Video clip summarization system and method
The video segment summarization system utilizes tools such as OpenCV, librosa, and NLTK to extract video features and adjust weights based on user feedback. This solves the problems of low efficiency and low automation in existing technologies, achieving efficient and personalized video summarization.
Patent Information
- Application Number
- CN202610534884.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-17
AI Technical Summary
Existing video summarization methods rely on manual editing, which is inefficient, has low automation, and lacks user interaction, failing to meet the needs of large-scale video processing and personalization.
A video segment summarization system is adopted, including a video content analysis module, a summarization generation module, a user interaction module, and a retrieval collaboration module. Visual, audio, and textual features of the video are extracted using tools such as OpenCV, librosa, and NLTK. Summarization is generated by combining feature quantization and temporal alignment, and feature weights are adjusted based on user feedback.
It improves the automation and efficiency of summary generation, reduces labor costs, ensures the accuracy and personalization of summaries, enhances the system's intelligence and flexibility, and meets the needs of large-scale video processing.
Smart Images
Figure CN122420622A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of video processing technology, and in particular to a video clip summarization system and method. Background Technology
[0002] In recent years, video, as a core carrier of information dissemination and content presentation, has seen explosive growth in output, widely penetrating into multiple fields such as video surveillance, film and television production, online education, and social media. The massive amount of video content provides users with a wealth of information choices, but it also brings the problem of information overload. Users need to sift through a huge video library to find content that meets their needs. Although existing video retrieval systems can help users quickly locate specific video clips, users still need to spend a lot of time browsing videos to obtain key information. Existing video summarization methods mainly rely on manual editing or simple editing tools, which have the following drawbacks: Manual editing is inefficient: The generation of abstracts relies heavily on manual screening and editing of videos segment by segment, which is time-consuming and labor-intensive and easily affected by subjective factors and fatigue, resulting in information omissions or redundancy, and cannot meet the needs of large-scale video processing. Low level of automation: Existing automated summarization methods are based on fixed editing rules and cannot accurately extract the core content of the video; Lack of user interaction mechanism: There is no real-time interaction channel in the existing technology. Users can only passively receive the summary results, cannot provide feedback on their needs, cannot dynamically adjust the summary content according to user interests and preferences, and it is difficult to adapt to personalized needs. Summary of the Invention
[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a video segment summarization system and method.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: The video clip summarization system includes a video content analysis module, a summarization generation module, a user interaction module, and a retrieval collaboration module. The video content analysis module is used to extract video features, including visual features, audio features, and text features; The summary generation module generates video segment summaries based on the extracted video features; The user interaction module is used to provide a user feedback entry point, enabling users to view and provide feedback on video clip summaries; The search collaboration module is integrated with the video search system, allowing users to directly view video clip summaries when searching for videos.
[0005] This invention also proposes a video segment summarization method, comprising the following steps: S1: Perform video content analysis and extract video features; The video features include visual features, audio features, and text features; S11: Extract visual features from the target video; The visual features include objects, scenes, etc.; S111: Identify objects and extract their features; The video content analysis module loads the target video through the video reading interface of the OpenCV library and extracts image frames frame by frame; the extracted image frames are preprocessed with grayscale conversion, Gaussian blur, etc.; the target video contains a unique video ID. The pre-trained classifier model built into the OpenCV library is called. The pre-trained classifier model contains target object features and confidence thresholds. The cv2.CascadeClassifier.detectMultiScale function is called to scan the preprocessed image frames. If the features of any region in the image frame match the features of the target object contained in the pre-trained classifier model, the cv2.CascadeClassifier.detectMultiScale function outputs the position coordinates and confidence of the object in this region. The output confidence score is compared with a confidence score threshold. Objects with a confidence score higher than the confidence score threshold are retained, and object features are extracted from the image frame. The object features include location coordinates, confidence score, and target object features. S112: Extract scene features; The average brightness of the preprocessed image frame is calculated using the cv2.mean function, and the background texture in the image frame is identified using the cv2.Canny algorithm to distinguish scene categories, which include indoor scenes or outdoor scenes. Based on the object features detected in step S111, the corresponding scene category is determined, including meeting scene, classroom scene, etc.; and scene features are output, including scene category, scene type, average brightness, background texture, etc. S12: Extract audio features from the target video; The video content analysis module extracts the audio track from the target video and converts it into a standard audio format to obtain an audio file; The librosa library is called to load audio files, and built-in functions are used to extract audio features from the audio files, including frequency features, amplitude features, rhythm features, timbre features, etc. S13: Extract text features from the target video; The text features include keywords, core sentences, semantic logic, etc. The video content analysis module obtains the title information and the corresponding formatted subtitle file of the target video, and uses a parsing tool to read the formatted subtitle file and title information and convert them into plain text strings; The plain text string was split into independent sentences using tools such as spaCy. The word frequency statistics function of NLTK tool was used to filter out words with high frequency as keywords. The entity recognition function of spaCy tool was used to identify entities in the text and add the corresponding nouns to the keywords. The independent sentences were vectorized using the spaCy tool, and the TF-IDF value of each independent sentence was obtained; the sentences with high TF-IDF values were selected as core sentences. Load the spaCy pre-trained model, sort out the logical relationships between core sentences and identify logical connectors between core sentences, construct a semantic network based on keywords and logical connectors, and obtain semantic logic based on the semantic network. Output text features such as keywords, core sentences, and semantic logic.
[0006] S2: Generate a summary of segments that reflect the core content of the video; The summary generation module pre-sets the weights of visual features, audio features, and text features, with the weight of text features > the weight of visual features > the weight of audio features. S21: Feature quantization and temporal alignment; The visual features, audio features, and text features extracted in step S1 are quantized to obtain visual quantized features, audio quantized features, text quantized features, and corresponding visual quantized values, audio quantized values, and text quantized values. The visual quantization value, audio quantization value, and text quantization value are aligned with the timestamp in the target video to obtain a feature group corresponding to each second. The feature group includes the visual quantization value, audio quantization value, and text quantization value. S22: Calculate the overall score of the video segments and select key segments; The target video is divided into multiple video segments according to a set duration. The average value of the feature groups contained in each video segment is taken, and the average value is the visual quantization value, audio quantization value, and text quantization value of each video segment. The visual quantization value is multiplied by the weight of the visual feature, the text quantization value is multiplied by the weight of the text feature, and the audio quantization value is multiplied by the weight of the audio feature. The scores are then added together to obtain the overall score of the video segment. The summary generation module filters out video segments that score above a pre-set threshold as key segments. S23: Obtain the abstract fragment; The visual quantization features, audio quantization features, and text quantization features of each selected key segment are concatenated into a unified feature vector using a clustering algorithm. The similarity of each unified feature vector is calculated using a clustering algorithm, and each unified feature vector is divided into a preset similarity interval. Each similarity interval corresponds to a cluster. The key segment with the highest overall score in each cluster is selected as the representative segment; The representative segments are arranged and spliced together according to their timestamps to obtain the abstract segments.
[0007] S3: User interaction and dynamic adjustment; The user interaction module provides a visual user interface, including a web interface, which includes a summary fragment display and a user feedback entry point. The summary fragment display is used to show summary fragments to users; The user feedback entry includes a text input box and various preset feedback tags, such as "key content missing", "too much redundant information", "unreasonable segment order", and "accurate content". Users submit feedback through text input boxes or by directly checking feedback tags; the system collects the feedback-related data submitted by users and stores it in the database, binding it with the corresponding summary fragments; if the cumulative number of feedbacks for the same content exceeds a set threshold, the system adjusts the weights of the corresponding visual features, audio features, and text features, and generates a new summary fragment based on the adjusted weights, replacing the original summary fragment in the visual user interface.
[0008] S4: Integrate abstract fragments with video retrieval systems; The summary fragment is associated with the video ID, and the summary fragment and the target video are saved to the database of the video retrieval system; The retrieval collaboration module extracts keywords from the summary fragments and establishes an index mapping between keywords and video IDs. When a user enters a search term in the video retrieval system, the video retrieval system matches the corresponding keywords and returns the corresponding target video and video summary. Furthermore, the summary fragments include timestamps for each representative segment; clicking on a timestamp will take you to the corresponding segment content.
[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This method improves automation by extracting features from the video content analysis module, eliminating reliance on manual segment-by-segment editing and significantly increasing the efficiency of summary generation. It also reduces labor costs and processing time, making it particularly suitable for handling massive amounts of video content. Furthermore, by designing reasonable feature weight settings and adjusting the weights based on user feedback, it improves the objectivity and accuracy of the generated summaries and reduces subjective bias. By setting up a user feedback entry point, users can provide real-time feedback via text or by checking tags. The system automatically records the feedback and adjusts the feature weights according to the user's needs, ensuring the personalization of the summary. The system can continuously improve the generated summary based on the user's personalized needs, thereby enhancing the intelligence and flexibility of the system. Attached Figure Description
[0010] Figure 1 This is a flowchart of the video clip summary generation method of the present invention. Detailed Implementation
[0011] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.
[0012] The video clip summarization system includes a video content analysis module, a summarization generation module, a user interaction module, and a retrieval collaboration module. The video content analysis module is used to extract video features, including visual features, audio features, and text features; The summary generation module generates video segment summaries based on the extracted video features; The user interaction module is used to provide a user feedback entry point, enabling users to view and provide feedback on video clip summaries; The search collaboration module is integrated with the video search system, allowing users to directly view video clip summaries when searching for videos.
[0013] like Figure 1 As shown, the present invention also proposes a video segment summarization method, comprising the following steps: S1: Perform video content analysis and extract video features; The video features include visual features, audio features, and text features; the visual features include objects, scenes, etc. S11: Extract visual features from the target video; S111: Identify objects and extract their features; The video content analysis module loads the target video through the video reading interface of the OpenCV library and extracts image frames one by one; the extracted image frames are preprocessed such as grayscale conversion and Gaussian blur. Specifically, the cv2.cvtColor function is used to uniformly convert image frames into grayscale images, reducing color interference and computational load; the cv2.GaussianBlur function is used to perform Gaussian blur processing on each grayscale image to remove noise in the grayscale image, such as image graininess and slight distortion. The pre-trained classifier model built into the OpenCV library is called. The pre-trained classifier model contains target object features and confidence thresholds. The cv2.CascadeClassifier.detectMultiScale function is called to scan the preprocessed image frames. If the features of any region in the image frame match the features of the target object contained in the pre-trained classifier model, the cv2.CascadeClassifier.detectMultiScale function outputs the position coordinates and confidence of the object in this region. The output confidence score is compared with a confidence score threshold. Objects with a confidence score higher than the confidence score threshold are retained, and object features are extracted from the image frame. The object features include location coordinates, confidence score, and target object features. S112: Extract scene features; The average brightness of the preprocessed image frame is calculated using the cv2.mean function, and the background texture in the image frame is identified using the cv2.Canny algorithm to distinguish between indoor and outdoor scenes (such as the regular texture of indoor walls and the open outline of the outdoor sky). Specifically, if the detected average brightness is higher than the set threshold, it is determined to be high brightness and identified as a dense and regular texture, thus determining it to be an indoor scene; if the detected average brightness is lower than the set threshold, it is determined to be low brightness and identified as an edge texture, thus determining it to be an outdoor scene. The corresponding scene category is determined by combining the object features detected in step S111, including meeting scene, classroom scene, etc. For example, if multiple people are detected clustered together, accompanied by objects such as tables, chairs, and whiteboards, it is determined to be a meeting scene; If a few people are detected in the foreground of the image and most people are detected in the background or sides of the image, and a blackboard, projection screen, or podium are detected in the foreground, then it is determined to be a classroom scene. Based on the object features detected in step S111, the corresponding scene category is determined, including meeting scene, classroom scene, etc.; and scene features are output, including scene category, scene type, average brightness, background texture, etc. S12: Extract audio features from the target video; The video content analysis module extracts the audio track from the target video and converts it into a standard audio format to obtain an audio file; Specifically, the video stream is removed from the target video using tools such as FFmpeg, the audio track is extracted, and the extracted audio track is converted into WAV format; The librosa library is called to load audio files, and built-in functions are used to extract audio features from the audio files, including frequency features, amplitude features, rhythm features, timbre features, etc. Specifically, the audio features include MFCC coefficients, etc. The librosa.load function of the librosa library is called to load the processed audio file, and the MFCC coefficients are extracted from the audio file using the librosa.feature.mfcc function. The amplitude characteristics include root-mean-square energy, etc. The librosa.feature.rms function is called to extract the root mean square energy from the audio file. The rhythmic features include beats, etc.; The librosa.beat.beat_track function is called to obtain the beat density from the audio file; The timbre features include chromaticity features, etc. The librosa.feature.chroma_stft function is called to extract chroma features from the audio file; S13: Extract text features from the target video; The text features include keywords, core sentences, semantic logic, etc. The video content analysis module obtains the title information and the corresponding formatted subtitle file of the target video. It reads the formatted subtitle file and title information and converts them into plain text strings using a parsing tool. Then, it performs semantic parsing on the plain text strings using tools such as spaCy to obtain text features such as keywords, core sentences, and semantic logic. Specifically, plain text strings are split into independent sentences using tools such as spaCy; high-frequency words are selected as keywords using the word frequency statistics function of NLTK tool; and entities in the text are identified using the entity recognition function of spaCy tool, and the nouns corresponding to the entities are added to the keywords. The independent sentences were vectorized using the spaCy tool, and the TF-IDF value of each independent sentence was obtained; the independent sentences with high TF-IDF values were selected as core sentences. Load the spaCy pre-trained model, sort out the logical relationships between core sentences (such as causal relationships, progressive relationships, etc.) and identify logical connectors between core sentences. Construct a semantic network based on keywords and logical connectors, and obtain semantic logic based on the semantic network. Output text features such as keywords, core sentences, and semantic logic.
[0014] S2: Generate a summary of segments that reflect the core content of the video; The summary generation module pre-sets the weights of visual features, audio features, and text features, with the weight of text features > the weight of visual features > the weight of audio features. S21: Feature quantization and temporal alignment; The visual features, audio features, and text features extracted in step S1 are quantized to obtain the corresponding visual quantization values, audio quantization values, and text quantization values. Specifically, visual feature quantification: the system includes an object matching degree table and a scene category matching degree table, wherein the object matching degree table and the scene category matching degree table contain corresponding object matching scores and scene matching scores; Based on the object matching degree table and the scene category matching degree table, the object matching score of the object features extracted in step S1 and the scene matching score of the scene features are obtained; the confidence, object matching score and scene matching score are normalized and mapped to the interval [0, 1], and the average value is taken as the visual quantization value. Audio feature quantization: The system includes an audio feature scoring table, which contains scores corresponding to different MFCC coefficients, scores corresponding to different root mean square energies, scores corresponding to different beat densities, and scores corresponding to different chromaticity features. Based on the audio feature scoring table, the scores corresponding to MFCC coefficients, root mean square energy, beat density, and chroma features are obtained, and then normalized and mapped to the interval [0, 1], and the average value is taken as the audio quantization value. Text feature quantification: Calculate the word frequency ratio of each keyword in the formatted caption file and title information, as well as the coverage rate of semantic logic in all core sentences (for example, if all core sentences have this logic, the coverage rate is 100%). The text quantization value is obtained by multiplying the preset word frequency weight, TF-IDF value weight, and semantic logic weight in the system by the corresponding word frequency weight, TF-IDF value, and semantic logic coverage, respectively. The visual quantization value, audio quantization value, and text quantization value are aligned with the timestamp in the target video to obtain a time series of one feature group per second, wherein the feature group contains the visual quantization value, audio quantization value, and text quantization value. S22: Calculate the overall score of the video segments and select key segments; The target video is divided into multiple video segments according to a set duration. The average value of the feature groups contained in each video segment is taken, and the average value is the visual quantization value, audio quantization value, and text quantization value of each video segment. The visual quantization value is multiplied by the weight of the visual feature, the text quantization value is multiplied by the weight of the text feature, and the audio quantization value is multiplied by the weight of the audio feature. The scores are then added together to obtain the overall score of the video segment. The summary generation module filters out video segments that score above a pre-set threshold as key segments. S23: Obtain the abstract fragment; The visual quantization features, audio quantization features, and text quantization features of each selected key segment are concatenated into a unified feature vector using a clustering algorithm. The similarity of each unified feature vector is calculated using a clustering algorithm, and each unified feature vector is divided into a preset similarity interval. Each similarity interval corresponds to a cluster. The key segment with the highest overall score in each cluster is selected as the representative segment; The representative segments are arranged and spliced together according to their timestamps to obtain the abstract segments.
[0015] S3: User interaction and dynamic adjustment; The user interaction module provides a visual user interface, including a web interface, which includes a summary fragment display and a user feedback entry point. The summary fragment display is used to show summary fragments to users; The user feedback entry includes a text input box and various preset feedback tags, such as "key content missing", "too much redundant information", "unreasonable segment order", and "accurate content". Users submit feedback through text input boxes or by directly checking feedback tags; the system collects the feedback-related data submitted by users and stores it in the database, binding it with the corresponding summary fragments; if the cumulative number of feedbacks for the same content exceeds a set threshold, the system adjusts the weights of the corresponding visual features, audio features, and text features, and generates a new summary fragment based on the adjusted weights, replacing the original summary fragment in the visual user interface.
[0016] S4: Integrate abstract fragments with video retrieval systems; The summary fragment is associated with the video ID, and the summary fragment and the target video are saved to the database of the video retrieval system; The retrieval collaboration module extracts keywords from the summary fragments and establishes an index mapping between keywords and video IDs. When a user enters a search term in the video retrieval system, the video retrieval system matches the corresponding keywords and returns the corresponding target video and video summary. Furthermore, the summary fragments include timestamps for each representative segment; clicking on a timestamp will take you to the corresponding segment content.
[0017] This method constructs an automated video content analysis module, leveraging tools such as OpenCV, librosa, and NLTK to extract multi-dimensional features from videos, replacing manual frame-by-frame screening. It employs feature quantization and temporal alignment, comprehensive score calculation to select key segments, and clustering algorithms to select representative segments and splice them by timestamp, eliminating the need for manual editing decisions. Integration with video retrieval systems eliminates the need for manual indexing. By reducing reliance on manual editing, it significantly lowers labor costs, improves processing efficiency, and meets the demands of large-scale video batch processing. It avoids information omissions or redundancy caused by subjective human judgment and fatigue, ensuring the stability and consistency of summary generation. Furthermore, it reduces the time cost for users' subsequent searches and browsing, eliminating the need for manual assistance in locating key segments and improving the efficiency of video information retrieval.
[0018] A feedback-driven dynamic adjustment mechanism is established, which binds and stores user feedback with summary fragments. When the cumulative number of the same feedback exceeds a set threshold, the weights of visual, audio, and text features are automatically adjusted, and a new summary is generated. This breaks the pattern of users passively receiving summaries, giving users the authority to participate in summary optimization and achieving personalized needs adaptation. It enhances the system's flexibility and adaptability, enabling it to cope with the different needs of different scenarios and user groups.
[0019] Seamlessly integrated with video retrieval systems, users can directly view summaries when searching for videos, determining whether a video meets their needs without having to watch the entire video, significantly improving video browsing and filtering efficiency; the timestamp jump function helps users quickly locate core segments of interest, reducing unnecessary browsing time and further optimizing the user experience.
[0020] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.
Claims
1. A method for generating video segment summaries, characterized in that: Includes the following steps: S1: Perform video content analysis and extract video features; S11: Extract visual features from the target video; S12: Extract audio features from the target video; S13: Extract text features from the target video; S2: Generate a summary of segments that reflect the core content of the video; The abstract generation module pre-sets the weights of visual features, audio features, and text features, and includes the following sub-steps: S21: Feature quantization and temporal alignment; S22: Calculate the overall score of the video segments and select key segments; S23: Obtain the abstract fragment; S3: User interaction and dynamic adjustment; S4: Integrate summary fragments with video retrieval systems.
2. The video segment summarization method as described in claim 1, characterized in that: The specific details of step S1 are as follows: S11: Extract visual features from the target video; S111: Identify objects and extract their features; The video content analysis module loads the target video through the video reading interface of the OpenCV library and extracts image frames one by one; the extracted image frames are then preprocessed. The pre-trained classifier model built into the OpenCV library is called. The pre-trained classifier model contains target object features and confidence thresholds. The cv2.CascadeClassifier.detectMultiScale function is called to scan the preprocessed image frames. If the features of any region in the image frame match the features of the target object contained in the pre-trained classifier model, the cv2.CascadeClassifier.detectMultiScale function outputs the position coordinates and confidence of the object in this region. The output confidence score is compared with a confidence score threshold. Objects with a confidence score higher than the confidence score threshold are retained, and object features are extracted from the image frame. The object features include location coordinates, confidence score, and target object features. S112: Extract scene features; The average brightness of the preprocessed image frame is calculated using the cv2.mean function, and the background texture in the image frame is identified using the cv2.Canny algorithm to distinguish between indoor and outdoor scenes. And determine the corresponding scene category by combining the object features detected in step S111; S12: Extract audio features from the target video; The video content analysis module extracts the audio track from the target video and converts it into a standard audio format to obtain an audio file; The librosa library is called to load audio files, and built-in functions are used to extract audio features from the audio files, including frequency features, amplitude features, rhythm features, and timbre features. S13: Extract text features from the target video; The text features include keywords, core sentences, and semantic logic; The video content analysis module obtains the title information and the corresponding formatted subtitle file of the target video. It reads the formatted subtitle file and title information and converts them into plain text strings using a parsing tool. The plain text string is then semantically parsed using the spaCy tool to obtain text features of keywords, core sentences, and semantic logic. Finally, the text features of keywords, core sentences, and semantic logic are output.
3. The video segment summarization method as described in claim 1, characterized in that: The specific details of step S2 are as follows: S21: Feature quantization and temporal alignment; The visual features, audio features, and text features extracted in step S1 are quantized to obtain the corresponding visual quantization values, audio quantization values, and text quantization values. The visual quantization value, audio quantization value, and text quantization value are aligned with the timestamp in the target video to obtain a feature group corresponding to each second. The feature group includes the visual quantization value, audio quantization value, and text quantization value. S22: Calculate the overall score of the video segments and select key segments; The target video is divided into multiple video segments according to a set duration. The average value of the feature groups contained in each video segment is taken, and the average value is the visual quantization value, audio quantization value, and text quantization value of each video segment. The visual quantization value is multiplied by the weight of the visual feature, the text quantization value is multiplied by the weight of the text feature, and the audio quantization value is multiplied by the weight of the audio feature. The scores are then added together to obtain the overall score of the video segment. The summary generation module filters out video segments that score above a pre-set threshold as key segments. S23: Obtain the abstract fragment; The visual quantization features, audio quantization features, and text quantization features of each selected key segment are concatenated into a unified feature vector using a clustering algorithm. The similarity of each unified feature vector is calculated using a clustering algorithm, and each unified feature vector is divided into a preset similarity interval. Each similarity interval corresponds to a cluster. The key segment with the highest overall score in each cluster is selected as the representative segment; The representative segments are arranged and spliced together according to their timestamps to obtain the abstract segments.
4. The video segment summarization method as described in claim 1, characterized in that: The specific details of step S3 are as follows: The user interaction module provides a visual user interface, which includes a summary fragment display and a user feedback entry point; The summary segment display is used to show summary segments to users; the user feedback entry includes a text input box and a variety of preset feedback tags; Users submit feedback through text input boxes or by directly checking feedback tags; the system collects the feedback-related data submitted by users and stores it in the database, binding it with the corresponding summary fragments; if the cumulative number of feedbacks for the same content exceeds a set threshold, the system adjusts the weights of the corresponding visual features, audio features, and text features, and generates a new summary fragment based on the adjusted weights, replacing the original summary fragment in the visual user interface.
5. The video segment summarization method as described in claim 1, characterized in that: The specific details of step S4 are as follows: The summary fragment is associated with the video ID, and the summary fragment and the target video are saved to the database of the video retrieval system; The retrieval collaboration module extracts keywords from the summary fragments and establishes an index mapping between keywords and video IDs. When a user enters search terms into the video retrieval system, the system matches the corresponding keywords and returns the corresponding target video and video summary.
6. A video segment summarization system for implementing the method of any one of claims 1-5, characterized in that: It includes a video content analysis module, a summary generation module, a user interaction module, and a search collaboration module; The video content analysis module is used to extract video features, including visual features, audio features, and text features; The summary generation module generates video segment summaries based on the extracted video features; The user interaction module is used to provide a user feedback entry point, enabling users to view and provide feedback on video clip summaries; The search collaboration module is integrated with the video search system, allowing users to directly view video clip summaries when searching for videos.