Video efficient splitting method based on picture similarity and semantic clustering

Abstract

The invention discloses a high-efficiency video splitting method based on picture similarity and semantic clustering, which comprises the following steps of: firstly, extracting picture frames in a video according to a fixed frame rate, calculating the similarity of adjacent frames by utilizing a picture coding model, and combining the frames with high similarity into a primary video clip; thirdly, extracting an audio and transferring the audio into a text, performing clustering analysis on text segments through text semantic vectorization and affinity clustering, and merging video segments according to a clustering result; and meanwhile, for video clips with insufficient lengths, intelligent merging is performed by dynamically calculating the similarity of adjacent clips, so that the continuity and accuracy of strip splitting are ensured. According to the method, the lightweight model and parallel processing are combined, the processing efficiency is greatly improved, the problems that in the prior art, strip splitting is too fine and depends on character recognition are solved, and the method has high universality and adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent editing and stripping of news videos, and in particular to an efficient video stripping method based on picture similarity and semantic clustering. Background Art

[0002] In the current information society, the amount of video data has increased dramatically, far exceeding the ability of individual humans to process information. In order to help users quickly obtain effective information, video stripping technology has emerged. It is mainly used for cutting and extracting content from long videos, such as news videos. By splitting long videos into multiple small clips, users can quickly obtain the core content. This technology is widely used in news videos, live playback, film and television editing and other scenarios.

[0003] Existing technologies usually use face recognition or scene change detection to determine the switching points of news video clips. The working principle is based on the characters or scene changes that appear in the video. For example, when the host or other characters are detected to appear on the screen, the system will use it as the segmentation point of the news clip. However, this method has significant problems in practical applications: 1. Over-reliance on the appearance of people: In many news clips, especially those that are purely news reports or street scenes, people do not always appear in the video. In this case, relying on character recognition to split the clips can easily lead to the inability to accurately segment the clips. For example, in the transitional screen between two news clips where no host appears, the system may not be able to correctly identify the segmentation point and miss important content.

[0004] 2. Insufficient ability to recognize non-face images: In some news, for privacy or copyright protection, the faces of people will be coded or hidden, causing the stripping technology that relies on face recognition to fail to work properly, affecting the segmentation effect.

[0005] 3. The processing effect is not refined: Existing methods can usually only roughly segment the video according to scene changes or the appearance of characters, and cannot further perform fine-grained segmentation based on the semantics of the content. Especially when faced with inadequate speech recognition and complex background sound effects, the coherence and accuracy of video segmentation are poor.

[0006] Therefore, how to improve the accuracy of news video segmentation, especially when no characters appear or the background is complex, by combining image similarity and semantic clustering to ensure accurate segmentation and merging of video segments, has become a technical problem to be solved by the present invention. Summary of the invention

[0007] The technical problem solved by the present invention is to provide an efficient video stripping method based on image similarity and semantic clustering in response to the defects existing in the above-mentioned prior art, so as to solve the problems proposed in the above-mentioned background technology that the prior art stripping is too dependent on character recognition, the segments are too fragmented, the speech recognition is inaccurate and the background noise interference is large.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows: A video efficient stripping method based on image similarity and semantic clustering, the method comprising the following steps: Step 1, extracting picture frames from the video to be processed at a fixed frame rate; Step 2: Encode the extracted picture frames using the picture coding model, calculate the similarity between adjacent frames, and merge adjacent frames whose similarity exceeds the threshold into a video clip based on a preset similarity threshold; Step 3, extracting corresponding audio content from the video clip and converting the audio into text; Step 4, encoding the text to convert the text into a semantic vector; Step 5: Based on the semantic vector results of the text, an affinity clustering algorithm is used to cluster the text to obtain clustering results of the text fragments; Step 6, according to the clustering result, merging the video segments corresponding to the consecutive text segments belonging to the same clustering category into a complete news segment; Step 7: When the length of a certain video segment is less than the preset minimum number of frames, the image similarity between the segment and the adjacent segments is calculated, and the segment is merged with the adjacent segment with higher similarity; Step 8, summarize the content of the merged news clips to complete the video stripping.

[0009] As a further solution of the present invention, the picture similarity calculation used in step 2 is to determine the similarity of adjacent frames by performing a vector product calculation on the vectorized representation of each picture frame, and compare it with a preset similarity threshold to decide whether to merge the frames.

[0010] As a further solution of the present invention, the affinity clustering algorithm used in step 5 does not require a preset number of categories, and can dynamically cluster according to text semantic similarity, and classify text segments with high similarity into the same cluster category.

[0011] As a further solution of the present invention, in step 7, when the number of frames of a certain video segment is lower than a preset minimum number of frames, the similarity between the segment and the last frame of the previous video segment and the first frame of the next video segment is calculated, and the video segment is merged with the segment with higher similarity.

[0012] As a further solution of the present invention, the preset minimum frame number is 150 frames, and video clips with a frame number below this number are regarded as too short clips and trigger a dynamic merging mechanism.

[0013] As a further solution of the present invention, the text semantic vector used in step 6 is generated by a lightweight text encoding model, which can efficiently process a large amount of text in the video and calculate the text semantic similarity in parallel.

[0014] As a further solution of the present invention, the steps of merging and cutting the video clips are performed by parallel processing technology, thereby achieving efficient processing with a processing speed and a ratio of one tenth of the video duration.

[0015] As a further solution of the present invention, the video strip splitting method further includes performing semantic summarization of the content of the video segments after merging the video segments, and generating corresponding summary information to help users quickly understand the main content of the news segments.

[0016] As a further solution of the present invention, the audio-to-text technology in step 3 extracts the video audio at fixed intervals through a speech recognition model to ensure time alignment with the video clip.

[0017] As a further solution of the present invention, the method is applicable to news videos from different sources, and can accurately split and merge video clips even when there is no host, characters or background sound effects in the news.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. In view of the situation where hosts often appear in news videos, the stripping method of the present invention does not rely on the appearance of the host or face recognition to determine the switching points of the news segments, but performs precise cutting through dual analysis of image similarity and semantic clustering. In traditional technology, when the host continues to appear between multiple news segments, it is often difficult to accurately judge the actual switching points of the segments, resulting in miscuts or segment merging errors. The present invention effectively solves this problem. Even if the host appears frequently in different news segments, the system can still identify the real news switching points through semantic differences and changes in video content, ensuring the accuracy and continuity of the news segments, especially in complex scenarios, greatly improving the accuracy and applicability of video stripping.

[0019] 2. Double-layer cropping mechanism to ensure the accuracy and coherence of video stripping: This application combines image similarity with text semantic similarity, preliminarily crops video clips based on image similarity, and then merges the clips again based on text semantic clustering. This double cropping mechanism effectively avoids the problem of excessive fragmentation caused by relying solely on image similarity, while ensuring the coherence and semantic integrity of the text, greatly improving the accuracy of video stripping.

[0020] 3. Dynamic merging strategy to solve the problem of too short segments: This application sets a minimum frame limit and intelligently merges too short segments by dynamically calculating the image similarity of adjacent segments. This strategy not only effectively solves the problem of too short segments caused by threshold settings, but also ensures the logical integrity and visual coherence of the stripping results, and realizes more intelligent dynamic merging processing.

[0021] 4. Flexible application of affinity clustering algorithm without presetting the number of categories: Compared with traditional clustering algorithms, the affinity clustering method used in this application can dynamically determine the clustering results of texts without setting the number of categories in advance. This method is particularly suitable for application scenarios such as news videos where the number of clips is uncertain. It can flexibly handle multi-source news content and ensure the efficiency and accuracy of text merging.

[0022] 5. Lightweight model and parallel processing greatly improve processing efficiency: This application uses a lightweight AI model and adopts parallel processing technology to ensure extremely high processing efficiency when performing large-scale video stripping. In actual tests, the processing time of this method is only one-tenth of the video length, which is far better than existing technologies, can significantly reduce processing time and improve user experience.

[0023] 6. Adaptability to diverse scenarios and strong versatility: In news videos, this application can ensure the accuracy and coherence of video stripping, regardless of whether there are hosts or characters, or even when there are complex background sound effects (such as military music or street noise). This versatility and robustness make this method widely applicable to news video content from different sources and forms, and has strong industry adaptability.

[0024] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0026] Figure 1 The present invention is a flow chart of the efficient video stripping method based on image similarity and semantic clustering. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] See also Figure 1 In an embodiment of the present invention, a video efficient stripping method based on image similarity and semantic clustering is provided, the method comprising the following steps: Step 1, extracting picture frames from the video to be processed at a fixed frame rate; Step 2: Encode the extracted picture frames using the picture coding model, calculate the similarity between adjacent frames, and merge adjacent frames whose similarity exceeds the threshold into a video clip based on a preset similarity threshold; Step 3, extracting corresponding audio content from the video clip and converting the audio into text; Step 4, encoding the text to convert the text into a semantic vector; Step 5: Based on the semantic vector results of the text, an affinity clustering algorithm is used to cluster the text to obtain clustering results of the text fragments; Step 6, according to the clustering result, merging the video segments corresponding to the consecutive text segments belonging to the same clustering category into a complete news segment; Step 7: When the length of a certain video segment is less than the preset minimum number of frames, the image similarity between the segment and the adjacent segments is calculated, and the segment is merged with the adjacent segment with higher similarity; Step 8, summarize the content of the merged news clips to complete the video stripping.

[0029] As a further solution of the present invention, the picture similarity calculation used in step 2 is to determine the similarity of adjacent frames by performing a vector product calculation on the vectorized representation of each picture frame, and compare it with a preset similarity threshold to decide whether to merge the frames.

[0030] As a further solution of the present invention, the affinity clustering algorithm used in step 5 does not require a preset number of categories, and can dynamically cluster according to text semantic similarity, and classify text segments with high similarity into the same cluster category.

[0031] As a further solution of the present invention, in step 7, when the number of frames of a certain video segment is lower than a preset minimum number of frames, the similarity between the segment and the last frame of the previous video segment and the first frame of the next video segment is calculated, and the video segment is merged with the segment with higher similarity.

[0032] As a further solution of the present invention, the preset minimum frame number is 150 frames, and video clips with a frame number below this number are regarded as too short clips and trigger a dynamic merging mechanism.

[0033] As a further solution of the present invention, the text semantic vector used in step 6 is generated by a lightweight text encoding model, which can efficiently process a large amount of text in the video and calculate the text semantic similarity in parallel.

[0034] As a further solution of the present invention, the steps of merging and cutting the video clips are performed by parallel processing technology, thereby achieving efficient processing with a processing speed and a ratio of one tenth of the video duration.

[0035] As a further solution of the present invention, the video strip splitting method further includes performing semantic summarization of the content of the video segments after merging the video segments, and generating corresponding summary information to help users quickly understand the main content of the news segments.

[0036] As a further solution of the present invention, the audio-to-text technology in step 3 extracts the video audio at fixed intervals through a speech recognition model to ensure time alignment with the video clip.

[0037] As a further solution of the present invention, the method is applicable to news videos from different sources, and can accurately split and merge video clips even when there is no host, characters or background sound effects in the news.

[0038] Embodiment 1: In the actual work of news media, journalists or media organizations usually need to process a large amount of news video content in order to quickly extract key information for broadcasting or re-editing. However, when processing long videos, traditional news video stripping technology has problems such as over-reliance on character recognition, inaccurate voice recognition, and severe background noise interference, resulting in unsatisfactory stripping effects. For example, background images, street scenes, and scenes without people appearing in the camera that often appear in news broadcasts make it difficult for existing technologies to accurately identify news switching points, and thus cannot correctly segment videos.

[0039] Aiming at this scenario, this embodiment provides an efficient video stripping method based on image similarity and semantic clustering, which significantly improves the editing efficiency and accuracy of news videos.

[0040] In a specific application, the picture frames are first extracted from the news video at a fixed frame rate. For example, for a 30-minute news video, the corresponding picture frames are extracted from the video at a frame rate of 10 frames per second. Then, these picture frames are vectorized and encoded using a lightweight picture coding model, and the similarity between adjacent frames is calculated. By comparing with the preset similarity threshold, frames with higher similarity are regarded as the same scene and merged into a preliminary video clip. In news scenes, such as clips that continuously broadcast the same content, although there are slight changes in the picture in the video, these pictures with smaller changes can still be effectively identified as the same news clip through similarity calculation.

[0041] Subsequently, for the video clips generated above, the corresponding audio content is extracted from the video, and the audio is transcribed into text through speech recognition technology. After the transcription is completed, the text is semantically vectorized using a text encoding model to generate a semantic representation of the text. Taking a certain news broadcast as an example, when the broadcast content involves similar topics, such as continuous reporting of different developments of the same event, these texts will be identified as semantically similar segments through semantic vector analysis. Then, the system clusters the text semantic vectors through an affinity clustering algorithm, and aggregates text segments with similar semantics into the same category. In this embodiment, the advantage of the affinity clustering algorithm is that there is no need to preset the number of news segments, and it can dynamically cluster according to the natural distribution of semantic content, avoiding misjudgments caused by irregular news content structure.

[0042] Furthermore, the system merges the video clips corresponding to the consecutive text clips belonging to the same cluster category into a complete news clip. For example, when a news video reports an international event, the system automatically aggregates the relevant clips of the event through semantic analysis to form a complete report. This method not only ensures semantic coherence, but also avoids the problem of overly fragmented traditional image similarity cropping.

[0043] To address the problem of insufficient length of some video clips, the system has also introduced a dynamic merging mechanism. During the video stripping process, if the length of a certain clip is less than the preset minimum number of frames (such as 150 frames), the system will calculate the image similarity between the clip and the previous and next clips, determine which clip it has a higher similarity with, and merge them intelligently. For example, in a news video scene, some short transition shots or background images are short in length. These short clips will be effectively merged into the corresponding content clips by matching the similarity with adjacent news clips, thus avoiding the impact of too short clips on the overall coherence.

[0044] The solution of this embodiment can handle complex news video scenes, including broadcasts without characters, and scenes with severe background music or noise interference. In situations that are difficult to handle with traditional technologies, this solution relies on the dual analysis of image similarity and semantic clustering to ensure the precise cutting and merging of news clips, effectively improving the accuracy and efficiency of video stripping. Through lightweight models and parallel processing technology, the application of this solution in news video processing can shorten the stripping time to one-tenth of the video length, achieving fast processing while ensuring high-quality segmentation effects.

[0045] Finally, this embodiment ensures accurate segmentation of news content and semantic coherence and visual consistency between video clips when processing news videos. Especially in multi-scene and cross-topic news reports, the system can flexibly cope with complex content structures, avoiding the common mis-segmentation or missed-segmentation problems in the prior art, and significantly improving the work efficiency of news media in video processing.

[0046] Embodiment 2: In the news media industry, processing a large number of long news videos often faces the problem of long time and poor effect, especially in some special scenarios, such as news broadcasts without hosts or characters. In this case, the existing face recognition technology cannot effectively segment news clips, resulting in inaccurate segment cutting, affecting subsequent editing and publishing work. In addition, in news videos from different sources, the way of switching news is different. Especially in the absence of characters or complex backgrounds, traditional segmentation technology based on face recognition and scene changes often does not work well, and it is difficult to quickly and efficiently obtain effective information.

[0047] In order to solve this problem, this embodiment adopts an efficient news video stripping method based on image similarity and semantic clustering. The specific implementation steps are as follows: First, extract the screen of the news video at a fixed frame rate to generate a series of picture frame data. For example, for a news video, extract picture frames at a frequency of 10 frames per second. Then, use the picture coding model to vectorize these frames and calculate the picture similarity between consecutive frames. When the similarity exceeds the preset threshold, the system merges these similar frames into a preliminary video clip. This method can effectively avoid relying solely on the appearance of characters or face recognition for cutting, and is suitable for news broadcasting scenarios without hosts or characters.

[0048] Next, for each extracted video clip, the corresponding audio content is extracted from the video, and the audio is transcribed into text using speech recognition technology. Then, the text is converted into semantic vectors using a text encoding model, and clustered based on these semantic vector results. Taking long reports in news videos as an example, news content that continuously broadcasts similar topics will be clustered into the same type of news clips through a semantic clustering algorithm, thereby achieving the merging of clips with coherent content.

[0049] In the process of semantic clustering, the affinity clustering algorithm is specially used. This algorithm does not need to pre-set the total number of news segments, and can automatically perform dynamic clustering based on semantic similarity. This method is particularly suitable for news videos with frequent content changes and uncertain number of news segments, and has significant advantages when processing long news. After semantic clustering, the system merges the video segments corresponding to the text segments belonging to the same clustering category into a complete news segment, thereby improving the accuracy and coherence of the segment splitting.

[0050] In order to further improve the accuracy of segmentation, the system also sets a minimum frame limit. When a video segment is less than 150 frames long, the system will calculate the image similarity between the segment and the previous and next segments, and intelligently merge it with the segments with higher similarity. For example, if the background image of a certain broadcast is short, by calculating the similarity with the previous and next scenes, the system can accurately determine whether the segment should be merged with the previous or next segment, ensuring the rationality and integrity of the video segmentation.

[0051] In addition, when processing news videos with a lot of background music and noise, relying solely on speech recognition and text semantic clustering is prone to interference, resulting in problems such as incoherent or incorrect merging of segments. This method effectively avoids the interference of background noise by first using the image similarity model to cut the video and then merging the segments using semantic similarity. For example, in news involving foreign visits or street scenes, background music or noise will affect the speech recognition effect of traditional technologies. However, this method ensures accurate merging of segments and avoids the problem of semantic incoherence by dual judgment of image and semantic similarity when splitting the video.

[0052] Finally, this example demonstrates that the efficient stripping of news videos can be achieved through the combined use of image similarity and text semantics, and that high accuracy and coherence can be maintained even when no characters appear or the background noise is complex. The entire process greatly reduces the video processing time through lightweight models and parallel processing, and can complete the stripping of long news videos in a short time, greatly improving the work efficiency of news editors.

[0053] Embodiment 3: In news video processing, hosts often appear between different news clips, and traditional stripping techniques often rely on the appearance of hosts as a sign of news clip switching. However, this method has obvious limitations: when there is no host in the news broadcast, or when the host appears continuously between multiple news clips, it is difficult for the system to accurately determine the switching point, resulting in inaccurate video segmentation. At the same time, for scenes where the host changes, especially when there are multiple hosts or multiple people talking, the traditional stripping method that relies on face recognition is difficult to accurately distinguish between the news clips, resulting in incorrect cutting. In addition, when background music or semantically non-semantic parts (such as street scenes, static images of emergencies) appear in news clips, speech recognition technology is also difficult to cope with, further reducing the effect and accuracy of stripping.

[0054] The stripping technology based on image similarity and semantic clustering proposed in the present invention can effectively solve this problem. In news videos with hosts, this technology does not simply rely on the host's face recognition or voice transcription results, but accurately identifies the actual switching points between news segments through dual judgments of image similarity and semantic similarity. For example, in a news program with multiple hosts, although the hosts appear frequently in the video, the system can still accurately identify the switching points of the news segments through image similarity and text semantic analysis without being affected by the host changes. This method effectively avoids misjudgments caused by the continuous appearance of hosts and ensures the accuracy of video segment cutting.

[0055] When the host appears continuously in the same news video, or there is no obvious host switching between news segments, traditional methods tend to over-cut or merge news segments. However, this technology calculates the image similarity of consecutive frames and combines the semantic clustering results to divide the segments according to the actual semantics of the video content. Even if the host appears continuously, the real switching points of the news segments can be identified, avoiding the problem of miscutting caused by relying on face recognition or scene changes. For example, in a scene where a host reports multiple news stories continuously, the system can accurately identify the semantic differences between different news segments through semantic clustering analysis, ensuring the accuracy of news cutting without relying on changes in the host.

[0056] In addition, when the news contains a lot of background sound effects or non-semantic parts (such as background music or static images of emergencies), traditional speech recognition technology often has difficulty coping, resulting in inaccurate segmentation results. This method uses image similarity for preliminary segmentation and then uses semantic clustering to merge segments. Even if background music or non-semantic content interferes with speech recognition, it can still ensure the accuracy of video segmentation. For example, in a breaking news report, background music appears frequently and has no semantic content. The system can ignore irrelevant background sound effects through dual analysis of images and semantics, and accurately identify the news switching points.

[0057] It can be seen from this embodiment that the technology of the present invention is a significant improvement over the existing news video stripping methods that rely on host or face recognition. Even if the host appears frequently or continuously in the news, the present technology can still accurately identify the news segments through comprehensive judgment of pictures and semantics, avoiding miscutting and missing cuts. At the same time, for the semantically non-semantic parts and background sound effects in the news, the present technology can effectively avoid these interferences and ensure the accuracy and coherence of the cutting results.

[0058] In the present invention, unless otherwise clearly stipulated and limited, the terms "installation", "setting", "connection", "fixation", "screw-on" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0059] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present invention.

Claims

1. An efficient video stripping method based on image similarity and semantic clustering, characterized by: The method comprises the following steps: Step 1, extracting picture frames from the video to be processed at a fixed frame rate; Step 2: Encode the extracted picture frames using the picture coding model, calculate the similarity between adjacent frames, and merge adjacent frames whose similarity exceeds the threshold into a video clip based on a preset similarity threshold; Step 3, extracting corresponding audio content from the video clip and converting the audio into text; Step 4, encoding the text to convert the text into a semantic vector; Step 5: Based on the semantic vector results of the text, an affinity clustering algorithm is used to cluster the text to obtain clustering results of the text fragments; Step 6, according to the clustering result, merging the video segments corresponding to the consecutive text segments belonging to the same clustering category into a complete news segment; Step 7: When the length of a certain video segment is less than the preset minimum number of frames, the image similarity between the segment and the adjacent segments is calculated, and the segment is merged with the adjacent segment with higher similarity; Step 8, summarize the content of the merged news clips to complete the video stripping.

2. The method for efficiently decomposing videos based on image similarity and semantic clustering according to claim 1, characterized in that: The image similarity calculation used in step 2 determines the similarity of adjacent frames by performing a vector product calculation on the vectorized representation of each image frame, and compares it with a preset similarity threshold to decide whether to merge the frames.

3. The method for efficiently decomposing videos based on image similarity and semantic clustering according to claim 1, characterized in that: The affinity clustering algorithm used in step 5 does not require a preset number of categories, and can dynamically cluster according to text semantic similarity, and classify text segments with high similarity into the same cluster category.

4. The method for efficiently decomposing videos based on image similarity and semantic clustering according to claim 1, characterized in that: In step 7, when the number of frames of a certain video segment is lower than the preset minimum number of frames, the similarity between the segment and the last frame of the previous video segment and the first frame of the next video segment is calculated, and the video segment is merged with the segment with higher similarity.

5. The method for efficiently decomposing videos based on image similarity and semantic clustering according to claim 1, characterized in that: The preset minimum frame number is 150 frames. Video clips with a frame number below this number are regarded as too short and trigger the dynamic merging mechanism.

6. The method for efficiently decomposing videos based on image similarity and semantic clustering according to claim 1, characterized in that: The text semantic vector used in step 6 is generated by a lightweight text encoding model, which can efficiently process a large amount of text in the video and calculate the text semantic similarity in parallel.

7. The method for efficiently decomposing videos based on image similarity and semantic clustering according to claim 1, characterized in that: The steps of merging and cutting the video clips are performed by parallel processing technology, achieving efficient processing with a processing speed ratio of one tenth of the video duration.

8. The method for efficiently de-stripping videos based on image similarity and semantic clustering according to claim 1, characterized in that: The video strip splitting method further includes performing semantic summarization of the segment contents after merging the video segments, and generating corresponding summary information to help users quickly understand the main contents of the news segments.

9. The method for efficiently decomposing videos based on image similarity and semantic clustering according to claim 1, characterized in that: The audio-to-text technology in step 3 extracts the video audio at fixed intervals through the speech recognition model to ensure time alignment with the video clip.

10. The method for efficiently de-stripping videos based on image similarity and semantic clustering according to claim 1, characterized in that: The method is applicable to news videos from different sources, and can accurately split and merge video clips even when there is no host, person or background sound effect in the news.

Images