Video scene recognition method, device, equipment and medium
By filtering keyframes and extracting multi-dimensional visual features, and combining them with a feature-mutually exclusive knowledge graph, the problem of low recognition accuracy caused by camera shake in video interview scenarios is solved, achieving high-accuracy video scene recognition.
Patent Information
- Application Number
- CN202611216289.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-11-03
Smart Images
Figure CN123033165A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and is applied to fintech and healthcare scenarios, particularly to a video scene recognition method, apparatus, device, and medium. Background Technology
[0002] In video interview scenarios for financial lending, online medical qualification applications, and other businesses, users are required to record on-site videos. The system then identifies the actual scene corresponding to the user's video to verify whether the scene type declared by the user matches the actual scene, thus preventing the risk of scenario fraud.
[0003] In related technologies, the video is typically first uniformly sampled at a fixed frame rate, and then each frame is classified into scenes. Finally, a majority vote mechanism is used to output the final video scene recognition result. However, because uniformly sampling at a fixed frame rate results in a large number of invalid frames generated by camera shake during user-captured footage being sampled with equal probability, the effective features are diluted, and the majority vote result is easily interfered with. Furthermore, drawing conclusions solely based on simple statistical analysis of single-frame classification results cannot accurately identify mixed scenes with contradictory features, resulting in low overall video scene recognition accuracy. Summary of the Invention
[0004] This invention provides an artificial intelligence video scene recognition method, apparatus, computer equipment, and medium, aiming to solve the technical problem that current technology cannot accurately identify mixed scenes with contradictory features and has low overall video scene recognition accuracy, thereby improving the accuracy of video scene recognition.
[0005] Firstly, a video scene recognition method is provided, including: Acquire the video to be identified, wherein the video to be identified includes a sequence of video frames to be identified; Based on the clarity of the video frames to be identified in the video frame sequence and the structural similarity between the video frames to be identified, a key video frame sequence is selected from the video frame sequence to be identified. Visual features are extracted from each key video frame in the key video frame sequence to obtain multi-dimensional visual features corresponding to each key video frame. The multi-dimensional visual features corresponding to each key video frame are classified to obtain multiple categories of visual features, and the frequency of occurrence of each category of visual features is counted to obtain visual feature statistics. Based on a pre-defined feature-mutually exclusive knowledge graph, the scene conflict relationships between various visual features are determined, and scene conflict relationship information is obtained. Based on the visual feature statistics and the scene conflict relationship information, the scene corresponding to the video to be identified is identified, and the identification result is obtained.
[0006] Secondly, a video scene recognition device is provided, comprising: The acquisition module is used to acquire the video to be identified, wherein the video to be identified includes a sequence of video frames to be identified; The filtering module is used to filter key video frame sequences from the video frame sequence to be identified based on the clarity of the video frames to be identified and the structural similarity between the video frames to be identified. The extraction module is used to extract visual features from each key video frame in the key video frame sequence to obtain multi-dimensional visual features corresponding to each key video frame. The classification module is used to classify the multi-dimensional visual features corresponding to each key video frame, obtain multiple categories of visual features, and count the frequency of occurrence of each category of visual features to obtain visual feature statistics. The determination module is used to determine the scene conflict relationships between various visual features based on a preset feature mutual exclusion knowledge graph, and obtain scene conflict relationship information; The recognition module is used to identify the scene corresponding to the video to be recognized based on the visual feature statistical information and the scene conflict relationship information, and obtain the recognition result.
[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the video scene recognition method described above.
[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the video scene recognition method described above.
[0009] The above-mentioned video scene recognition method, device, computer equipment, and storage medium can be used to acquire a video to be recognized, which includes a video frame sequence; based on the clarity of the video frames and the structural similarity between the video frames, a key frame sequence is selected from the video frame sequence; visual features are extracted from each key frame in the key frame sequence to obtain multi-dimensional visual features corresponding to each key frame; the multi-dimensional visual features corresponding to each key frame are classified to obtain multiple types of visual features, and the frequency of occurrence of each type of visual feature is counted to obtain visual feature statistics; based on a preset feature mutual exclusion knowledge graph, the scene conflict relationship between various types of visual features is determined to obtain scene conflict relationship information; based on the visual feature statistics and the scene conflict relationship information, the scene corresponding to the video to be recognized is identified to obtain the recognition result. In this invention, by selecting key frames based on the clarity of video frames and the structural similarity between frames, invalid image interference is effectively eliminated; multi-dimensional visual features of key frames are extracted and classified and statistically analyzed, and the scene is comprehensively judged by combining the scene conflict relationship between features, which can identify mixed scenes with contradictory features, and the overall video scene recognition accuracy is high. Attached Figure Description
[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0011] Figure 1 This is a schematic diagram of an application environment for a video scene recognition method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a video scene recognition method according to an embodiment of the present invention; Figure 3 yes Figure 2 A schematic diagram of a specific implementation method for step S20; Figure 4 yes Figure 3 A flowchart illustrating a specific implementation method prior to step S22; Figure 5 yes Figure 2 A schematic diagram of a specific implementation method for step S30; Figure 6 yes Figure 2 A schematic diagram of a specific implementation of step S40; Figure 7 yes Figure 2 A schematic diagram of a specific implementation method for step S50; Figure 8 yes Figure 2 A schematic diagram of a specific implementation method for step S60; Figure 9 This is a schematic diagram of a video scene recognition device according to an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 11 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation
[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0013] The video scene recognition method provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can capture a video to be identified via a client. The video to be identified includes a sequence of video frames. Based on the clarity and structural similarity between the video frames in the sequence, key video frame sequences are selected. Visual features are extracted from each key video frame to obtain multi-dimensional visual features. These multi-dimensional visual features are categorized to obtain multiple categories of visual features, and the frequency of occurrence of each category is statistically analyzed to obtain visual feature statistics. Based on a pre-defined feature mutual exclusion knowledge graph, scene conflict relationships between various visual features are determined to obtain scene conflict relationship information. Based on the visual feature statistics and scene conflict relationship information, the scene corresponding to the video to be identified is identified to obtain the identification result. In this invention, by selecting key frames based on video frame clarity and inter-frame structural similarity, invalid image interference is effectively eliminated. By extracting multi-dimensional visual features from key frames and classifying and statistically analyzing them, and combining the scene conflict relationships between features to comprehensively determine the scene, mixed scenes with contradictory features can be identified, resulting in high overall accuracy of video scene recognition.
[0014] Those skilled in the art should understand that the above-described video scene recognition method can also be executed entirely locally on the client side, without uploading the video to be recognized to the server, and video scene recognition can be completed independently on the client side.
[0015] The present invention will now be described in detail through specific embodiments.
[0016] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the video scene recognition method provided in this embodiment of the invention includes the following steps: S10: Obtain the video to be identified, which includes the sequence of video frames to be identified.
[0017] The following will use the verification of the business premises declared by users in a video interview scenario for financial credit and other businesses as an example to illustrate the implementation of the present invention in detail.
[0018] The video to be identified is continuous dynamic image data collected by the client's camera, which includes environmental information of the business premises declared by the user.
[0019] The video frame sequence to be identified is a series of consecutive static images that constitute the video to be identified. Each static image is called a video frame, and the video frames are arranged in chronological order to form a sequence.
[0020] In this step, when business premises verification is required, a prompt message can be displayed asking the user to film the interior and surrounding environment of their business premises. The user can then automatically access their camera via the client to record a video, which becomes the video to be identified. Subsequently, by identifying the scene corresponding to this video, the business premises presented in the video can be determined. This allows for further verification of whether the identified business premises match the user's declared business premises, thus enabling the verification of the user's declared business premises in video interview scenarios for financial lending and other businesses.
[0021] For example, the prompt message could be: "Please hold your phone and slowly rotate it 360 degrees to fully display the interior and surrounding environment of your business premises," allowing users to record a video for verification. Alternatively, it could prompt users to "hold their phone and slowly rotate it 360 degrees to fully display the treatment room, pharmacy, disinfection area, and the location where the medical institution's practice license is displayed," allowing them to record a video of the clinic's interior for qualification verification.
[0022] S20: Based on the clarity of the video frames to be identified in the video frame sequence and the structural similarity between the video frames to be identified, select key video frame sequences from the video frame sequence to be identified.
[0023] Structural similarity refers to the similarity relationship between different video frames, and is used to determine whether there is duplicate information between video frames.
[0024] In this step, each video frame in the sequence of video frames to be identified can be processed sequentially. The key video frame sequence is obtained by filtering based on the clarity of the video frames and the structural similarity between them.
[0025] In some embodiments, candidate video frame sequences can be first filtered based on structural similarity, and then key video frame sequences can be filtered based on sharpness. For example, Figure 3 As shown, step S20, which is to filter key video frame sequences from the video frame sequence to be identified based on the sharpness of the video frames to be identified and the structural similarity between the video frames to be identified, includes the following steps S21-S22: S21: Based on the structural similarity of adjacent video frames to be identified in the video frame sequence to be identified, candidate video frame sequences are selected from the video frame sequence to be identified.
[0026] Structural similarity is an objective indicator that measures the degree of structural similarity between two images. Its value ranges from 0 to 1. The closer the value is to 1, the more similar the content of the two images are. The closer the value is to 0, the greater the difference in content between the two images.
[0027] In this step, the video frames to be identified in the sequence are read sequentially. Using the first frame as the reference frame, the structural similarity index (SSIM) between the current frame and the previous reference frame is calculated. When the calculated structural similarity is less than a preset similarity threshold, the current frame is determined to have undergone effective viewpoint displacement compared to the reference frame, containing new scene information. It is then marked as a candidate frame and updated as the new reference frame. When the structural similarity is greater than or equal to the preset threshold, the current frame is determined to have highly overlapping content with the reference frame and is discarded. This process is repeated until all video frames have been processed, resulting in a candidate video frame sequence.
[0028] The preset similarity threshold is a quantification threshold used to determine whether there is a valid viewpoint shift between adjacent video frames and whether new scene information is contained. When the structural similarity between adjacent video frames is less than the preset similarity threshold, it is determined that a valid scene change has occurred, and the corresponding video frame is retained; when the structural similarity between adjacent video frames is greater than or equal to the preset similarity threshold, it is determined that the content is highly repetitive and is discarded. The preset similarity threshold can be set by those skilled in the art according to actual needs. For example, the value range of the preset similarity threshold can be 0.7-0.9.
[0029] For example, in the scenario of verifying the premises of a financial lending business, assuming a structural similarity threshold of 0.75, a user's phone is used to photograph the premises in a circular motion. When the structural similarity between adjacent frames is 0.92, it is determined that the user is slowly scanning a blank wall, resulting in highly repetitive content, and is discarded. When the structural similarity between adjacent frames is 0.68, it is determined that the user is rotating the phone to scan the company sign area, resulting in a valid change in the scene, and the frame is retained as a candidate video frame. In the scenario of verifying the premises for online medical qualification applications, the threshold is set to 0.75: adjacent frames with a similarity of 0.90 (scanning a blank wall) are discarded; those with a similarity of 0.65 (scanning a medical institution's practice license) are retained as candidate frames.
[0030] S22: Based on the clarity of candidate video frames in the candidate video frame sequence, select key video frame sequences from the candidate video frame sequence.
[0031] Among them, the clarity of a video frame is an indicator of how clear a video frame is. The higher the clarity, the clearer the video frame; the lower the clarity, the blurrier the video frame.
[0032] In this step, when the clarity of a candidate video frame is greater than or equal to a preset clarity threshold, the candidate video frame is determined to have high clarity and is identified as a key video frame; when the clarity of a candidate video frame is less than the preset threshold, the candidate video frame is determined to have low clarity and is discarded. This process is repeated until all candidate video frames have been processed, resulting in a key video frame sequence.
[0033] The preset sharpness threshold is a quantization threshold used to determine whether a video frame has motion blur or out-of-focus blur, and whether it meets the quality requirements for subsequent feature extraction. When the sharpness of a video frame is greater than or equal to the preset sharpness threshold, it is determined to be a sharp image and the frame is retained; when the sharpness of a video frame is less than the preset sharpness threshold, it is determined to be a blurry image and is discarded. The preset sharpness threshold can be set by those skilled in the art according to actual needs.
[0034] For example, in the scenario of verifying the premises of a financial credit business, blurry frames captured when a user quickly rotates their phone are deemed low-quality frames and discarded because their clarity is less than a preset clarity threshold; clear frames captured when a customer smoothly rotates their phone are deemed qualified frames and retained in the key video frame sequence because their clarity is greater than or equal to the preset clarity threshold. In the scenario of verifying the premises for online medical qualification applications, blurry license frames captured by rapidly shaking the phone are discarded; clear frames of the treatment table captured smoothly are retained; a 1-minute video (approximately 1800 frames) is condensed into 12-18 key frames (covering four core areas: the treatment area, pharmacy, disinfection area, and qualification plaque).
[0035] In one embodiment, through the above steps S21-S22, a 1-minute video to be identified (approximately 1800 frames) can be condensed into a high-quality key video frame sequence (approximately 10-15 frames) containing a full-view perspective.
[0036] The above steps S21-S22 employ a two-layer filtering method: first, candidate frames are selected based on inter-frame structural similarity, and then key frames are selected based on clarity. This method can quickly eliminate redundant images with highly repetitive content, reducing the volume of video data. It can also accurately filter out blurry, jittery, and invalid images, retaining effective frames with complete information and excellent image quality. This not only reduces the amount of data processing in subsequent feature extraction and inference stages, lowering the overall computation time and resource consumption, but also ensures the accuracy and reliability of subsequent visual feature extraction results, thereby improving the overall efficiency and accuracy of video scene recognition from the source.
[0037] In some embodiments, the variance of the Laplacian operator can be calculated as sharpness. For example, Figure 4 As shown, before step S22, which is to select the key video frame sequence from the candidate video frame sequence based on the sharpness of the candidate video frames in the candidate video frame sequence, the following steps S23-S24 are also included: S23: Perform grayscale processing on the candidate video frames in the candidate video frame sequence.
[0038] Grayscale conversion refers to the image processing procedure of converting a color image containing information from the red, green, and blue channels into a single-channel grayscale image containing only brightness information. Grayscale images retain the structure, contours, and details of the original image while reducing the data volume to one-third of that of the original color image.
[0039] Understandably, candidate video frames are color images containing red, green, and blue channel information. To simplify the image data structure, eliminate the interference of color information on edge detection, and improve the efficiency and accuracy of subsequent sharpness calculation, candidate video frames can be converted into grayscale images containing only brightness information.
[0040] For example, in the scenario of verifying the premises of financial credit operations, a color candidate video frame containing "company sign + office desks and chairs" can be processed into grayscale to obtain a grayscale candidate video frame containing "company sign + office desks and chairs".
[0041] S24: Calculate the variance of the Laplacian operator of the candidate video frames after grayscale processing to obtain the sharpness of the candidate video frames.
[0042] The Laplacian operator variance is the statistical variance of the Laplacian convolution result. The Laplacian operator is sensitive to edge and detail information in the image; the clearer the image, the richer the edges and details, and the larger the calculated variance value; the blurrier the image, the more blurred the edges and details, and the smaller the variance value.
[0043] For example, a 3×3 Laplacian kernel can be used to convolve the grayscale image to obtain the second derivative response map, which highlights the edges and details in the image. Then, the variance of all pixel values in the entire response map is calculated, and this variance is used as the sharpness of the candidate video frame. Understandably, the sharper the image, the richer the edges and details, the greater the difference in pixel values in the Laplacian response map, and the higher the calculated variance; conversely, the blurrier the image, the smoother the edges and details, the smaller the difference in pixel values in the Laplacian response map, and the lower the calculated variance.
[0044] For example, after sequentially converting candidate video frames in a candidate video frame sequence to grayscale, the variance of the Laplacian operator is calculated frame by frame. Specifically, for blurry frames captured when a user rapidly rotates their phone, the variance of the Laplacian operator is less than a preset sharpness threshold, indicating that edges and details are almost completely lost, and these frames can be discarded. For sharp frames captured when a user smoothly rotates their phone, the variance of the Laplacian operator is greater than or equal to the preset sharpness threshold, indicating that details such as sign text and object textures are clearly visible, and these frames can be retained.
[0045] The above steps S23-S24 simplify the image data by first performing grayscale processing on the candidate video frames, and then calculating the Laplacian operator variance quantization of sharpness. This can efficiently and accurately determine the degree of image sharpness, simplify the calculation process, improve the sharpness detection efficiency, and stably distinguish between motion blur and sharp images. This provides objective numerical basis for subsequent selection of high-quality key frames and ensures that the frame selection results are rigorous and reliable.
[0046] S30: Extract visual features from each key video frame in the key video frame sequence to obtain the multi-dimensional visual features corresponding to each key video frame.
[0047] Multi-dimensional visual features refer to the feature information that describes the content of video frames from multiple different semantic dimensions, with each dimension corresponding to a specific attribute of the scene. For example, dimensions describing the spatial layout of the scene, dimensions describing the objects present in the scene, and dimensions describing the text information in the scene.
[0048] In this step, visual features can be extracted from each key video frame in the key video frame sequence in sequence to extract multi-dimensional visual features that can reflect different aspects of the scene.
[0049] In some embodiments, a visual feature extraction model can be pre-trained to extract multi-dimensional visual features. For example, Figure 5 As shown, step S30, which involves extracting visual features from each key video frame in the key video frame sequence to obtain the multi-dimensional visual features corresponding to each key video frame, includes the following steps S31-S32: S31: Input each key video frame into the visual feature extraction model.
[0050] Before step S31, a visual feature extraction model can be pre-trained. After obtaining the key video frame sequence, each key video frame in the sequence can be input into the visual feature extraction model. In some embodiments, the visual feature extraction model can be the Qwen-VL visual large model, which has powerful image detail recognition and text generation capabilities, and can output structured features that meet the requirements based on prompt words.
[0051] S32: The visual feature extraction model is controlled by preset prompts to output multi-dimensional visual features corresponding to each key video frame. The multi-dimensional visual features include spatial features, object features, text label features, and abnormal clue features.
[0052] Here, preset prompts refer to pre-written text instructions used to guide the visual feature extraction model to output results in a specific format and content. This embodiment employs the Chain of Thought (CoT) prompting engineering technique, which standardizes the model's output behavior through clear task requirements and output format constraints.
[0053] Multidimensional visual features refer to a set of structured features that describe the content of a video frame from multiple independent semantic dimensions, specifically including: Spatial characteristics: Describe the physical spatial attributes of the scene, such as floor height, lighting conditions, and spatial layout type; Object characteristics: Describe the various physical objects that exist in the scene, such as office equipment, daily necessities, and display items; Textual signage features: Describes identifiable textual information in a scene, such as any readable signboard text, poster text, sign text, etc. Abnormal clue characteristics: Describes unusual situations in the scene that do not conform to the logic of a normal business place, such as being too crowded or having a strong sense of daily life.
[0054] In this step, pre-designed prompts can force the visual feature extraction model to ignore overall scene type guesses and objectively extract only multi-dimensional detailed features from the image, ensuring the objectivity and accuracy of feature extraction. Specifically, when inputting key video frames, pre-designed prompts are also input into the visual feature extraction model. The prompts explicitly require the visual feature extraction model, acting as an environmental observation expert, to describe only objectively existing detailed information in the image and output results according to a specified field format. Based on the prompts, the visual feature extraction model analyzes the content of the key video frames and extracts visual features in the corresponding dimensions.
[0055] For example, in a scenario involving the verification of financial credit business premises, the preset prompts input into the visual feature extraction model could be: "As an environmental observation expert, please ignore the overall scene type guesses and only objectively describe the details of the scene. Please output a strict JSON format containing the following fields: {'spatial_features':['Is the ceiling height high?','Is there natural lighting?','Spatial layout type'],'objects':['Names of displayed items','Names of office equipment','Names of household items'],'text_signs':['Any readable signs, posters, or logos in the scene'],'abnormal_clues':['Is it too crowded?','Is there a noticeable sense of life?']}. If a field has no corresponding content, please output 'None'." For a key video frame containing "desk, laptop, bunk bed", the model outputs the following multi-dimensional visual features: {'spatial_features':['low ceiling height','natural lighting','single room layout'],'objects':['desk','laptop','bunk bed,'folding table'],'text_signs':['none'],'abnormal_clues':['obvious signs of daily life']}. Alternatively, please act as an expert observing medical settings, ignoring the overall scene type guesses and objectively describing only the details of the image. Please output in strict JSON format, including the following fields: {"spatial_features":["Does the floor height meet the standard?","Is there an independent treatment area?","Is there a disinfection area sign?"],"medical_objects":["Treatment table","Stethoscope","Blood pressure monitor","Medicine cabinet","Disinfection equipment"],"text_signs":["Medical institution practice license","Drug name","Disinfection record board"],"abnormal_clues":["Are there kitchen utensils?","Are there bedding?","Are there miscellaneous items stored?"]} If a field has no corresponding content, please output "None".
[0056] Steps S31-S32 above, by sending key video frames into the visual feature extraction model and using preset prompts to standardize the output content, can selectively extract four types of multi-dimensional visual information: space, objects, text labels, and abnormal clues, achieving comprehensive and structured acquisition of image details. By relying on prompts to constrain the model's output format, the visual model illusion problem is effectively reduced, ensuring that feature extraction is objective and regular, and providing comprehensive, standard, and reliable raw feature data for subsequent feature classification and conflict determination.
[0057] S40: Classify the multi-dimensional visual features corresponding to each key video frame to obtain multiple types of visual features, and count the frequency of occurrence of each type of visual feature to obtain visual feature statistics.
[0058] Among them, visual feature statistics refer to the set of information that records each type of visual feature and its frequency of occurrence in all key video frames, which is used to reflect the various types of visual features appearing in the video frames to be identified and their degree of occurrence in the video to be identified.
[0059] In this step, the multi-dimensional visual features corresponding to all key video frames are analyzed, and visual features with the same or similar semantics are grouped into the same category to obtain multiple categories of visual features. The total number of times each category of visual features appears in all key video frames is counted to generate visual feature statistics that include feature categories and corresponding frequencies.
[0060] For example, suppose we obtain four types of visual features, namely visual feature F1, visual feature F2, visual feature F3, and visual feature F4. Among them, visual feature F1 appears a total of 3 times in all key video frames, visual feature F2 appears a total of 2 times in all key video frames, visual feature F3 appears a total of 1 time in all key video frames, and visual feature F4 appears a total of 4 times in all key video frames. Then the visual feature statistics can be: [visual feature F1: 3 times, visual feature F2: 2 times, visual feature F3: 1 time, visual feature F4: 4 times].
[0061] In some embodiments, cosine similarity can be used to classify the visual features of each key video frame across various dimensions. For example, Figure 6 As shown, step S40, which involves classifying the multi-dimensional visual features corresponding to each key video frame to obtain multiple categories of visual features, includes the following steps S41-S43: S41: Vectorize the visual features of each key video frame in each dimension to obtain multiple visual feature vectors.
[0062] In this step, the multi-dimensional visual features of all keyframes are flattened to extract all independent text descriptions under the four fields spatial_features, objects, text_signs, and abnormal_clues in the JSON, forming a visual feature list. A pre-trained lightweight text embedding model is then used to convert each visual feature in the list into a fixed-dimensional feature vector. In some embodiments, this lightweight text embedding model can be a Text2Vec or BGE-m3 lightweight text embedding model.
[0063] For example, in a financial credit operation site verification scenario, multiple visual features such as "a large desk," "a laptop on the desk," "bunk beds," "a folding table," "a computer," "a workstation with a computer," and "a cash register" can be extracted from the structured JSON output of multiple key video frames, resulting in 32 independent visual features. The BGE-m3 model is then used to convert each visual feature description into a high-dimensional feature vector, resulting in multiple visual feature vectors, such as 32.
[0064] S42: Calculate the cosine similarity between any two visual feature vectors.
[0065] In this step, the cosine similarity between any two visual feature vectors can be calculated. For example, for the text feature descriptions in the previous example, the cosine similarity between the visual feature vectors corresponding to "a large desk" and "a laptop on the desk," the cosine similarity between the visual feature vectors corresponding to "a large desk" and "a folding bed," and the cosine similarity between the visual feature vectors corresponding to "a folding bed" and "a computer" can be calculated. The cosine similarity value ranges from [-1, 1]. The closer the value is to 1, the more similar the semantics of the two features are; the closer the value is to -1, the more opposite the semantics are.
[0066] S43: Based on cosine similarity, the multi-dimensional visual features corresponding to each key video frame are classified to obtain multiple types of visual features.
[0067] In this step, visual features with the same or similar semantics can be merged into the same category, eliminating duplicate features and forming a concise global feature set. Specifically, when the cosine similarity of two visual feature vectors is greater than or equal to a threshold, their corresponding visual features are determined to be the same semantic feature and merged into one category; when the cosine similarity of two visual feature vectors is less than the threshold, their corresponding visual features are determined to be different semantic features and classified separately. During the merging process, the frequency information of all features in the key video frame sequence is preserved.
[0068] For example, multiple visual features can be merged into multiple categories of visual features, and the frequency of occurrence of each category of visual features can be counted to obtain visual feature statistics: [desk: 8 times, laptop: 6 times, bunk bed: 5 times, folding table: 3 times, low ceiling height: 12 times, natural lighting: 10 times, single room layout: 12 times, obvious living atmosphere: 7 times]. Or [treatment table: 7 times, blood pressure monitor: 5 times, gas stove: 4 times, folding table: 3 times, no independent treatment area: 14 times, no disinfection area signage: 14 times, kitchen utensils: 6 times, piled-up miscellaneous items: 9 times].
[0069] The above steps S41-S43 first convert various visual features into feature vectors, then measure the semantic similarity through cosine similarity and complete feature aggregation and classification. This can quickly merge duplicate semantic features, eliminate redundant information, and form a concise and unified global feature set. This not only compresses the data volume and reduces the pressure of subsequent processing, but also organizes and sorts out various valid evidence in the scene, laying a solid and reliable data foundation for determining subsequent conflict relationships.
[0070] S50: Based on a pre-defined feature-mutually exclusive knowledge graph, determine the scene conflict relationships between various visual features and obtain scene conflict relationship information.
[0071] Scene conflict refers to contradictory relationships between different categories of visual features that do not conform to the logic of a typical scene. For example, in a normal office building setting, it is unusual for a large number of office equipment and personal items to appear simultaneously; if they do, it constitutes a scene conflict.
[0072] Scene conflict relationship information refers to the set of information that records all detected combinations of scene conflict features, which is used to reflect the abnormal scene features present in the video to be identified.
[0073] This step analyzes whether there are conflicting relationships between different categories of visual features that do not conform to the logic of a typical scene. When a conflicting feature combination is detected, a corresponding conflict record is generated, and all conflict records are aggregated to obtain scene conflict relationship information.
[0074] In some embodiments, a feature-mutually exclusive knowledge graph can be pre-constructed, and scene conflict relationship information can be obtained based on this feature-mutually exclusive knowledge graph. For example, Figure 7 As shown, step S50, which is to determine the scene conflict relationships between various visual features based on the preset feature mutual exclusion knowledge graph and obtain scene conflict relationship information, includes the following steps S51-S52: S51: Based on a pre-defined feature-mutually exclusive knowledge graph, analyze whether there are mutually exclusive relationships between visual features of different categories, and obtain the analysis results.
[0075] The pre-defined feature-mutually exclusive knowledge graph refers to a pre-constructed feature-mutually exclusive relationship that should not occur simultaneously in a normal business setting. Specifically, feature relationships that would not normally occur simultaneously in a single, formal business setting, but might occur simultaneously in unusual scenarios such as mixed-use commercial / residential spaces or home offices, can be defined as weakly mutually exclusive relationships. For example, commercial nodes (shelves, cash registers) and residential nodes (beds, kitchen stoves) have a weakly mutually exclusive relationship.
[0076] In this step, multiple visual features can be matched and traversed with a preset mutually exclusive knowledge graph to determine whether any two visual features have a mutually exclusive relationship, and thus obtain the analysis results.
[0077] For example, in the verification scenario of financial credit business premises, multiple visual features include two categories: cash registers and bunk beds. Through a pre-set feature mutual exclusion knowledge graph matching, it is determined that "cash register" and "bunk beds" have a weak mutual exclusion relationship. The simultaneous appearance of both does not conform to the logic of a legitimate business premises, and the analysis result indicates that they have a weak mutual exclusion relationship. Similarly, "treatment table" and "gas stove" have a weak mutual exclusion relationship, as do "medicine cabinet" and "food," and their simultaneous appearance does not conform to the logic of a legitimate medical premises.
[0078] S52: Based on the analysis results, generate scenario conflict relationship information.
[0079] In this step, when the analysis results indicate a weak mutual exclusion relationship between two types of visual features, a conflict marker can be triggered to record the mutually exclusive features and their corresponding frame positions. All conflict markers are then aggregated to generate scene conflict relationship information. If no mutual exclusion relationship exists, the scene conflict relationship information will be empty.
[0080] For example, in a scenario where financial credit business premises are being verified, there is a weak mutual exclusion relationship between the two visual features "cashier counter" and "bunk bed", and a weak mutual exclusion relationship between the two visual features "desk" and "wardrobe". Then, the generated scenario conflict relationship information can be: {"Conflict_List":["'cashier counter found in frame 3', 'bunk bed found in frame 8'", "'desk found in frame 5', 'wardrobe found in frame 9'"]}.
[0081] In the online medical service qualification application site verification scenario, “Conflict_List”: [“'Frame 2 found medical station', 'Frame 7 found gas stove', 'Frame 4 found medicine cabinet', 'Frame 10 found instant noodle box']}.
[0082] Steps S51-S52 above rely on a pre-set feature-mutually exclusive knowledge graph to complete feature relationship comparison and judgment, quickly identify mutually exclusive feature combinations in the scene and generate standardized conflict information, accurately discover illogical scene contradictions in the business premises, objectively retain the location of conflict and corresponding content, automatically sort out risk control anomaly clues, without the need for manual comparison and investigation, efficiently determine hidden anomalies in the scene, and provide intuitive and effective conflict basis for subsequent reasoning and judgment.
[0083] S60: Based on visual feature statistical information and scene conflict relationship information, the scene corresponding to the video to be identified is identified to obtain the recognition result.
[0084] The video scene recognition result refers to the final output judgment on the scene type corresponding to the video to be recognized. In verifying the scene at the business premises, the recognition result includes, but is not limited to, "office building", "street-front shop", "home office", and "commercial and residential use".
[0085] In this step, a comprehensive analysis can be performed based on visual feature statistical information and scene conflict relationship information to finally obtain the video scene recognition result corresponding to the video to be recognized.
[0086] In some embodiments, a video scene recognition model can be pre-trained, and the video scene recognition result can be output through this model. For example, Figure 8 As shown, step S60, which is to identify the scene corresponding to the video to be identified based on visual feature statistical information and scene conflict relationship information, and obtain the identification result, includes the following steps S61-S62: S61: Generate video scene recognition data based on visual feature statistics, scene conflict relationship information, and preset video scene recognition rules.
[0087] The preset video scene recognition rules can be business judgment logics pre-set for corresponding scenarios, such as video interview scenarios for financial lending and other businesses, used to quantitatively distinguish scene types such as formal business premises, mixed-use commercial and residential premises, and home disguised as business premises. For example, the preset video scene recognition rules may include: Rule 1: If there are a large number of daily necessities and beds, and a small amount of office equipment, even if the user claims it is an "office building", it must be judged as "home office" or "mixed-use commercial and residential"; Rule 2: If there is a cashier, price tags, and a street-facing view with natural light, it should be preferentially judged as a "street-facing shop".
[0088] In this step, visual feature statistics, scene conflict relationship information, and preset video scene recognition rules can be combined to generate video scene recognition data that is adapted to the input format of the video scene recognition model.
[0089] S62: Input the video scene recognition data into the video scene recognition model to recognize the scene corresponding to the video to be recognized and obtain the recognition result.
[0090] In this step, video scene recognition data can be input into the video scene recognition model, which is then forced to output a two-dimensional result: a standardized final judgment scene label and a text-based arbitration reasoning report. Finally, the scene label and reasoning report are integrated to generate a complete video scene recognition result. The video scene recognition model can be a large text model such as Qwen3 or GPT-4.
[0091] For example, in a financial credit business premises verification scenario, video scene recognition data is input into the Qwen3 text model. The Qwen3 text model reads the visual feature statistics and finds a large number of household items such as beds and wardrobes, while only a small number of office facilities, triggering rule 1; combining the mutual exclusion and conflict between the two sets of items, it finally outputs a two-dimensional recognition result: 1. Scenario tag: Working from home; 2. Arbitration Reasoning Report: "Although office items such as desks and documents appeared in the footage, the video scene recognition data detected personal furniture such as bunk beds and wardrobes. The proportion of personal items was too high, and there was a conflict between commercial items and personal items, triggering Rule 1. Therefore, it was determined that the user's declaration of an office building was untrue, and the actual business location was working from home."
[0092] In the scenario of venue verification for online medical service qualification applications, output two-dimensional results: Scene tag: Home disguised as a medical facility; The reasoning report states: "Although basic medical items such as examination tables and blood pressure monitors appear in the footage, the video scene recognition data detects household items such as gas stoves and food boxes. There are no separate treatment areas and disinfection areas marked, and there is a conflict between medical items and household items, which triggers Article 3 of the medical qualification verification rules. Therefore, it is determined that the compliant clinic declared by the user is not true." The system automatically triggers a qualification review warning, marks the case as high-risk, and transfers it to the health supervision department for manual review.
[0093] In some embodiments, when it is determined that the actual business location is inconsistent with the information declared by the user, a risk control warning is automatically triggered, the case is marked as a risk case and transferred to manual review.
[0094] Steps S61-S62 above abandon the traditional simple voting method and rely on visual feature statistical information combined with preset video scene recognition rules to complete logical reasoning. This can accurately capture subtle features that appear in a short period of time, effectively improving the discrimination accuracy of complex scenes. At the same time, it achieves multimodal task decoupling, with the visual model dedicated to extracting objective features and the large text model responsible for business logic deduction. Coupled with intelligent frame extraction strategy, it reduces computational costs and the risk of model illusion, as well as reducing interface call overhead. In addition, by synchronously outputting scene labels and complete reasoning reports, the judgment basis is clear and intuitive, which can directly assist financial institution staff in verifying conflicting points, greatly improving the efficiency of manual review and the overall reliability of risk control.
[0095] As can be seen, in the above scheme, by filtering key frames based on video frame clarity and inter-frame structural similarity, invalid image interference can be effectively eliminated; multi-dimensional visual features of key frames are extracted and classified and statistically analyzed, and the scene is comprehensively judged by combining the scene conflict relationship between features. This can identify mixed scenes with contradictory features, and the overall video scene recognition accuracy is high.
[0096] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0097] In one embodiment, a video scene recognition device is provided, which corresponds one-to-one with the video scene recognition method described in the above embodiments. For example... Figure 9 As shown, the video scene recognition device includes an acquisition module 101, a filtering module 102, an extraction module 103, a classification module 104, a determination module 105, and a recognition module 106. Detailed descriptions of each functional module are as follows: The acquisition module 101 is used to acquire the video to be identified, wherein the video to be identified includes a sequence of video frames to be identified; The filtering module 102 is used to filter key video frame sequences from the video frame sequence to be identified based on the clarity of the video frames to be identified and the structural similarity between the video frames to be identified. Extraction module 103 is used to extract visual features from each key video frame in the key video frame sequence to obtain multi-dimensional visual features corresponding to each key video frame. The classification module 104 is used to classify the multi-dimensional visual features corresponding to each key video frame, obtain multiple types of visual features, and count the frequency of occurrence of each type of visual feature to obtain visual feature statistics. The determination module 105 is used to determine the scene conflict relationships between various visual features based on a preset feature mutual exclusion knowledge graph, and obtain scene conflict relationship information. The recognition module 106 is used to recognize the scene corresponding to the video to be recognized based on the visual feature statistical information and the scene conflict relationship information, and obtain the recognition result.
[0098] In one embodiment, the filtering module 102 is specifically used for: Based on the structural similarity of adjacent video frames to be identified in the video frame sequence to be identified, candidate video frame sequences are selected from the video frame sequence to be identified; Based on the clarity of the candidate video frames in the candidate video frame sequence, a key video frame sequence is selected from the candidate video frame sequence.
[0099] In one embodiment, the filtering module 102 is specifically used for: The candidate video frames in the candidate video frame sequence are converted to grayscale. The Laplacian operator variance of the candidate video frames after grayscale processing is calculated to obtain the resolution of the candidate video frames.
[0100] In one embodiment, the extraction module 103 is specifically used for: Input each key video frame into the visual feature extraction model; The visual feature extraction model is controlled by preset prompts to output multi-dimensional visual features corresponding to each key video frame. The multi-dimensional visual features include spatial features, object features, text label features, and abnormal clue features.
[0101] In one embodiment, the classification module 104 is specifically used for: The visual features of each key video frame in each dimension are vectorized to obtain multiple visual feature vectors. Calculate the cosine similarity between any two visual feature vectors; Based on the cosine similarity, the multi-dimensional visual features corresponding to each key video frame are classified to obtain multiple categories of visual features.
[0102] In one embodiment, the determining module 105 is specifically used for: Based on a pre-defined feature-mutually exclusive knowledge graph, the system analyzes whether there are mutually exclusive relationships between different categories of visual features and obtains the analysis results. Based on the analysis results, scenario conflict relationship information is generated.
[0103] In one embodiment, the identification module 106 is specifically used for: Based on the visual feature statistics, the scene conflict relationship information, and the preset video scene recognition rules, video scene recognition data is generated. The video scene recognition data is input into the video scene recognition model to identify the scene corresponding to the video to be recognized and obtain the recognition result.
[0104] This invention provides a video scene recognition device that effectively eliminates invalid image interference by filtering key frames based on video frame clarity and inter-frame structural similarity; extracts multi-dimensional visual features of key frames and performs classification and statistics, and comprehensively judges the scene by combining the scene conflict relationship between features, which can identify mixed scenes with contradictory features, and the overall video scene recognition accuracy is high.
[0105] Specific limitations regarding the video scene recognition device can be found in the limitations of the intelligent question-answering method described above, and will not be repeated here. Each module in the aforementioned video scene recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.
[0106] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a video scene recognition method on the server side.
[0107] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 11 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a video scene recognition method on the client side.
[0108] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Acquire the video to be identified, wherein the video to be identified includes a sequence of video frames to be identified; Based on the clarity of the video frames to be identified in the video frame sequence and the structural similarity between the video frames to be identified, a key video frame sequence is selected from the video frame sequence to be identified. Visual features are extracted from each key video frame in the key video frame sequence to obtain multi-dimensional visual features corresponding to each key video frame. The multi-dimensional visual features corresponding to each key video frame are classified to obtain multiple categories of visual features, and the frequency of occurrence of each category of visual features is counted to obtain visual feature statistics. Based on a pre-defined feature-mutually exclusive knowledge graph, the scene conflict relationships between various visual features are determined, and scene conflict relationship information is obtained. Based on the visual feature statistics and the scene conflict relationship information, the scene corresponding to the video to be identified is identified, and the identification result is obtained.
[0109] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire the video to be identified, wherein the video to be identified includes a sequence of video frames to be identified; Based on the clarity of the video frames to be identified in the video frame sequence and the structural similarity between the video frames to be identified, a key video frame sequence is selected from the video frame sequence to be identified. Visual features are extracted from each key video frame in the key video frame sequence to obtain multi-dimensional visual features corresponding to each key video frame. The multi-dimensional visual features corresponding to each key video frame are classified to obtain multiple categories of visual features, and the frequency of occurrence of each category of visual features is counted to obtain visual feature statistics. Based on a pre-defined feature-mutually exclusive knowledge graph, the scene conflict relationships between various visual features are determined, and scene conflict relationship information is obtained. Based on the visual feature statistics and the scene conflict relationship information, the scene corresponding to the video to be identified is identified, and the identification result is obtained.
[0110] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.
[0111] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.
[0113] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.
[0114] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A video scene recognition method, characterized in that, include: Acquire the video to be identified, wherein the video to be identified includes a sequence of video frames to be identified; Based on the clarity of the video frames to be identified in the video frame sequence and the structural similarity between the video frames to be identified, a key video frame sequence is selected from the video frame sequence to be identified. Visual features are extracted from each key video frame in the key video frame sequence to obtain multi-dimensional visual features corresponding to each key video frame. The multi-dimensional visual features corresponding to each key video frame are classified to obtain multiple categories of visual features, and the frequency of occurrence of each category of visual features is counted to obtain visual feature statistics. Based on a pre-defined feature-mutually exclusive knowledge graph, the scene conflict relationships between various visual features are determined, and scene conflict relationship information is obtained. Based on the visual feature statistics and the scene conflict relationship information, the scene corresponding to the video to be identified is identified, and the identification result is obtained.
2. The video scene recognition method according to claim 1, characterized in that, The step of filtering key video frame sequences from the video frame sequence to be identified based on the sharpness of the video frames to be identified and the structural similarity between the video frames to be identified includes: Based on the structural similarity of adjacent video frames to be identified in the video frame sequence to be identified, candidate video frame sequences are selected from the video frame sequence to be identified; Based on the clarity of the candidate video frames in the candidate video frame sequence, a key video frame sequence is selected from the candidate video frame sequence.
3. The video scene recognition method according to claim 2, characterized in that, Before filtering key video frame sequences from the candidate video frame sequence based on the sharpness of the candidate video frames in the candidate video frame sequence, the method further includes: The candidate video frames in the candidate video frame sequence are converted to grayscale. The Laplacian operator variance of the candidate video frames after grayscale processing is calculated to obtain the resolution of the candidate video frames.
4. The video scene recognition method according to any one of claims 1 to 3, characterized in that, The step of extracting visual features from each key video frame in the key video frame sequence to obtain multi-dimensional visual features corresponding to each key video frame includes: Input each key video frame into the visual feature extraction model; The visual feature extraction model is controlled by preset prompts to output multi-dimensional visual features corresponding to each key video frame. The multi-dimensional visual features include spatial features, object features, text label features, and abnormal clue features.
5. The video scene recognition method according to any one of claims 1 to 3, characterized in that, The multi-dimensional visual features corresponding to each key video frame are categorized to obtain multiple categories of visual features, including: The visual features of each key video frame in each dimension are vectorized to obtain multiple visual feature vectors. Calculate the cosine similarity between any two visual feature vectors; Based on the cosine similarity, the multi-dimensional visual features corresponding to each key video frame are classified to obtain multiple categories of visual features.
6. The video scene recognition method according to any one of claims 1 to 3, characterized in that, The method, based on a pre-defined feature-mutually exclusive knowledge graph, determines scene conflict relationships among various visual features, obtaining scene conflict relationship information, including: Based on a pre-defined feature-mutually exclusive knowledge graph, the system analyzes whether there are mutually exclusive relationships between different categories of visual features and obtains the analysis results. Based on the analysis results, scenario conflict relationship information is generated.
7. The video scene recognition method according to any one of claims 1 to 3, characterized in that, The process of identifying the scene corresponding to the video to be identified based on the visual feature statistical information and the scene conflict relationship information, and obtaining the identification result, includes: Based on the visual feature statistics, the scene conflict relationship information, and the preset video scene recognition rules, video scene recognition data is generated. The video scene recognition data is input into the video scene recognition model to identify the scene corresponding to the video to be recognized and obtain the recognition result.
8. A video scene recognition device, characterized in that, include: The acquisition module is used to acquire the video to be identified, wherein the video to be identified includes a sequence of video frames to be identified; The filtering module is used to filter key video frame sequences from the video frame sequence to be identified based on the clarity of the video frames to be identified and the structural similarity between the video frames to be identified. The extraction module is used to extract visual features from each key video frame in the key video frame sequence to obtain multi-dimensional visual features corresponding to each key video frame. The classification module is used to classify the multi-dimensional visual features corresponding to each key video frame, obtain multiple categories of visual features, and count the frequency of occurrence of each category of visual features to obtain visual feature statistics. The determination module is used to determine the scene conflict relationships between various visual features based on a preset feature mutual exclusion knowledge graph, and obtain scene conflict relationship information; The recognition module is used to identify the scene corresponding to the video to be recognized based on the visual feature statistics and the scene conflict relationship information, and obtain the recognition result.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the video scene recognition method as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the video scene recognition method as described in any one of claims 1 to 7.