An intelligent editing style learning and shot arrangement method and device

Abstract

The application discloses an intelligent editing style learning and shot arrangement method and device, first, the original material video is shot segmented to obtain a plurality of independent video clips, each video clip is analyzed in multiple dimensions, and a structured label text containing picture semantic features and shooting syntax features is generated, then the to-be-edited manuscript is semantically segmented to obtain a plurality of text paragraphs, at least one editing example is matched for each text paragraph, finally, a prompt word is constructed and input into a generative language model to arrange shots, and the generative language model outputs a shot arrangement sequence of the current paragraph, the error accumulation problem caused by the dependence of a traditional intelligent editing method on an intermediate link of an editing script is solved, and the defect that the prior art is difficult to dynamically learn a specific editing style is solved, semantic matching and editing syntax are cooperatively optimized, and the accuracy of shot selection, arrangement fluency and style adaptability are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a smart editing style learning and shot arrangement method and device. BACKGROUND

[0002] In the field of news video editing, traditional methods highly rely on manual operations of professional editors. Editors need to manually select matching pictures from a large amount of original materials based on the content of news manuscripts, and perform accurate shot arrangement to ensure that the final film content is accurate, the narrative is smooth, and the news value is reflected. This method is tedious, time-consuming and labor-intensive, resulting in low overall editing efficiency.

[0003] To improve efficiency, early intelligent editing technology has been explored. This method usually slices the news manuscript text and performs shot retrieval based on the semantics of the text segments. However, this method has significant limitations: on the one hand, there is often a lack of strict semantic alignment between news manuscript text descriptions and required video pictures, making it difficult to accurately reflect the core content of the news. In addition, this method usually directly assembles the semantic retrieval results in sequence to form the shot arrangement results, without considering the fluency and aesthetics of the editing (i.e., editing grammar, such as scene transition, picture motion transition, camera movement transition, etc.), thereby affecting the quality of the final editing.

[0004] With the development of large language model (LLM) technology, a new generation of intelligent editing method has emerged. The core idea is to use LLM to analyze news manuscripts (or combine material information) to generate detailed editing scripts, which contain multiple shots, each containing picture descriptions, scene guidance information, and then match specific shots based on the script through cross-modal retrieval. Using scripts as an intermediary, to some extent, improves the semantic matching effect from text to shots. However, this method still has many problems, for example: first, the quality of the editing script, as a key intermediate variable, lacks a unified and reliable evaluation standard, making it difficult to quantify its impact on the final editing result, resulting in deviations from human editing intentions. Second, the method mainly focuses on semantic matching of shot content, and the processing of editing fluency and aesthetics usually relies on pre-set fixed rules, making it difficult to adaptively adjust according to the diversity of manuscript content and specific needs. Third, it is difficult to effectively learn and adapt to specific editing styles, making it difficult to meet the individualized editing needs of different media or scenarios.

[0005] In summary, the traditional manual editing method is inefficient, and the existing intelligent editing method has significant deficiencies in the accuracy of semantic matching of shot content, the flexible adaptability of editing grammar, and the alignment of personalized style. Therefore, there is an urgent need for a more direct and end-to-end intelligent editing solution that can effectively simplify operations, improve efficiency, and comprehensively consider content semantic matching and editing grammar rules to achieve more accurate shot selection and arrangement, overcoming the limitations of existing technologies in accuracy, smoothness, and personalization. SUMMARY

[0006] The purpose of the present application is to overcome the technical defects of the prior art and provide an intelligent editing style learning and shot arrangement method and device, which solves the error accumulation problem caused by the dependence on the intermediate link of the editing script in the traditional intelligent editing method, and the defect that the prior art is difficult to dynamically learn a specific editing style.

[0007] The purpose of the present application is achieved by the following technical solutions:

[0008] In a first aspect, the present application provides an intelligent editing style learning and shot arrangement method, comprising:

[0009] Performing shot segmentation on the original material video to obtain a plurality of independent video segments, performing multi-dimensional analysis on each video segment, and generating a structured label text containing picture semantic features and shooting grammar features;

[0010] Performing semantic segmentation on the to-be-edited manuscript to obtain a plurality of text paragraphs, matching at least one editing example for each text paragraph, and the editing example containing historical news paragraph text and its corresponding structured label sequence of the finished shot;

[0011] Constructing a prompt word and inputting a generative language model to perform shot arrangement, and outputting a shot arrangement sequence of the current paragraph from the generative language model, the prompt word including a text paragraph, an editing example, and a structured label text, and the shot arrangement sequence being an ordered set of shot labels selected from a label text pool.

[0012] In one possible implementation, the step of performing shot segmentation on the original material video to obtain a plurality of independent video segments comprises:

[0013] The original material video is segmented into a plurality of video segments according to a single shot by a shot segmentation algorithm: wherein represents a video shot segmentation algorithm, is a material video, is a plurality of video segments obtained after algorithm processing, is a single shot.

[0014] In one possible implementation, multidimensional analysis is... ,in This indicates the method for obtaining multi-dimensional tags. These are individual video clips obtained by segmenting the original footage. To obtain Multi-dimensional tags on;

[0015] Multi-dimensional analysis includes image content dimension and shooting information dimension. Image content dimension includes extracting image description, background scene, visual element entities, on-screen characters, and subject behavior; shooting information dimension includes extracting shot type, shooting method, camera movement mode, image brightness, and shot duration.

[0016] In one possible implementation, semantic segmentation is as follows: , This is a method for implementing semantic segmentation. This is the complete manuscript to be edited. After division One paragraph.

[0017] In one possible implementation, semantic segmentation can be achieved through a large language model: ,in Representing a large language model; This indicates the prompt word used here.

[0018] In one possible implementation, the step of matching at least one clip example for each text paragraph includes:

[0019] Building a historical database , Indicates the history database; It is an index set composed of previously edited news articles; News The manuscript, This indicates the corresponding finished video. All manuscripts and finished videos have undergone semantic segmentation, shot segmentation, and multi-dimensional tag analysis.

[0020] For each paragraph of the manuscript to be edited Calculate paragraphs using a text semantic vector encoding model Semantic similarity with all historical segments in the historical database: ,in Indicates the implementation method. All paragraphs obtained by segmenting all manuscripts in the historical database. This indicates that the paragraph is from here. Found with paragraph The paragraphs with the most similar meanings;

[0021] Select the historical paragraphs with the highest similarity As a matching result, extract and The corresponding sequence of finished shots and a set of multi-dimensional tags form an editing example.

[0022] In one possible implementation, the sequence of shots is as follows: The multi-dimensional tag set is as follows: ,in Indicates taking the result AND The corresponding part; for The lens To the lens ; ，……， For the lens To the lens Multi-dimensional tag information, This represents a multi-dimensional tag sequence in professionally edited footage.

[0023] In one possible implementation, the steps of constructing cue words and inputting them into a generative language model for shot arrangement, and having the generative language model output the shot arrangement sequence of the current paragraph, include:

[0024] Dynamically construct a pool of potential shots for the current segment to be edited. The shot pool contains a multi-dimensional tag set of all original footage shots or a subset of filtered shots:

[0025] Construct cue words, input them into a generative language model, which learns from the editing style of clip examples, directly selects and sorts the shot numbers from the pool of candidate shots, and outputs the shot arrangement sequence of the current segment. .

[0026] Secondly, this application proposes an intelligent editing style learning and shot arrangement device, the device comprising:

[0027] The material analysis module is used to segment the original video footage into multiple independent video segments. Each video segment is then analyzed in multiple dimensions to generate structured tag text containing semantic features of the scene and grammatical features of the shooting.

[0028] The manuscript analysis module is used to semantically segment the manuscript to be edited to obtain multiple text segments, and match at least one editing example for each text segment. The editing example contains a structured tag sequence of historical news segment text and its corresponding finished shot.

[0029] The shot arrangement module is used to construct cue words and input them into a generative language model for shot arrangement. The generative language model outputs the shot arrangement sequence of the current paragraph. The cue words include text paragraphs, clip examples, and structured tag text. The shot arrangement sequence is an ordered set of shot tags selected from the tag text pool.

[0030] The main solution and its various further alternatives described above can be freely combined to form multiple solutions, all of which are solutions that can be adopted and are claimed in this application; furthermore, the (non-conflicting alternatives) can also be freely combined with each other and with other alternatives. Those skilled in the art, after understanding the solution of this application, will realize from the prior art and common general knowledge that there are many combinations, all of which are technical solutions to be protected in this application, and will not be exhaustively listed here.

[0031] This application discloses an intelligent editing style learning and shot arrangement method and apparatus. First, the original video footage is segmented into multiple independent video segments. Each video segment is analyzed in multiple dimensions to generate structured tag text containing semantic features of the scene and shooting grammar features. Then, the manuscript to be edited is semantically segmented to obtain multiple text segments. At least one editing example is matched to each text segment. Finally, prompt words are constructed and input into a generative language model for shot arrangement. The generative language model outputs the shot arrangement sequence of the current segment. This solves the problem of error accumulation caused by the reliance on intermediate steps of the editing script in traditional intelligent editing methods, as well as the deficiency of existing technologies in dynamically learning specific editing styles. It achieves synergistic optimization of semantic matching and editing grammar, significantly improving the accuracy of shot selection, the smoothness of arrangement, and the style adaptability. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 The diagram shows a flowchart of an intelligent editing style learning and shot arrangement method proposed in an embodiment of this application.

[0034] Figure 2 A flowchart of the method proposed in an embodiment of this application is shown. Detailed Implementation

[0035] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0036] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] Traditional news editing methods suffer from high time and labor costs and low efficiency. Furthermore, existing intelligent editing methods have significant limitations in semantic matching of shot content and editing grammar. Therefore, there is an urgent need to find a more direct and end-to-end intelligent editing approach. This approach not only simplifies manual operations and improves efficiency but also emulates the thinking of professional human editors, comprehensively analyzing and considering both semantic matching and editing grammar. This intelligent editing method can more accurately select and arrange shots, effectively addressing the shortcomings of existing methods in terms of accuracy, fluency, and personalization, further improving the quality and efficiency of the editing results.

[0038] Please refer to Figure 1 , Figure 1 This paper illustrates a flowchart of an intelligent editing style learning and shot arrangement method proposed in an embodiment of this application. The method includes:

[0039] Step S1: Segment the original video footage to obtain multiple independent video segments. Perform multi-dimensional analysis on each video segment to generate structured tag text containing semantic features of the scene and grammatical features of the shooting.

[0040] First, the original video footage is segmented into individual video clips. Then, each clip is analyzed in depth from various dimensions, such as content, style, and object recognition. Finally, based on these analysis results, multi-dimensional tagged text information is generated. These tags enable quick location and retrieval of specific content, enriching the video footage with deeper semantic and structural information.

[0041] Step S1, which involves segmenting the original video footage into multiple independent video segments, includes:

[0042] The original video footage is divided into multiple video segments based on individual shots using a shot segmentation algorithm. ,in This represents a video shot segmentation algorithm. For a source video, The result after algorithm processing Each video clip consists of a single shot.

[0043] Each original video clip can be segmented using a shot segmentation algorithm, dividing the video into multiple independent video segments, which can be represented as: ,in This represents a video shot segmentation algorithm. For a source video, The result after algorithm processing Each video clip is a single shot. If the original footage already consists of a single, independent shot, this process can be omitted.

[0044] Multidimensional analysis for ,in This indicates the method for obtaining multi-dimensional tags. These are individual video clips obtained by segmenting the original footage. To obtain Multi-dimensional tags on;

[0045] Multi-dimensional analysis includes image content dimension and shooting information dimension. Image content dimension includes extracting image description, background scene, visual elements, on-screen characters, and subject behavior; shooting information dimension includes extracting shot type, shooting method, camera movement mode, image brightness, and shot duration. Table 1 shows the multi-dimensional analysis:

[0046] Table 1

[0047]

[0048] There are no restrictions on how tags are acquired; they can be obtained through large visual models, multimodal large models, or other algorithms. The three mentioned above are just some examples. The tag system mainly includes the dimensions of "image content" and "shooting information," covering image semantics and image grammar. These tags are represented in text form, converting visual information in the video into text data that can be processed and understood by large language models, providing an important data foundation for subsequent end-to-end intelligent editing. Each dimension includes, but is not limited to, the tags mentioned above. This process can be represented as: ,in This indicates the method for obtaining multi-dimensional labels, which can include large visual models, multimodal large models, multiple algorithms, etc. These are individual video clips obtained by segmenting the original material described above. They can be in video format or composed of... The resulting image sequence; To obtain Multi-dimensional tags on.

[0049] Step S2: Semantically segment the manuscript to be edited to obtain multiple text segments, and match at least one editing example for each text segment.

[0050] The clip examples contain structured tag sequences of historical news clips and their corresponding finished shots. The manuscript is semantically segmented to generate multiple text segments for precise matching with video footage. Then, at least one clip example is matched to each text segment. Each example contains a multi-dimensional tag sequence of historical news clips and their corresponding finished shots. Through matching, a semantic connection is established between the text and the video, and the multi-dimensional tags enhance the intelligence of video editing.

[0051] Semantic segmentation is as follows: , This is a method for implementing semantic segmentation. This is the complete manuscript to be edited. After division One paragraph.

[0052] For manuscripts, they need to be divided into multiple paragraphs according to semantics. If the manuscript has already been divided into paragraphs, this step can be omitted. This process can be represented as:

[0053] in, This indicates the implementation method of semantic segmentation here; The complete manuscript to be edited; After division One paragraph.

[0054] The text paragraphs were also obtained through a large language model: ,in Representing a large language model; This indicates the prompt word used here.

[0055] The large language model divides the text into multiple paragraphs based on semantics, describes the semantic segmentation task in prompt words, and provides the complete text in the prompt words. The large language model then outputs the segmented paragraphs. This process can be represented as:

[0056] in, Representing a large language model; This indicates the prompt word used here. Note that there are no restrictions on the prompt word; it only needs to fulfill the task.

[0057] The clip example refers to a paragraph from a news article, and the finished clip edited from that paragraph by a professional editor. The finished clip then needs to be further converted into multi-dimensional tagged text information for each shot, so that it can be used and understood by the large language model later. It's worth noting that the matching implementation method is not limited and various methods can be used.

[0058] The steps for matching at least one clip example for each text paragraph include:

[0059] Building a historical database , Indicates the history database; It is an index set composed of previously edited news articles; News The manuscript, This indicates the corresponding finished video. All manuscripts and finished videos have undergone semantic segmentation, shot segmentation, and multi-dimensional tag analysis.

[0060] For each paragraph of the manuscript to be edited Calculate paragraphs using a text semantic vector encoding model Semantic similarity with all historical segments in the historical database: ,in Indicates the implementation method. All paragraphs obtained by segmenting all manuscripts in the historical database. This indicates that the paragraph is from here. Found with paragraph The paragraphs with the most similar meanings;

[0061] Select the historical paragraphs with the highest similarity As a matching result, extract and The corresponding sequence of finished shots and a set of multi-dimensional tags form an editing example.

[0062] The sequence of shots in the final cut is as follows: The multi-dimensional tag set is as follows:

[0063]

[0064] in Indicates taking the result AND The corresponding part; for The lens To the lens ; ，……， For the lens To the lens Multi-dimensional tag information, This represents a multi-dimensional tag sequence in professionally edited footage.

[0065] Television stations or other media organizations typically maintain a historical database that stores previously edited news articles (drafts and corresponding finished products). ,in Indicates the history database; It is an index set composed of previously edited news articles; News The manuscript, This indicates the corresponding finished video.

[0066] All news articles (and corresponding footage) in the historical database should have undergone the same processing as described in step one. That is:

[0067] ;

[0068] ;

[0069] ;

[0070] in, For the historical database of complete videos After the shot split A single shot clip; For the lens Multi-dimensional tags; For the manuscript After semantic segmentation One paragraph.

[0071] For each paragraph of the manuscript to be edited obtained in step one Taking matching a clip example, we can find the most semantically similar paragraph from the history database, as well as the corresponding full-length clip. This process can be represented as:

[0072] ;

[0073] in, This indicates the implementation method here. All paragraphs obtained by segmenting all manuscripts in the historical database. This indicates that from Found with The paragraphs with the most similar meanings are news articles. Paragraphs of the manuscript .

[0074] One approach is through text semantic encoding: A text semantic vector encoding model is used to convert the text of a paragraph into a vector form based on its semantics. This is done for all manuscripts in the historical database and the manuscript to be edited. Then, for each paragraph in the manuscript to be edited, its vector encoding is compared pairwise with the vector encodings of all paragraphs in the historical database. The paragraph with the highest similarity is the target paragraph. This process can be represented as follows:

[0075] ;

[0076] ;

[0077] in, This represents a text semantic vector encoding model; Paragraph Vector encoding. This indicates that each paragraph of the manuscript to be edited will be... All with Each of them Calculate similarity. This represents the similarity calculation function. There are no restrictions here, and there can be multiple calculation methods, such as cosine similarity.

[0078] exist Take the maximum value from the middle, corresponding to That is the goal If matching multiple clip examples, such as two, then here for... Sort by value and take the two largest.

[0079] Depend on You can find the complete finished product corresponding to manuscript x. , Tag analysis for each shot has already been performed in the aforementioned processing; it is still necessary to find the tags that match the target segment. Corresponding finished film From the sequence of shots, extract structured, multi-dimensional label data for these shot sequences. For example, suppose... Divided into One shot, one of the shots To the lens Corresponding paragraphs So let's take a shot. To the lens Multi-dimensional tag information. This process can be represented as:

[0080] ;

[0081] ;

[0082] in, Indicates taking the result AND The corresponding part; for The lens To the lens ; , ..., For the lens To the lens Multi-dimensional tag information, This represents a multi-dimensional tag sequence in professionally edited footage.

[0083] Finally, the paragraph Professionally edited finished film with multi-dimensional tag sequences Use this as an example clip.

[0084] Step S3: Construct prompt words and input them into the generative language model for shot arrangement. The generative language model outputs the shot arrangement sequence of the current paragraph.

[0085] The cue words include text paragraphs, clip examples, and structured tag text. The shot arrangement sequence is an ordered set of shot tags selected from the tag text pool.

[0086] A cue word set containing text paragraphs, clip examples, and structured tag text is constructed and input into a generative language model. This model learns from the clip examples to understand the shot arrangement task, and then selects appropriate shot tags from a tag text pool based on the needs of the text paragraphs, generating an ordered sequence of shots. This sequence is a list of shots selected and arranged for the current script paragraph based on the model's learning from the examples, aiming to assist video editing work and improve editing efficiency.

[0087] The steps of constructing cue words and inputting them into a generative language model for shot arrangement, and then having the generative language model output the shot arrangement sequence for the current segment, include:

[0088] Dynamically construct a pool of potential shots for the current segment to be edited. The shot pool contains a multi-dimensional tag set of all original footage shots or a subset of filtered shots:

[0089] Construct cue words, input them into a generative language model, which learns from the editing style of clip examples, directly selects and sorts the shot numbers from the pool of candidate shots, and outputs the shot arrangement sequence of the current segment. .

[0090] The large language model is used to perform shot arrangement tasks based on the results of steps one and two. Using a text segment from the script to be edited as a unit, cue words are constructed, explaining the editing style learning and shot arrangement tasks. The cue words also provide an editing example, the current segment of the script to be edited, and structured text information of the video footage. After learning from the editing example, the large language model selects and arranges a shot list for the current segment.

[0091] Specifically, taking a paragraph as a unit, the multi-dimensional tags of several shots obtained from the analysis of the original material in step one need to be used to construct a pool of candidate shots. Shots entering the pool can be all shots or a subset of all shots. Taking shots a, b, c, d, f, g, and h as an example, then...

[0092] ;

[0093] in, ~ These represent the multi-dimensional labels for lenses a through h. This indicates the pool of lenses to be selected.

[0094] Next, the cue words are constructed. The cue words explain the editing style learning and shot arrangement tasks. The editing examples obtained in step two (if multiple examples exist, list them sequentially) are then provided in the cue words, along with the text of the segment to be edited and the pool of potential shots. The specific text content of the cue words is not strictly limited. Next, the cue words are input into a large language model, which learns the editing style of the editing examples and selects and arranges the shot sequence for the segment to be edited from the pool of potential shots. The final output is the shot arrangement result. This process can be represented as:

[0095] ;

[0096] Among them, for the segments to be edited , This indicates the result obtained in step two. Clip examples; The shot arrangement results output for the large model; This indicates the prompt word used here.

[0097] Continue with For example, the shot arrangement result output by the large model means a shot sequence, such as "[d, h, f, b, e, g]", which means that the large model has selected six shots for this segment: shot d, shot h, shot f, shot b, shot e, and shot g. The shot arrangement order is "shot d → shot h → shot f → shot b → shot e → shot g".

[0098] The intelligent editing result is achieved by converting the shot arrangement output from the large model into a corresponding shot video sequence for display. Once all segments of the manuscript to be edited have been processed, the intelligent editing of the entire manuscript is complete.

[0099] In one possible embodiment, Figure 2 The flowchart of the method proposed in this application is shown. First, the system segments the original video footage into multiple independent shots and performs multi-dimensional analysis on each shot to extract semantic and structural information, forming multi-dimensional tag text. Simultaneously, the script to be edited is semantically segmented into multiple text paragraphs. Next, the system matches editing examples to each paragraph from a historical database, including historical news paragraphs and their corresponding complete edits. Finally, the system constructs prompt words containing paragraph text, editing examples, and tag text, inputs them into a large language model, and the model learns editing rules based on these prompts to select and arrange a list of shots for the current paragraph, generating the final shot arrangement result. The entire process achieves deep integration of text and video, and efficiently completes video editing tasks by leveraging the generation capabilities of a large language model.

[0100] Compared with the prior art, the embodiments of this application have the following beneficial effects:

[0101] First, by generating shot arrangement results directly from a large model, end-to-end intelligent editing is achieved, eliminating the intermediate step of traditional editing scripts. This not only simplifies the editing process but also effectively solves the problem of difficulty in evaluating script quality, reduces error propagation, and significantly improves editing efficiency and the reliability of the final result.

[0102] Secondly, by utilizing multi-dimensional analysis of source material information, comprehensively covering multiple dimensions such as image content and shooting information, the large model can fully integrate semantics and syntax when arranging shots. This comprehensive consideration results in more precise shot selection, smoother and more natural choreography, and effectively improves the intelligence level of video editing and the quality of the final product.

[0103] Third, by incorporating editing examples into the model's prompts, the large model can learn and adapt to specific editing styles. In actual editing tasks, the model can flexibly adjust shot selection and arrangement based on the learned styles, meeting users' needs for personalized editing and greatly enhancing the flexibility and diversity of video content creation.

[0104] The following provides a possible implementation of an intelligent editing style learning and shot arrangement device, which is used to perform the various execution steps and corresponding technical effects of the intelligent editing style learning and shot arrangement method shown in the above embodiments and possible implementations.

[0105] The device includes:

[0106] The material analysis module is used to segment the original video footage into multiple independent video segments. Each video segment is then analyzed in multiple dimensions to generate structured tag text containing semantic features of the scene and grammatical features of the shooting.

[0107] The manuscript segmentation module is used to semantically segment the manuscript to be edited to obtain multiple text segments, and to match at least one editing example for each text segment. The editing example contains a structured tag sequence of historical news segment text and its corresponding finished shot.

[0108] The shot arrangement module is used to construct cue words and input them into a generative language model for shot arrangement. The generative language model outputs the shot arrangement sequence of the current paragraph. The cue words include text paragraphs, clip examples, and structured tag text. The shot arrangement sequence is an ordered set of shot tags selected from the tag text pool.

[0109] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for intelligent editing style learning and shot arrangement, characterized in that, include: The original video footage is divided into multiple video segments based on individual shots using a shot segmentation algorithm. ,in This represents a video shot segmentation algorithm. For a source video, The algorithm generates n video clips, each a single shot. Each clip undergoes multi-dimensional analysis to generate structured tag text containing semantic features and shooting grammar features. The multi-dimensional analysis is... ,in This indicates the method for obtaining multi-dimensional tags. These are individual video clips obtained by segmenting the original footage. To obtain Multi-dimensional tags on; Multi-dimensional analysis includes two dimensions: image content and shooting information. Image content includes extracting image descriptions, background scenes, visual elements, on-screen characters, and main actions. Shooting information includes extracting shot type, shooting method, camera movement mode, image brightness, and shot duration. The manuscript to be edited is semantically segmented to obtain multiple text paragraphs. Each text paragraph is matched with at least one editing example. These editing examples contain structured tag sequences of historical news paragraphs and their corresponding finished shots. The semantic segmentation is as follows: , This is a method for implementing semantic segmentation. This is the complete manuscript to be edited. After division One paragraph; The steps for matching at least one clip example for each text paragraph include: Building a historical database , Indicates the history database; It is an index set composed of previously edited news articles; News The manuscript, This indicates the corresponding finished video. All manuscripts and finished videos have undergone semantic segmentation, shot segmentation, and multi-dimensional tag analysis. For each paragraph of the manuscript to be edited Calculate paragraphs using a text semantic vector encoding model Semantic similarity with all historical segments in the historical database: ,in Indicates the implementation method. All paragraphs obtained by segmenting all manuscripts in the historical database. This indicates that the paragraph is from here. Found with paragraph The paragraphs with the most similar meanings; Select the historical paragraphs with the highest similarity As a matching result, extract and The corresponding sequence of finished shots and a set of multi-dimensional tags form an editing example; a pool of candidate shots is dynamically constructed for the current segment to be edited, and the shot pool contains a set of multi-dimensional tags for all original footage shots or a subset of selected shots. Construct cue words and input them into a generative language model. The model learns the editing style based on the clip examples and directly selects and sorts the shot numbers from the pool of candidate shots. It then outputs the shot arrangement sequence of the current segment. The cue words include text paragraphs, clip examples, and structured label text. The shot arrangement sequence is an ordered set of shot labels selected from the label text pool.

2. The intelligent editing style learning and shot arrangement method as described in claim 1, characterized in that, Semantic segmentation is achieved through a large language model: ,in Representing a large language model; This indicates the prompt word used here.

3. The intelligent editing style learning and shot arrangement method as described in claim 1, characterized in that, The sequence of shots in the final cut is as follows: The multi-dimensional tag set is as follows: ； in Indicates taking the result AND The corresponding part, for The lens To the lens , For the lens To the lens Multi-dimensional tag information, This represents a multi-dimensional tag sequence in professionally edited footage.

4. A device for intelligent editing style learning and shot arrangement, characterized in that, The device includes: The material analysis module is used to segment the original video footage into multiple video segments based on individual shots using a shot segmentation algorithm. ,in This represents a video shot segmentation algorithm. For a source video, The algorithm generates n video clips, each a single shot. Each clip undergoes multi-dimensional analysis to generate structured tag text containing semantic features and shooting grammar features. The multi-dimensional analysis is... ,in This indicates the method for obtaining multi-dimensional tags. These are individual video clips obtained by segmenting the original footage. To obtain Multi-dimensional tags on; Multi-dimensional analysis includes image content dimension and shooting information dimension. Image content dimension includes extracting image description, background scene, visual element entities, on-screen characters, and subject behavior; shooting information dimension includes extracting shot type, shooting method, camera movement mode, image brightness, and shot duration. The manuscript analysis module is used to semantically segment the manuscript to be edited, resulting in multiple text paragraphs. For each text paragraph, at least one editing example is matched. The editing example contains structured tag sequences of historical news paragraphs and their corresponding finished shots. The semantic segmentation is as follows: , This is a method for implementing semantic segmentation. This is the complete manuscript to be edited. After division One paragraph; Building a historical database , Indicates the history database; It is an index set composed of previously edited news articles; News The manuscript, This indicates the corresponding finished video. All manuscripts and finished videos have undergone semantic segmentation, shot segmentation, and multi-dimensional tag analysis. For each paragraph of the manuscript to be edited Calculate paragraphs using a text semantic vector encoding model Semantic similarity with all historical segments in the historical database: ,in Indicates the implementation method. All paragraphs obtained by segmenting all manuscripts in the historical database. This indicates that the paragraph is from here. Found with paragraph The paragraphs with the most similar meanings; Select the historical paragraphs with the highest similarity As a matching result, extract and The corresponding sequence of finished shots and a set of multi-dimensional tags form an editing example; The shot arrangement module is used to dynamically construct a pool of potential shots for the current segment to be edited. The shot pool contains a multi-dimensional tag set of all original footage shots or a subset of selected shots: Construct cue words and input them into a generative language model. The model learns the editing style based on the clip examples and directly selects and sorts the shot numbers from the pool of candidate shots. It then outputs the shot arrangement sequence of the current segment. The cue words include text paragraphs, clip examples, and structured label text. The shot arrangement sequence is an ordered set of shot labels selected from the label text pool.

Images