the degree of learning of the reasoning basis, thereby reducing the efficiency of utilizing the valuable information of the LVLM. Therefore, the purpose of setting θ is to find a balance point that can solve these two problems at the same time. The results show that this balance converges to 0.8. The horizontal axis represents the different value changes of the hyperparameter θ, and the vertical axis represents the model performance evaluation value, in percentage.

[0095] The method for detecting fake news short videos with large model-assisted reasoning of the present invention uses a deep learning model for feature extraction, and realizes the effective fusion of multimodal features through self-attention mechanism and cross-modal interaction technology, so as to fully understand the news content. Multimodal prompt words are constructed through prompt word engineering (role-playing, context learning and thought chain) and short video key frames, and a large visual language model (LVLM) is used to assist in reasoning of news short videos from the perspective of video style, video content and information matching, and the reasoning basis and confidence level are obtained. In addition, in order to reduce the negative impact of the hallucination generation problem and strengthen the perception of auxiliary information by the news detector so that it can better understand the reasoning basis, the present invention designs a key information selector and a video-information interactor, so as to establish an effective connection between a large language model and a fine-tuned small model. Finally, after a large number of experimental verifications, the detection effect of the present invention on the real-world fake news short video dataset exceeds the existing methods.

[0096] Based on the same inventive concept, the present invention also proposes a short video release information detection system, comprising:

[0097] The acquisition module is used to collect information data released by short videos and extract multimodal features of the information data released by short videos; multimodal features include text features, audio features and visual features; the text features, audio features and visual features of the short video information data are integrated to obtain a unified representation of the short video features.

[0098] The generation module is used to extract key frames of short videos, combine the key frames with the prompt word engineering including role-playing, context learning and thought chain, and construct multimodal prompt words; use multimodal prompt words to drive the large visual language model LVLM to generate reasoning basis and confidence level for judging the authenticity of short videos.

[0099] The enhancement module is used to extract features of the reasoning basis and interactively enhance the extracted features with the short video features after unified representation; the importance of different reasoning bases is evaluated according to the confidence level;

[0100] The detection module is used to aggregate the enhanced short video features and reasoning basis features based on importance; the information released by the short video is detected by inputting the aggregated features into a multi-layer perceptron.

[0101] The present invention also proposes a short video release information detection computer device, comprising: a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the short video release information detection method when executing the computer program.

[0102] The present invention also proposes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the short video release information detection method.

[0103] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A short video release information detection method, characterized in that: The following steps are involved: Collecting information data published by short videos and extracting multimodal features of the information data published by short videos; the multimodal features include text features, audio features and visual features; The text features, audio features and visual features of the information data released by the short video are integrated to obtain a unified representation of the short video features; Extract key frames from short videos, combine key frames with prompt word engineering including role-playing, context learning and thought chain, and construct multimodal prompt words; The multimodal prompt words are used to drive the large visual language model LVLM to generate the reasoning basis and confidence level for judging the authenticity of the short video; the key frames of the short video are specifically obtained by sampling the key frames at time intervals, and the sampling process is expressed as: in, is a set of key frames; F j represents the jth frame in a short video; s is the sampling interval; t is the index of the key frame, ranging from 0 to N represents the total number of frames in the short video; Extract features from the inference basis, and interactively enhance the extracted features with the short video features after unified representation; evaluate the importance of different inference bases according to the confidence level; the interactive enhancement of the inference basis and the short video features is expressed as follows: Where Co-Attn(·) represents the co-attention mechanism, Respectively represent the initial features of the i-th reasoning basis and the features after interactive enhancement, represents the video features obtained after learning the i-th reasoning basis, represents enhanced video features; Based on the importance, the enhanced short video features and the inference basis features are aggregated; the information released by the short video is detected by inputting the aggregated features into a multi-layer perceptron; based on the importance, the enhanced short video features and the inference basis are aggregated to obtain features for final classification; specifically, the following steps are included: Importance r It is expressed as: Where σ represents the control quantization trend, b represents the control quantization range, θ represents the hyperparameter based on the average level of inference, and α r Indicates the confidence level corresponding to the inference basis R; After quantifying the importance of the reasoning basis, it is expressed as in, They are three reasoning evidence sets outputted from the perspectives of video style, video content, and information matching using LVLM; Multimodal features for video enhancement and Aggregate to obtain the final aggregate feature E for classification cls : The information released by the short video is detected by inputting the aggregated features into a multi-layer perceptron, which specifically includes the following steps: The aggregate feature E cls Input into the multi-layer perceptron MLP to predict the probability that the short video release information is false The cross entropy loss function is used to measure the difference between the probability distribution of the predicted label and the actual label. The loss function is expressed as: Among them, y is the actual label of the short video release information; is the predicted probability that the information released in the short video is false, 1 represents a false news video and 0 represents a true news video.

2. A short video release information detection method according to claim 1, characterized in that: Using BERT to extract text features t , extract audio features using the pre-trained VGGish model Use deep neural network VGG19 to extract static visual frame features from visual features Extracting dynamic visual video clip features from visual features using C3D models 3. A short video release information detection method according to claim 2, characterized in that: The text features, audio features and visual features of the information data released by the short video are integrated to obtain a unified representation of the short video features as E v : in, They are the text features, audio features, static visual frame features, and dynamic visual video clip features of the short video release information data enhanced by the co-attention mechanism. Trm represents the Transformer encoder layer, and AP represents average pooling.

4. A short video release information detection system, characterized in that: include: A collection module, used to collect information data released by short videos and extract multimodal features of the information data released by short videos; the multimodal features include text features, audio features and visual features; The text features, audio features and visual features of the information data released by the short video are integrated to obtain a unified representation of the short video features; The generation module is used to extract key frames of short videos and combine the key frames with the prompt word engineering including role-playing, context learning and thought chain to construct multimodal prompt words; The multimodal prompt words are used to drive the large visual language model LVLM to generate the reasoning basis and confidence level for judging the authenticity of the short video; the key frames of the short video are specifically obtained by sampling the key frames at time intervals, and the sampling process is expressed as: in, is a set of key frames; F j represents the jth frame in a short video; s is the sampling interval; t is the index of the key frame, ranging from 0 to N represents the total number of frames in the short video; The enhancement module is used to extract features from the reasoning basis, and interactively enhance the extracted features with the short video features after unified representation; the importance of different reasoning bases is evaluated according to the confidence level; the interactive enhancement of the reasoning basis and the short video features is expressed as follows: Where Co-Attn(·) represents the co-attention mechanism, Respectively represent the initial features of the i-th reasoning basis and the features after interactive enhancement, represents the video features obtained after learning the i-th reasoning basis, represents enhanced video features; The detection module is used to aggregate the enhanced short video features and the reasoning basis features based on the importance; the information released by the short video is detected by inputting the aggregated features into a multi-layer perceptron; the features for the final classification are obtained by aggregating the enhanced short video features and the reasoning basis based on the importance; specifically, the following steps are included: Importance r It is expressed as: Where σ represents the control quantization trend, b represents the control quantization range, θ represents the hyperparameter based on the average level of inference, and α r Indicates the confidence level corresponding to the inference basis R; After quantifying the importance of the inference basis, it is expressed as in, They are three reasoning evidence sets outputted from the perspectives of video style, video content, and information matching using LVLM; Multimodal features for video enhancement and Aggregate to obtain the final aggregate feature E for classification cls : The information released by the short video is detected by inputting the aggregated features into a multi-layer perceptron, which specifically includes the following steps: The aggregate feature E cls Input into the multi-layer perceptron MLP to predict the probability that the short video release information is false The cross entropy loss function is used to measure the difference between the probability distribution of the predicted label and the actual label. The loss function is expressed as: Among them, y is the actual label of the short video release information; is the predicted probability that the information released in the short video is false, 1 represents a false news video and 0 represents a true news video.

5. A computer device for detecting short video release information, characterized in that: include: A memory, a processor, and a computer program stored in the memory, wherein the processor implements the steps of the short video release information detection method described in any one of claims 1 to 3 when executing the computer program.

6. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the short video release information detection method described in any one of claims 1 to 3.

Images