A video tag classification method, system and computer readable storage medium
Through the combination of multimodal fusion and two-stage model and the correction of map information, the problems of less data labeling and poor fine-grained recognition in video tag classification are solved, achieving higher recognition accuracy and fine-grainedness.
Patent Information
- Application Number
- CN202210465333.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-29
- Publication Date
- 2025-04-11
- Estimated Expiration
- 2042-04-29
AI Technical Summary
The existing video tag classification model has problems such as few data labels and insufficient physical tags, resulting in low recognition accuracy.
Multimodal fusion is performed by extracting the visual features and text features of the video data, multi-task prediction is performed using a two-stage model, and the prediction results are corrected through graph information to achieve finer granular identification.
It improves the accuracy and fine-grained recognition ability of video tag classification, and can still achieve good classification performance under a small number of data samples.
Smart Images
Figure CN114758283B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence and deep learning, and in particular to a video tag classification method, system and computer-readable storage medium. Background Art
[0002] With the development of communication and network speed, video has become another popular medium for information. The surge in the number of videos has made it an effective way to manage video content and label it hierarchically. Based on the diversity of video perception methods, a complete video label hierarchical classification algorithm should include two parts: video modality extraction and multimodal content understanding. For the understanding of multimodal content, we generally start with visual and textual modalities, and train a joint model to jointly infer video content. Based on practical experience, we have found that the current video label classification model in film and television scenarios has the following two problems:
[0003] 1) Few annotated data. Videos uploaded by Internet users vary greatly in content and quality. User-generated titles are often incomplete or ambiguous and may contain errors. Therefore, although there are many videos and movies, there is little high-quality annotated data.
[0004] 2) The entity labels are not fine-grained enough and the recognition accuracy is not high. Existing research is usually applicable to coarse-grained classification, such as considering the granularity of type and subject matter. Although some current baseline models of video classification have achieved good results in overall label recognition in film and television scenes, they are not good for some visually similar dramas. It is necessary to add features for further recognition. Factors such as characters and scenes also need to be considered.
[0005] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention
[0006] The purpose of the present invention is to provide a video tag classification method, system and computer-readable storage medium, which can provide a deeper understanding of the video content through graph information, and then achieve more fine-grained recognition through training and inference of a two-stage model.
[0007] To achieve the above object, the present invention provides a video tag classification method, the method comprising:
[0008] Extract visual and text features from video data;
[0009] Performing multimodal fusion of the visual features and the text features to obtain fused features;
[0010] Performing multi-task prediction on the fused features to obtain a prediction result, wherein the prediction result is a classified video label;
[0011] Acquire semi-structured data of the video data, and generate graph information according to the semi-structured data; the graph information represents association information between various types of nodes in the video;
[0012] The prediction result is corrected based on the atlas information to obtain a corrected result.
[0013] Preferably, extracting visual features and text features of video data includes:
[0014] Performing frame extraction processing on the video data to obtain a video frame image, and extracting the visual feature according to the video frame image;
[0015] The text information in the video data is preprocessed to form a corpus set, and the text features are extracted based on the corpus set.
[0016] Preferably, the preprocessing includes: word segmentation processing and stop word processing.
[0017] Preferably, extracting the visual features according to the video frame image comprises: obtaining a picture code {r1, r2, ..., r n} and divide the picture encoding into several image branches; obtain visual features Emb based on the image branches i ;
[0018] Extracting the text features based on the corpus set includes: obtaining a text sequence {w1, w2, ..., w n} and dividing the text sequence into multiple text branches; obtaining text features Emb based on the text branches t .
[0019] Preferably, performing multimodal fusion of the visual features and text features to obtain fused features comprises:
[0020] Learn interactive features of visual features and text features;
[0021] A fusion feature is obtained based on the interaction feature.
[0022] Preferably, performing multi-task prediction on the fusion feature to obtain a prediction result comprises:
[0023] The fused features are respectively input into the classification network and the similarity network after passing through several fully connected layers, so as to train the classification network and the similarity network simultaneously;
[0024] The training results outputted by the classification network and the similarity network are weightedly combined to obtain a prediction result.
[0025] Preferably, the method further comprises: the classification network trains the fusion features through a classification loss function, and the similarity network trains the fusion features through a metric learning loss function.
[0026] Preferably, correcting the prediction result based on the atlas information to obtain a corrected result includes:
[0027] Constructing an adjacency matrix, wherein the adjacency matrix includes neighbor nodes and non-neighbor nodes, wherein the attention of the neighbor nodes is set to 1, and the attention of the non-neighbor nodes is set to 0; the neighbor nodes are nodes having a connection relationship between the graph information and the prediction results;
[0028] The probability distribution of the adjacency matrix is calculated, and a correction result is obtained according to the node with the highest probability.
[0029] The present invention provides a video tag classification system, the system comprising:
[0030] A feature extraction unit, used to extract visual features and text features of video data;
[0031] A fusion unit, used for multimodally fusing the visual features and text features to obtain fused features;
[0032] A label classification unit, used for performing multi-task prediction on the fusion feature to obtain a prediction result, wherein the prediction result is a classified video label;
[0033] A data acquisition unit, used to acquire semi-structured data of the video data, and generate graph information according to the semi-structured data; the graph information represents the association information between various nodes in the video;
[0034] A correction unit is used to correct the prediction result based on the atlas information to obtain a corrected result.
[0035] The present invention provides a computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, the steps of a video tag classification method as described in any one of the above items are implemented.
[0036] Compared with the problems of insufficient data annotation and low fine-grained recognition accuracy in the prior art, the video label classification method according to the present invention completes more fine-grained recognition through the training and inference of label prediction models and entity error correction models and achieves better classification performance without the need for large-scale data sets. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 is a flow chart of a video tag classification method according to an embodiment of the present invention;
[0038] Figure 2 is a structural diagram of a multi-task network layer according to an embodiment of the present invention;
[0039] Figure 3 It is a structural diagram of a video tag classification system according to one embodiment of the present invention. DETAILED DESCRIPTION
[0040] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation modes.
[0041] Unless explicitly stated otherwise, throughout the specification and claims, the term “comprise” or variations such as “include” or “comprising”, etc., will be understood to include the stated elements or components but not to exclude other elements or components.
[0042] like Figure 1 As shown, according to a preferred embodiment of the present invention, a video tag classification method comprises:
[0043] Step 101, extracting visual features and text features of video data; wherein the video data is training data in an embodiment of the present invention, the visual features represent image information in the video data, the text features represent text information in the video data, and the visual features and text features belong to different modalities.
[0044] Step 102, multi-modally fusing the visual features and text features to obtain fused features; fusing the features of the two modalities of the visual features and text features extracted in step 101 to generate fused features.
[0045] Step 103, performing multi-task prediction on the fused features to obtain a prediction result, wherein the prediction result is a classified video label; wherein the fused features generated in step 102 are fed to a downstream classification task layer for multi-task prediction, and the loss value is minimized by tuning the model parameters to obtain the final prediction result.
[0046] Step 104, obtaining semi-structured data of the video data, and generating graph information according to the semi-structured data; the graph information represents the association information between various nodes in the video; obtaining semi-structured data of the video data, such as the name of a film or TV series, actors, etc., from various knowledge sources on the Internet, and constructing graph information related to film and TV knowledge, the graph information including various semi-structured data association information, such as "actor-series-actor" and "series-actor-series" and other film and TV knowledge graph information of the acting relationship. Among them, the way to obtain semi-structured data can be obtained from knowledge sources such as Douban and Encyclopedia. Semi-structured data includes field information such as the name of a film or TV series, release date, director, actor, theme type, brief description, etc.
[0047] Step 105, the prediction result is corrected based on the graph information to obtain a corrected result. In this step, the entity error correction model expanded by the local attention head introduces external graph information to correct the prediction result obtained by the front model to obtain a more accurate corrected result.
[0048] The method described in the embodiment of the present invention can be widely applied to multimodal video classification prediction problems based on image-text pair data, especially classification problems with a small number of data samples, which helps to improve the operating efficiency of video websites.
[0049] In a video tag classification method described in an embodiment of the present invention, preferably, extracting visual features and text features of video data includes:
[0050] The video data is subjected to frame extraction processing to obtain a video frame image, and the visual feature is extracted according to the video frame image; wherein the visual feature is also the image frame feature of the video frame image.
[0051] The text information in the video data is preprocessed to form a corpus set, and the text features are extracted based on the corpus set. In a specific embodiment, a multimodal pre-trained model trained based on large-scale general data is used to extract visual features and text features.
[0052] In the video tag classification method described in the embodiment of the present invention, preferably, the preprocessing includes: word segmentation processing and stop word processing.
[0053] In a video tag classification method according to an embodiment of the present invention, preferably, extracting the visual features according to the video frame image in a feature extraction network comprises: obtaining the picture encoding {r1, r2, ..., r n} and divide the picture encoding into several image branches; obtain visual features Emb based on the image branches i ;
[0054] Extracting the text features based on the corpus set includes: obtaining a text sequence {w1, w2, ..., w n} and dividing the text sequence into multiple text branches; obtaining text features Emb based on the text branches t .
[0055] In a specific embodiment, the text sequence {w1, w2, ..., w n} and the picture encoding of the video frame image {r1, r2, ..., r n} will be divided into multiple text branches and image branches. Each text branch and image branch after segmentation is recorded as a token and encoded. Based on the text branch and the image branch, the representations trained in the same semantic space are extracted, that is, the text feature Emb t and image features Emb i In the text branch, we first transform the text sequence {w1, w2, ..., w n The word vector and position vector of each token in} are added together as the input vector. Since the text branches are ordered, the position vector refers to the encoding of the word order information of the token token into the form of a feature vector, so that the position relationship information of the words can be introduced into the model, thereby obtaining the inherent order information of the text. The input vector is then fed to the feature extraction Encoder module of the Transformer model for encoding input. Each Encoder contains two sublayers. The first sublayer is a multi-head self-attention mechanism, including 12 attention heads; the second sublayer is a forward propagation layer, which is used to improve the nonlinear fitting ability of the model. Each sublayer uses a residual network, which can alleviate the problem of gradient disappearance during training, so that a deeper network can be built. The entire Transformer model stacks 12 identical feature extraction Encoder module structures, and finally outputs the text feature Emb t In the image branch, the input is a single image. First, the size of the input image is adjusted. The given image is scaled to 224*224, and then the image is cropped proportionally, retaining the middle image part, and then the pixel values of each channel of the image are normalized. Here, the ResNet-50 network composed of several residual modules is used as the backbone structure for visual feature extraction. The convolution and pooling operations are mainly performed in space, and finally the visual feature Emb is obtained. i .
[0056] In a video tag classification method described in an embodiment of the present invention, preferably, multi-modally fusing the visual features and text features to obtain fused features includes:
[0057] Learn interactive features of visual features and text features;
[0058] A fusion feature is obtained based on the interaction feature.
[0059] In a specific embodiment, the text feature Emb obtained by the previous feature extraction network t and visual features Emb i They are regarded as text tokens and visual tokens respectively. After concatenating the two tokens, the interactive feature Emb is obtained, and then fed into the feature fusion Encoder module of the Transformer model to fuse the two tokens. In essence, the fusion of features based on the feature fusion Encoder module can be interpreted as performing dot product calculations on the query vector Query, the key vector Key, and the value vector Value in multiple output heads of the attention, and obtaining adaptive weighting schemes respectively. For information in different areas, different weighted values are combined to finally obtain the overall vector expression. Here, the main function of the feature fusion Encoder is to learn the interactive features of the text token and the visual token in the same sample, and finally obtain a fixed-length vector representation Emb. attention . Preferably, the feature fusion encoder stacks 6 layers of the same encoding layer, and each layer uses an 8-head attention module, which is equivalent to fusion learning of image and text modality features in 8 different channels.
[0060] In a video label classification method described in an embodiment of the present invention, preferably, performing multi-task prediction on the fusion feature to obtain a prediction result includes:
[0061] The fused features are respectively input into the classification network and the similarity network after passing through several fully connected layers, so as to train the classification network and the similarity network simultaneously;
[0062] The training results outputted by the classification network and the similarity network are weighted and combined to obtain the prediction result. Specifically, in the embodiment of the present invention, the classification network and the similarity network are combined in the same large network for multi-task training and joint optimization. Figure 2 The figure shows the structure diagram of the multi-task network layer. After modal fusion, the fused features first pass through several fully connected layers, and then are input into the classification network and similarity network for simultaneous training.
[0063] A video tag classification method described in an embodiment of the present invention, preferably, the method also includes: the classification network trains the fusion feature through a classification loss function, and the similarity network trains the fusion feature through a metric learning loss function.
[0064] In a specific embodiment, the classification loss function may be a cross entropy loss function, which is a classic classification loss function. In this embodiment, the fusion feature is used as a training sample. The classification network trains the fusion feature through the classification loss function as follows: first, the softmax formula and the cross entropy loss function are combined. Assume that there are N training samples. Belongs to C categories, where each sample r i Labeled as category label l i As shown in Formula 1, the calculation formula of the cross entropy loss function is as follows:
[0065]
[0066] where f s (r i , c) indicates that the final fully connected layer has an input of r i Next, the output of the cth class, l i is the category label of the corresponding input.
[0067] Preferably, in the embodiment of the present invention, in addition to using the classification loss function for constraint, the metric learning loss function is also used to merge with the classification loss function for feature representation learning. For example, in the calculation, there are three samples represented as (r i , p i , n i ), where r i is a reference sample from a specific class, p i is from the same sample, n i are samples from different classes. Given a reference sample r i , the metric learning network can generate a feature vector f t (r i )∈R D , where the hyperparameter D is the feature dimension. Ideally, for each reference sample r i , it is expected that it is different from any different class sample n i The distance is greater than that of similar samples p i The margin length is m, and m>0, that is, D(r i , p i )+m<D(r i , n i ). The metric learning loss function is shown in Formula 2:
[0068]
[0069] Where D is the square Euclidean distance between two fused features in the metric learning network after L2 norm normalization, and m is the minimum margin, that is, the distance between the reference sample and samples of different classes minus the distance between the reference sample and samples of the same class. The minimum value of m needs to be found during the model training process.
[0070] Finally, these two types of losses are integrated through weighted combination, as shown in Formula 3, where λ is the weight coefficient of the two calculations, ranging from 0 to 1:
[0071] Loss = λ*Loss s (r, l)+(1-λ)Loss t (r, p, n, m) Formula 3
[0072] Introducing similarity tasks into the multi-task learning network increases the difficulty of training the classification model and obtains multi-category labels for the video. At the same time, the learned features can better express the sample differences.
[0073] In a video tag classification method according to an embodiment of the present invention, preferably, correcting the prediction result based on the graph information to obtain a corrected result includes:
[0074] Constructing an adjacency matrix, wherein the adjacency matrix includes neighbor nodes and non-neighbor nodes, wherein the attention of the neighbor nodes is set to 1, and the attention of the non-neighbor nodes is set to 0; the neighbor nodes are nodes having a connection relationship between the graph information and the prediction results;
[0075] The probability distribution of the adjacency matrix is calculated, and a correction result is obtained according to the node with the highest probability.
[0076] In a specific embodiment, before correcting the prediction result, the graph information is first obtained. Semi-structured data related to film and television are obtained, and the semi-structured data at least includes: series, release date, director, actor, theme type, brief description and other field information; series and actors are extracted as two types of nodes. The name of the film and television drama constitutes the film and television name node in the knowledge graph, and the performers constitute the actor node in the knowledge graph. The performers in the play include the protagonist, supporting role, and the masses. Here, the actors refer to the main performers; a graph information of film and television knowledge containing two acting relationships of "actor-series-actor" and "series-actor-series" is constructed. The relationship is an undirected edge connecting the film and television name node and the actor node, that is, forming a one-to-many relationship between the film and television drama name and the actor. The co-occurrence information of the external knowledge graph information is introduced to correct the prediction results of the front model to obtain more accurate entity label prediction results.
[0077] Then, a batch of training samples G = {g1, g2, ..., g n}, where g∈{n1,n2,n3,...,N}. For each input g, a portion of the nodes are randomly masked to obtain g∈{n1,[mask],n3,...,N}, where node n i Including the movie name node and the actor node, as well as a special MASK node. The goal of this step is to train a model to reconstruct the missing nodes to calculate the probability distribution and predict the node n with the highest probability of appearing. i ∈g.
[0078] In a preferred embodiment, the attention head of the Transformer model is expanded. In the Encoder part of the traditional Transformer, the calculation of the Query vector, Key vector and Value vector of the self-attention mechanism is for all tokens, so it is equivalent to calculating a global attention score; however, in this step, we want to pay more attention to the role of adjacent nodes, that is, to calculate the attention score between the two nodes of actor and film and television drama name, we only need to have a starring relationship and a cooperative relationship, and do not need to calculate all tokens. Therefore, part of the attention head in the Encoder is replaced by local attention, and the graph information is added to correct the prediction result, so as to force the attention between non-neighboring nodes to be set to zero. For a video to be predicted, the candidate film and television names with predicted probabilities are regarded as film and television name nodes, the actors identified by the characters appearing on the screen are regarded as actor nodes, and each node is regarded as a token, which is converted into a vector and then input into the model. During training, the two types of nodes, film and television name-actor, are extracted from the film and television knowledge graph information, and an adjacency matrix of a bipartite graph is constructed and input into the model. In the adjacency matrix, there is an interactive relationship between TV series and actors. The names of TV series and actors in the relationship are regarded as neighbor nodes, and the attention of neighbor nodes is set to 1. The other nodes are regarded as non-neighbor nodes and set to 0. Finally, the cross entropy loss function is used to train the network to calculate the probability distribution of all categories to predict the output node category. The node with the highest probability is the corrected label, and the correction result is finally obtained.
[0079] The embodiment of the present invention provides a video tag classification system, such as Figure 3 As shown, the system comprises:
[0080] A feature extraction unit 301, used to extract visual features and text features of video data;
[0081] A fusion unit 302, configured to perform multimodal fusion of the visual features and the text features to obtain fused features;
[0082] A label classification unit 303 is used to perform multi-task prediction on the fusion feature to obtain a prediction result, where the prediction result is a classified video label;
[0083] The data acquisition unit 304 is used to acquire the semi-structured data of the video data and generate graph information according to the semi-structured data; the graph information represents the association information between various nodes in the video;
[0084] The correction unit 305 is used to correct the prediction result based on the graph information to obtain a corrected result.
[0085] An embodiment of the present invention provides a computer-readable storage medium, characterized in that a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the steps of a video tag classification method as described in any of the above specific embodiments are implemented.
[0086] The film and television video label classification model that integrates graph information proposed in the specific embodiment of the present invention is a new attempt to use image-text pair data in the multimodal classification problem. It does not require a large-scale data set and achieves good classification performance.
[0087] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0088] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0089] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1A function specified in one or more boxes.
[0090] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0091] The foregoing description of specific exemplary embodiments of the present invention is for the purpose of illustration and demonstration. These descriptions are not intended to limit the present invention to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical application, so that those skilled in the art can realize and utilize various different exemplary embodiments of the present invention and various different selections and changes. The scope of the present invention is intended to be limited by the claims and their equivalents.
Claims
1. A video tag classification method, characterized in that: The method comprises: Extract visual and text features from video data; Performing multimodal fusion of the visual features and the text features to obtain fused features; Performing multi-task prediction on the fused features to obtain a prediction result, wherein the prediction result is a classified video label; Acquire semi-structured data of the video data, and generate graph information according to the semi-structured data; the graph information represents association information between various types of nodes in the video; The prediction result is corrected based on the graph information to obtain a corrected result, including: constructing an adjacency matrix, the adjacency matrix including neighbor nodes and non-neighbor nodes, wherein the attention of the neighbor nodes is set to 1, and the attention of the non-neighbor nodes is set to 0; the neighbor nodes are nodes that have a connection relationship between the graph information and the prediction result; calculating the probability distribution of the adjacency matrix, and obtaining the corrected result according to the node with the highest probability.
2. A video tag classification method as claimed in claim 1, characterized in that: Extracting visual features and text features from video data includes: Performing frame extraction processing on the video data to obtain a video frame image, and extracting the visual feature according to the video frame image; The text information in the video data is preprocessed to form a corpus set, and the text features are extracted based on the corpus set.
3. A video tag classification method as claimed in claim 2, characterized in that: The preprocessing includes: word segmentation processing and stop word processing.
4. A video tag classification method as claimed in claim 2, characterized in that: Extracting the visual features according to the video frame image includes: obtaining the picture code {r1, r2, ..., r n } and divide the picture encoding into several image branches; obtain visual features Emb based on the image branches i ; Extracting the text features based on the corpus set includes: obtaining a text sequence {w1, w2, ..., w n } and dividing the text sequence into multiple text branches; obtaining text features Emb based on the text branches t .
5. A video tag classification method as claimed in claim 1, characterized in that: The multimodal fusion of the visual features and the text features to obtain the fusion features includes: Learn interactive features of visual features and text features; A fusion feature is obtained based on the interaction feature.
6. A video tag classification method as claimed in claim 1, characterized in that: Performing multi-task prediction on the fusion feature to obtain a prediction result includes: The fused features are respectively input into the classification network and the similarity network after passing through several fully connected layers, so as to train the classification network and the similarity network simultaneously; The training results outputted by the classification network and the similarity network are weightedly combined to obtain a prediction result.
7. A video tag classification method as claimed in claim 6, characterized in that: The method also includes: the classification network trains the fusion features through a classification loss function, and the similarity network trains the fusion features through a metric learning loss function.
8. A video label classification system, characterized in that: The system comprises: A feature extraction unit, used to extract visual features and text features of video data; A fusion unit, used for multimodally fusing the visual features and text features to obtain fused features; A label classification unit, used for performing multi-task prediction on the fusion feature to obtain a prediction result, wherein the prediction result is a classified video label; A data acquisition unit, used to acquire semi-structured data of the video data, and generate graph information according to the semi-structured data; the graph information represents the association information between various nodes in the video; A correction unit, for correcting the prediction result based on the graph information to obtain a corrected result, includes: constructing an adjacency matrix, the adjacency matrix includes neighbor nodes and non-neighbor nodes, wherein the attention of the neighbor nodes is set to 1, and the attention of the non-neighbor nodes is set to 0; the neighbor nodes are nodes that have a connection relationship between the graph information and the prediction result; calculating the probability distribution of the adjacency matrix, and obtaining the corrected result according to the node with the highest probability.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of a video tag classification method according to any one of claims 1 to 7 are implemented.