A judgment method and system based on AI automatic editing of video materials

Abstract

The present invention discloses a judgment method and system for automatically editing video materials based on AI, which relates to the field of artificial intelligence technology. The judgment method includes the following steps: obtaining multimodal video data; generating a collection of abnormal editing segments; evaluating the degree of influence of the collection of abnormal editing segments on the quality of finished editing; judging whether the collection of abnormal editing segments meets the video editing standard, screening out materials that meet the video editing standard, and generating a high-quality material library for editing for storage. By evaluating the influence of abnormal editing segments on the quality of finished editing, the present invention can accurately control the quality of the material during the editing process, avoid the influence of bad material on the final effect, and automatically screen materials that meet the video editing standard based on the evaluation results, generate a high-quality material library for storage, thereby avoiding the subjectivity of manual selection of materials, ensuring the uniformity and quality of the materials, and providing a high-standard material library for subsequent editing.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for determining automatic editing of video materials based on AI. Background Art

[0002] Video editing is the nonlinear editing process of video sources and is part of multimedia production software. Using video editing software, users can perform a variety of complex operations on videos, including cutting, merging, adjusting speed, adding special effects, color grading, and dubbing, to achieve the desired visual effect. The editing process involves more than simple deletion and splicing; it requires in-depth processing of the video footage, removing unnecessary sections, and reorganizing and recreating it to suit creative needs. After video capture, it often contains a significant amount of "waste" footage. This waste can be caused by shooting errors, duplicate shots, blank frames, or irrelevant content. The presence of waste footage significantly impacts the smoothness and expressiveness of the video, so the primary task of editing software is to filter and delete these waste clips. Waste clips include, but are not limited to, redundant shots, incorrectly shot content, and meaningless blank frames, all of which need to be removed during the initial editing phase.

[0003] However, in the existing technology, it is not convenient to quickly screen out invalid clips and abnormal materials in video materials, which increases the time and manpower cost of video editing. It is also not convenient to compare the generated abnormal editing clips with the high-quality video material standards, and the quality of the final product cannot be ensured. As a result, the quality of the material cannot be accurately controlled during the editing process to avoid the impact of bad materials on the final effect. At the same time, it is not convenient to automatically screen materials that meet the video editing standards, which reduces the efficiency and quality of video editing and increases the difficulty of manual operation.

[0004] Currently, no effective solutions have been proposed for the problems in related technologies. Summary of the Invention

[0005] In response to the deficiencies in the prior art, the present invention proposes a judgment method and system for automatic editing of video materials based on AI, which solves the problem raised in the above background technology that it is inconvenient to quickly screen out invalid clips and abnormal materials in video materials, increases the time and manpower cost of video editing, and is inconvenient to compare the generated abnormal editing clips with the high-quality video material standards, and cannot ensure the quality of the final product, so that the quality of the material cannot be accurately controlled during the editing process, and poor material is prevented from affecting the final effect. At the same time, it is not convenient to automatically screen materials that meet the video editing standards, thereby reducing the efficiency and quality of video editing and increasing the difficulty of manual operation.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0007] According to one aspect of the present invention, a method for determining whether video material can be automatically edited based on AI is provided. The method comprises the following steps:

[0008] S1. Obtain video material data to be edited, and pre-process the video material data to obtain multimodal video data;

[0009] S2. Analyze the obtained multimodal video data based on an anomaly recognition algorithm, mark abnormal material data, and generate a set of abnormal clips;

[0010] S3. Compare the generated set of abnormal editing segments with known high-quality video material standards to evaluate the degree of influence of the abnormal editing segments on the quality of the finished editing;

[0011] S4. Based on the evaluation results, determine whether the set of abnormal clips meets the video editing standards, screen out materials that meet the video editing standards, and generate a high-quality material library for editing and store it;

[0012] Based on the anomaly recognition algorithm, the obtained multimodal video data is analyzed and the abnormal material data is marked. The generation of a set of abnormal clips includes the following steps:

[0013] S21, quantizing the obtained multimodal video data, dividing the quantized multimodal video data into a training set and a test set, and setting a maximum number of iterations and a size of a candidate parameter set for the anomaly recognition algorithm;

[0014] S22. Generate an initial parameter combination for the anomaly recognition algorithm using a random initialization strategy combined with the distribution characteristics of the feature space, and configure key parameters of the anomaly detection model;

[0015] S23. Dynamically adjust the weight of the objective function of the anomaly detection model based on the changing pattern of the abnormal features of the video clips, combine the temporal characteristics of the abnormal clips, and set the feature priority;

[0016] S24. Using the anomaly detection model objective function, evaluate the recognition accuracy of the anomaly detection model for video data anomalies under each parameter combination to obtain a fitness value, and sort the fitness values ​​of all parameter combinations to select the currently optimal anomaly detection model parameters;

[0017] S25. Based on the distribution of abnormal features, calculate the nonlinear deviation of the anomaly detection model. If the nonlinear deviation is less than the threshold, adopt a local optimization strategy to adjust the parameter combination of the anomaly detection model. Otherwise, adopt a full local optimization strategy to adjust the direction of all parameter combinations of the anomaly detection model and adjust the weights of the auxiliary features.

[0018] S26. Based on the parameter selection algorithm, compare the optimization result of the current anomaly detection model with the known optimal parameter combination, and select the anomaly detection model parameters with the best performance;

[0019] S27. Randomly perturb the current optimal anomaly detection model parameters to generate new candidate anomaly detection model parameter combinations, and optimize the perturbed anomaly detection model parameters using a model optimization algorithm to obtain the local optimal anomaly detection model parameters in the perturbed area.

[0020] S28. If the local optimal anomaly detection model parameters are better than the current optimal anomaly detection model parameters, then record the current optimal anomaly detection model and update the anomaly segment set; otherwise, return to step S22 to continue optimization;

[0021] S29: Output the optimal anomaly detection model parameters and the anomaly labeling results of the anomaly detection model, and generate a set of clip anomaly segments based on the anomaly labeling results of the anomaly detection model.

[0022] Furthermore, based on the parameter selection algorithm, the optimization results of the current anomaly detection model are compared with the known optimal parameter combination, and the selection of the anomaly detection model parameters with the best performance includes the following steps:

[0023] S261. Define the key parameter search range and feature selection range of the anomaly detection model;

[0024] S262. Randomly generate several key parameter configurations and feature subsets, each combination of a key parameter configuration and a feature subset represents a candidate parameter configuration, and all combinations constitute a candidate parameter configuration group;

[0025] S263. Train the anomaly detection model using each candidate parameter configuration, calculate its fitness value, and record the performance results of the current candidate parameter configuration;

[0026] S264: Perform a neighborhood search near the key parameter configuration with the highest fitness in the candidate parameter configuration set, and update and record the key parameter configuration and feature subset with the highest fitness;

[0027] S265: Distribute the remaining candidate parameter configurations to random key parameter configurations and feature subsets in the global range, calculate their fitness values, and update the global optimal key parameter configuration;

[0028] S266. Repeat steps S263 to S265 until the maximum number of iterations is reached, and compare the current optimal key parameter configuration with the known optimal parameter configuration. If the current optimal key parameter configuration is better than the known optimal parameter configuration, replace it with the new optimal parameter configuration as the anomaly detection model parameter with the best performance.

[0029] Furthermore, performing a neighborhood search near the key parameter configuration with the highest fitness in the candidate parameter configuration set, and updating and recording the key parameter configuration and feature subset with the highest fitness include the following steps:

[0030] S2641. Identify the key parameter configuration with the highest fitness value in the current candidate parameter configuration set as the center of the neighborhood search;

[0031] S2642. Define a neighborhood range according to preset rules, set a perturbation amplitude for each dimension of the key parameter with the current optimal parameter configuration as the center, and calculate a neighborhood separately for each key parameter configuration to form a multi-dimensional search space;

[0032] S2643. Randomly generate several new key parameter configurations within the neighborhood, retrain the anomaly detection model using the new parameter configurations, and calculate the fitness value of each new key parameter configuration;

[0033] S2644. Compare the fitness values ​​of all new key parameter configurations within the domain with the fitness value of the current optimal parameter configuration. If the fitness value of the new key parameter configuration is greater than the fitness value of the current optimal parameter configuration, the new key parameter configuration is used as the key parameter configuration with the highest fitness. At the same time, the feature subset corresponding to the key parameter configuration with the highest fitness is updated and recorded. Otherwise, the current optimal parameter configuration is retained unchanged.

[0034] Furthermore, the current optimal anomaly detection model parameters are randomly perturbed to generate new candidate anomaly detection model parameter combinations. The perturbed anomaly detection model parameters are optimized using a model optimization algorithm. Obtaining the local optimal anomaly detection model parameters in the perturbed area includes the following steps:

[0035] S271. Initialize the parameters of the model optimization algorithm and randomly select a number of candidate anomaly detection model parameter configurations within the disturbance region;

[0036] S272: Input the candidate anomaly detection model parameter configuration into the anomaly detection model, and calculate the fitness value of each candidate anomaly detection model parameter configuration through the objective function;

[0037] S273. Select the anomaly detection model parameter configuration with the highest fitness value from all candidate anomaly detection model parameter configurations as the current optimal anomaly detection model parameter solution;

[0038] S274. Update the position of each anomaly detection model parameter configuration according to the position update formula;

[0039] S275. If the fitness value of any candidate parameter is greater than the fitness value of the current optimal anomaly detection model parameter, then update the optimal anomaly detection model parameter solution, remove the anomaly detection model parameter configuration with the lowest fitness value from the current candidate anomaly detection model parameter set, and randomly generate a new anomaly detection model parameter configuration within the disturbance region of the current optimal anomaly detection model parameter.

[0040] S276. Check whether the maximum number of iterations is reached. If so, output the current optimal anomaly detection model parameter configuration as the local optimal anomaly detection model parameter of the disturbance area.

[0041] Furthermore, the position update formula is:

[0042] ;

[0043] Where, Y i ( s +1) indicates the i The anomaly detection model parameters are configured in s +1 position at iteration time; rand Represents a random number; Y Pc represents the location of the optimal anomaly detection model parameter scheme within the disturbance region; PC Represents the optimal anomaly detection model parameter scheme.

[0044] Furthermore, the generated set of abnormal editing segments is compared with known high-quality video material standards to evaluate the impact of the abnormal editing segments on the quality of the finished editing, including the following steps:

[0045] S31, importing the generated abnormal editing segment set and known high-quality video material standards, setting evaluation parameters, and pre-processing the abnormal editing segment set and the high-quality video material;

[0046] S32. Screening out the editing features that have the greatest impact on the quality of the finished editing through a feature screening algorithm;

[0047] S33, dividing the abnormal clip set into a training set and a test set;

[0048] S34, randomly initializing the parameter configuration of the clip evaluation model, defining the parameter search space as multi-dimensional, and using each set of parameter configurations as a combination of evaluation criteria;

[0049] S35, using a cross-validation method to evaluate the fitness value of each parameter configuration in the training set, recording the fitness value of the current parameter configuration, and marking the optimal fitness value among all parameter configurations;

[0050] S36. Using a heuristic algorithm to adjust the evaluation model parameter configuration to generate a new parameter configuration;

[0051] S37, using the new parameter configuration to evaluate the training set, calculate the fitness value of each abnormal clip set under the new parameter configuration; and update the optimal fitness value and its corresponding parameter configuration among all parameter configurations;

[0052] S38, check whether the maximum number of iterations has been reached, if so, proceed to step S39, otherwise return to step S36, continue to adjust the evaluation model parameter configuration; output the evaluation model parameter configuration with the highest fitness value;

[0053] S39. Use the optimal parameter configuration to evaluate the abnormal editing segments in the test set, quantify the matching degree between the abnormal editing segments and the high-quality standard, calculate the impact score of the editing abnormality segment set on the quality of the finished editing, and output the evaluation results.

[0054] Furthermore, the following steps are involved in selecting the editing features that have the greatest impact on the quality of the finished editing through a feature selection algorithm:

[0055] S321, initializing all parameters in the feature screening algorithm;

[0056] S322, randomly select a feature combination as an initial feature candidate set, and initialize the feature weights using chaos thinking;

[0057] S323, evaluating the impact of each candidate feature combination on the editing quality, calculating the fitness value of the candidate feature combination, and finding the feature combination with the best performance at present;

[0058] S324: Update the weight of the current feature combination to generate a new candidate feature combination;

[0059] S325. Check whether the updated feature combination meets the feature quantity threshold. If so, update the feature combination according to the update rule. Otherwise, keep the current feature combination unchanged.

[0060] S326, recalculate the fitness value of the updated feature combination, and compare it with the previous feature combination to find the feature combination with the highest fitness value;

[0061] S327, comparing the fitness values ​​of the current feature combination and the historical optimal feature combination, and recording the feature combination with the highest fitness as the new optimal feature combination;

[0062] S328. Check whether the maximum number of iterations has been reached. If so, output the optimal feature combination as the editing feature that has the greatest impact on the quality of the finished editing. Otherwise, return to step S324 and continue iterative optimization.

[0063] Furthermore, the heuristic algorithm is used to adjust the evaluation model parameter configuration. Generating a new parameter configuration includes the following steps:

[0064] S361. Set the parameters of the heuristic algorithm and randomly generate a parameter configuration population matrix, where each row represents a candidate parameter configuration and each column represents an evaluation model parameter;

[0065] S362, sequentially evaluating the fitness value of each parameter configuration to check whether the parameter configuration exceeds the upper and lower limits. If so, adjusting the parameter configuration to the boundary value; otherwise, retaining the current parameter configuration and calculating the fitness value of each parameter configuration using the fitness function;

[0066] S363. During the initial iteration, the parameter configurations are sorted from small to large according to the fitness values. The sorted parameter configuration population is the optimal parameter configuration population. The first parameter configuration in the optimal parameter configuration population is selected as the optimal parameter configuration. During subsequent iterations, the current optimal parameter configuration population is merged with the previous optimal parameter configuration population, and the population is re-sorted according to the fitness values ​​to select a new optimal parameter configuration population. The first parameter configuration in the new optimal parameter configuration population is selected as the optimal parameter configuration.

[0067] S364. Update the parameter configurations in the parameter configuration population one by one according to the update rule of the heuristic algorithm;

[0068] S365. Check whether the maximum number of iterations is reached. If so, output the optimal parameter configuration and the corresponding fitness value as the new parameter configuration. Otherwise, return to step S362 and continue executing until the maximum number of iterations is reached.

[0069] Furthermore, the formula for updating the rule is:

[0070] ;

[0071] Where, A e Indicates the e Parameter configuration; M n Indicates the n Target parameter configuration; G e Indicates the e Parameter configuration and n The difference between the target parameter configurations; b represents the constant that defines the shape of the logarithmic spiral; t Represents a random number; B represents the spiral function.

[0072] According to another aspect of the present invention, a judgment system for automatically editing video materials based on AI is also provided, and the judgment system includes:

[0073] The data acquisition module is used to obtain the video material data to be edited and pre-process the video material data to obtain multimodal video data;

[0074] Anomaly recognition module, which is used to analyze the obtained multimodal video data based on the anomaly recognition algorithm, mark the abnormal material data, and generate a set of abnormal clips;

[0075] A quality comparison module is used to compare the generated set of abnormal editing clips with known high-quality video material standards to evaluate the impact of the abnormal editing clips on the quality of the finished editing;

[0076] An evaluation and judgment module is used to determine whether the set of abnormal editing clips meets the video editing standards based on the evaluation results, screen out materials that meet the video editing standards, and generate a high-quality material library for editing for storage;

[0077] Among them, the data acquisition module is connected through the abnormality recognition module and the quality comparison module, and the quality comparison module is connected with the evaluation and judgment module.

[0078] The beneficial effects of the present invention are:

[0079] 1. The present invention uses automated multimodal video data processing and anomaly recognition algorithms to quickly screen out invalid clips and abnormal materials in video materials, greatly saving the time and labor costs of video editing. The generated abnormal editing clips are compared with the high-quality video material standards, which helps to ensure the quality of the final product. By evaluating the impact of abnormal editing clips on the editing quality of the finished product, the quality of the material can be accurately controlled during the editing process to avoid the impact of poor material on the final effect. Based on the evaluation results, the material that meets the video editing standards is automatically screened and a high-quality material library is generated for storage, thereby avoiding the subjectivity of manual material selection, ensuring the uniformity and quality of the material, and providing a high-standard material library for subsequent editing.

[0080] 2. The present invention uses an anomaly recognition algorithm to perform a detailed analysis of video data, and adjusts the parameter combination of the anomaly detection model through multiple rounds of optimization, so that it can accurately identify abnormal segments in the video material. By marking and eliminating these parts, the quality of the video material is ensured. By using a random initialization strategy and multiple iterative optimizations, the optimal anomaly detection model parameters are generated, making the screening of abnormal materials in the editing process more efficient and accurate. The final generated set of editing abnormal segments can automatically screen out irrelevant or low-quality materials, thereby avoiding omissions during manual inspection and reducing manual intervention.

[0081] 3. The present invention uses a feature screening algorithm to screen out the editing features that have the greatest impact on the quality of the finished editing, which can ensure that the most critical quality factors are focused on during the editing process. This not only improves the accuracy of the editing process, but also makes the evaluation results more targeted, thereby ensuring that the quality of the finished video meets the expected standards. By comparing with known high-quality video material standards, it can ensure that the video editing results meet industry standards and ensure the quality consistency of the finished video. This not only improves the professionalism of the editing, but also avoids artificial quality differences and inconsistencies, thereby ensuring the reliability of the video product. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. 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 any creative work.

[0083] Figure 1 is a flowchart of a method for determining automatic editing of video materials based on AI according to an embodiment of the present invention;

[0084] Figure 2 This is a principle block diagram of a judgment system based on AI automatic editing of video materials according to an embodiment of the present invention.

[0085] In the picture:

[0086] 1. Data acquisition module; 2. Abnormal identification module; 3. Quality comparison module; 4. Evaluation and judgment module. DETAILED DESCRIPTION

[0087] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0088] In the description of the present invention, unless otherwise specified, "plurality" means two or more. In addition, the terms "first", "second", "third", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0089] According to an embodiment of the present invention, a method and system for determining automatic editing of video materials based on AI are provided.

[0090] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to the method for determining whether to automatically edit video materials based on AI according to an embodiment of the present invention, the method includes the following steps:

[0091] S1. Obtain video material data to be edited, and pre-process the video material data to obtain multimodal video data;

[0092] Specifically, the video material data to be edited refers to the original video clips obtained from the video source or video library, including:

[0093] Video frame data: Each frame of video material is usually stored in an image format (such as JPEG, PNG, etc.) and contains the visual information of the video.

[0094] Audio data: The audio tracks in the video (background music, dialogue, sound effects, etc.), stored in audio formats (such as WAV, MP3, AAC, etc.).

[0095] Subtitles / text data: Subtitles, captions, or other text information in a video, which may exist in the form of text files (such as SRT, VTT, etc.) or embedded in the video.

[0096] Metadata: Additional information associated with a video, such as timestamp, shooting location, camera settings, video resolution, frame rate, duration, etc.

[0097] Specifically, multimodal video data refers to the combination of multiple types of information (i.e., multiple "modalities") of video materials. This data includes not only traditional visual and auditory data, but also other types of data. Specifically, it includes:

[0098] Visual information (image data): video frames, image features (such as color, texture, shape, etc.), objects and people in the image, etc. These features can be extracted through image recognition algorithms or deep learning.

[0099] Audio information: The audio data itself, including the audio waveform, frequency information, and audio features (such as pitch, rhythm, and tone). Through sound processing and analysis algorithms, the audio's speech features (such as the speaker's emotions, language content, background noise, etc.) can be extracted.

[0100] Text information: subtitles, speech-to-text, annotations in the video, or other related text documents. This text information can provide additional context for more accurate analysis and understanding of the video content.

[0101] Emotion and mood analysis data: Analyzes the emotional or sentiment information in videos, inferring emotional states based on audio (speech sentiment analysis) or visual information (facial expression analysis, gesture recognition, etc.), and helps understand the emotional tone of the video.

[0102] Object and scene detection information in images or videos: Object detection algorithms (such as YOLO and Faster R-CNN) are used to analyze objects, scenes, or actions in video frames and extract specific content information, such as people, places, objects, and activities.

[0103] Time series data: The timing information of video frames and audio, and their synchronization on the timeline. It is especially important for analyzing editing, transitions, and rhythm.

[0104] Structured and unstructured metadata: Additional information related to the video is extracted through analysis of video materials, such as visual attributes such as shooting angle, focal length, and light intensity, or meta-information such as the theme, content, and genre of the video.

[0105] Special effects / transition data: Special effects data or shot switching information that the original video may contain, especially for video materials with special effects or animations.

[0106] Video editing information: If the material is obtained from an edited video, it may include pre-editing markers and editing point information, indicating which parts of the video material have been selected or removed.

[0107] S2. Analyze the obtained multimodal video data based on an anomaly recognition algorithm, mark abnormal material data, and generate a set of abnormal clips;

[0108] Specifically, abnormal material data includes: abnormal visual data, abnormal time series data, irrelevant or low-quality clips, materials that do not meet editing standards, abnormal scenes and characters, abnormal audio data, technical problems, etc.

[0109] S3. Compare the generated set of abnormal editing segments with known high-quality video material standards to evaluate the degree of influence of the abnormal editing segments on the quality of the finished editing;

[0110] S4. Based on the evaluation results, determine whether the set of abnormal clips meets the video editing standards, screen out materials that meet the video editing standards, and generate a high-quality material library for editing and store it;

[0111] It should be explained that the evaluation results of the collection of abnormal editing clips are compared with the known standards of high-quality video materials to screen out clips that meet the standards. For example, if a clip meets the standard requirements in terms of picture quality, sound quality, rhythm, emotional expression, etc., then the clip is considered qualified material; by comparing with the standards, those abnormal clips that fail to meet the editing standards are identified. These clips may be excluded due to poor quality, incoherent plot, unclear picture, distorted sound, etc.; according to the set editing standards (such as video clarity, sense of rhythm, effectiveness of information transmission, emotional expression, etc.), the matching degree of each clip is evaluated and the materials that meet these standards are screened out; after evaluation and comparison, the remaining clips that meet the editing standards will be screened as high-quality materials; the screened material collection that meets the standards will be stored in a special high-quality material library for subsequent video editing.

[0112] Based on the anomaly recognition algorithm, the obtained multimodal video data is analyzed and the abnormal material data is marked. The generation of a set of abnormal clips includes the following steps:

[0113] S21, quantizing the obtained multimodal video data, dividing the quantized multimodal video data into a training set and a test set, and setting a maximum number of iterations and a size of a candidate parameter set for the anomaly recognition algorithm;

[0114] S22. Utilize the random initialization strategy (i.e., convex lens imaging reverse learning) combined with the distribution characteristics of the feature space to generate the initial parameter combination of the anomaly recognition algorithm and configure the key parameters of the anomaly detection model;

[0115] S23. Dynamically adjust the weight of the objective function of the anomaly detection model based on the changing pattern of the abnormal features of the video clips (i.e., the changing pattern of the nonlinear s-shaped features of the Sigmoid function), combine the temporal characteristics of the abnormal clips, and set the feature priority;

[0116] S24. Using the anomaly detection model objective function, evaluate the recognition accuracy of the anomaly detection model for video data anomalies under each parameter combination to obtain a fitness value, and sort the fitness values ​​of all parameter combinations to select the currently optimal anomaly detection model parameters;

[0117] S25. Based on the distribution of abnormal features, calculate the nonlinear deviation of the anomaly detection model (i.e., the escape energy of the prey in the improved golden jackal optimization algorithm). If the nonlinear deviation is less than a threshold, adopt a local optimization strategy to adjust the parameter combination of the anomaly detection model. Otherwise, adopt a full local optimization strategy to adjust the direction of all parameter combinations of the anomaly detection model and adjust the weights of the auxiliary features.

[0118] S26. Based on the parameter selection algorithm, compare the optimization result of the current anomaly detection model with the known optimal parameter combination, and select the anomaly detection model parameters with the best performance;

[0119] S27. Randomly perturb the current optimal anomaly detection model parameters to generate new candidate anomaly detection model parameter combinations, and optimize the perturbed anomaly detection model parameters using a model optimization algorithm to obtain the local optimal anomaly detection model parameters in the perturbed area.

[0120] S28. If the local optimal anomaly detection model parameters are better than the current optimal anomaly detection model parameters, then record the current optimal anomaly detection model and update the anomaly segment set; otherwise, return to step S22 to continue optimization;

[0121] S29: Output the optimal anomaly detection model parameters and the anomaly labeling results of the anomaly detection model, and generate a set of clip anomaly segments based on the anomaly labeling results of the anomaly detection model.

[0122] Specifically, the anomaly recognition algorithm improves the golden jackal optimization algorithm. First, it integrates the convex lens imaging reverse learning strategy to explore a better initial solution position; introduces the Sigmoid function to change the escape energy and balance the exploration and development phases; and integrates the spiral encirclement mechanism of the whale optimization algorithm to enhance exploration capabilities and improve convergence accuracy. In the traditional golden jackal optimization algorithm, in the golden jackal combination, the male golden jackal is the first leader position, which is the optimal exploration position; the female golden jackal is a follower and is in a suboptimal position. In the golden jackal optimization algorithm, the golden jackal's hunting process is divided into three stages: (1) searching for prey and approaching the prey; (2) encircling the prey and stimulating the prey until it stops moving; (3) attacking the prey.

[0123] Specifically, in basic convex lens imaging reverse learning, the scaling factor is typically used as a fixed value, which can easily lead to local optima. This paper proposes a scaling factor strategy based on dynamic decrement. This strategy achieves a larger value in the early stages of iteration, allowing the jackal to conduct a wide-scale search across different dimensional regions, improving the diversity of the population. It then achieves a smaller value in the later stages of iteration, allowing for a more refined search near the optimal individual, improving local optimization capabilities.

[0124] Specifically, the prey's escape energy is the decisive factor controlling the prey's position update, responsible for balancing the algorithm's global search and local exploitation behaviors. The nonlinear sigmoid function's skewing characteristics allow for continuous global search to find the optimal region, thereby improving the slow convergence rate in the early stages and increasing the speed of convergence.

[0125] In order to facilitate understanding of the above technical solutions of the present invention, the following describes in detail how the present invention analyzes the obtained multimodal video data based on an abnormality recognition algorithm in a practical process, marks abnormal material data, and generates a set of abnormal clips.

[0126] Step 1: Multimodal video data quantization and data division

[0127] 1) Multimodal video data acquisition: Assume that a set of multimodal video data has been obtained from a video editing system, including visual data (video frames) and audio data (sound waveforms).

[0128] Visual data: The image size of each frame of the video is 1280×720, the total length of the video is 100 seconds, the frame rate is 30fps, and there are 3000 frames in total.

[0129] Audio data: Assume the audio sampling rate is 44.1 kHz, 44,100 samples per second, the total audio data duration is 100 seconds, and the total number of samples is 4,410,000.

[0130] 2) Quantification

[0131] For visual data, each frame of the image is converted into a grayscale image, and edge detection (such as the Canny algorithm) is performed to quantify the edge features of each frame of the image.

[0132] Perform spectral analysis on audio data to extract spectral features of the audio, such as MFCC (Mel Frequency Cepstral Coefficient).

[0133] 3) Data partitioning: All data are divided into a training set (80%) and a test set (20%). The training set contains 2400 frames of video and corresponding audio data, and the test set contains 600 frames of video and corresponding audio data.

[0134] 4) Set parameters

[0135] Maximum number of iterations: set to 100 iterations.

[0136] Candidate parameter set size: set to 10 candidate parameter sets.

[0137] Step 2: Initialize the anomaly recognition algorithm parameters

[0138] 1) Random Initialization: Use a random initialization strategy to generate an initial parameter combination from the distribution characteristics of the feature space. Assume that a support vector machine (SVM)-based anomaly detection model is used, and its key parameters include:

[0139] C (Penalty parameter): Initial value is 1.0; c (Kernel function parameters): The initial value is 0.01.

[0140] Randomly generate 10 sets of candidate parameter combinations, which are:

[0141] Parameter combination 1: C =1.0, c =0.01.

[0142] Parameter combination 2: C =2.0, c =0.05.

[0143] Parameter combination 3: C =0.5, c =0.02.

[0144] And so on, 10 candidate combinations are generated.

[0145] 2) Configure the model: Apply these initial parameter combinations to the anomaly detection model.

[0146] Step 3: Dynamically adjust objective function weights and feature priorities

[0147] 1) Objective function weight adjustment: Dynamically adjust the weight of the objective function in the anomaly detection model based on the changing patterns of clip anomaly features. Suppose that visual features (image clarity, edge detection) are more important, while audio features (spectral features) are less important. Therefore, the weights can be adjusted as follows:

[0148] Visual feature weight: 0.7.

[0149] Audio feature weight: 0.3.

[0150] 2) Temporal feature merging: Based on the temporal characteristics of the clips, a time window parameter can be added to model the continuity between video frames and increase the priority of temporal features. Assume that the time window size is 10 frames.

[0151] Step 4: Evaluate model fitness value

[0152] Fitness evaluation: Based on the training set data, evaluate the recognition accuracy of each parameter combination. Suppose we use the cross-validation method to evaluate the accuracy of the model, and the results are as follows:

[0153] Parameter combination 1: 95% accuracy.

[0154] Parameter combination 2: 92% accuracy.

[0155] Parameter combination 3: 89% accuracy.

[0156] Similarly, after the fitness values ​​of all parameter combinations are sorted, the optimal parameter combination is selected.

[0157] Step 5: Nonlinear deviation calculation and optimization strategy selection

[0158] 1) Nonlinear deviation calculation: Based on the distribution of abnormal features, the nonlinear deviation is calculated. Assume that the calculation result is: deviation value = 0.05, which is lower than the set threshold (0.1).

[0159] 2) Local optimization strategy: Since the deviation is less than the threshold, a local optimization strategy is used to further adjust the parameter combination of the anomaly detection model. Assume that after local optimization, the following updated parameter combination is obtained: C =1.2, c =0.03.

[0160] Step 6: Compare with the optimal parameter combination

[0161] 1) Optimal parameter comparison: Compare the current optimal parameters with the known optimal parameter combination. Assume that the optimal parameter combination is known from historical data: C =1.0, c =0.02.

[0162] Current optimal parameter combination C =1.2, c =0.03 has a slight improvement in accuracy, so the current parameter combination is selected.

[0163] Step 7: Parameter perturbation and local optimality acquisition

[0164] 1) Parameter perturbation: Perturb the parameters of the current optimal anomaly detection model to generate new candidate parameter combinations. For example:

[0165] Parameter perturbation 1: C =1.0, c =0.04

[0166] Parameter perturbation 2: C =1.3, c =0.01.

[0167] 2) Local optimal calculation: Use the optimization algorithm to optimize the parameters after the disturbance. Assume that the parameter combination obtained through local optimization is: C =1.1, c =0.025, the accuracy rate increased to 96%.

[0168] Step 8: Local Optimal Parameter Update

[0169] 1) Update the optimal model: Since the accuracy of the local optimal parameter combination is better than the current optimal model, update the optimal model and record the current optimal parameter combination: C =1.1, c =0.025.

[0170] 2) Update the set of abnormal clips: Based on the current optimal parameters, apply the anomaly detection model to analyze the video data and generate a new set of abnormal clips. Assume that the model has marked 200 abnormal clips (such as blurry images, excessive background noise, editing errors, etc.).

[0171] Step 9: Output the optimal model parameters and anomaly marking results

[0172] 1) Output the optimal parameter combination: C =1.1, c =0.025.

[0173] 2) Output abnormal marking results:

[0174] Marking abnormal segments: 200 segments (e.g., numbered 105, 205, 405, etc.) were identified as abnormal clips.

[0175] 3) Generate a collection of abnormal clips: Based on the model labeling results, generate a collection of abnormal clips, which contains all clips marked as abnormal for further processing (such as removal or replacement).

[0176] Preferably, based on a parameter selection algorithm, the optimization result of the current anomaly detection model is compared with a known optimal parameter combination, and the selection of the anomaly detection model parameters with the best performance includes the following steps:

[0177] S261. Define the key parameter search range and feature selection range of the anomaly detection model;

[0178] S262. Randomly generate several key parameter configurations and feature subsets, each combination of a key parameter configuration and a feature subset represents a candidate parameter configuration, and all combinations constitute a candidate parameter configuration group;

[0179] S263. Train the anomaly detection model using each candidate parameter configuration, calculate its fitness value, and record the performance results of the current candidate parameter configuration;

[0180] S264: Perform a neighborhood search near the key parameter configuration with the highest fitness in the candidate parameter configuration set, and update and record the key parameter configuration and feature subset with the highest fitness;

[0181] S265: Distribute the remaining candidate parameter configurations to random key parameter configurations and feature subsets in the global range, calculate their fitness values, and update the global optimal key parameter configuration;

[0182] S266. Repeat steps S263 to S265 until the maximum number of iterations is reached, and compare the current optimal key parameter configuration with the known optimal parameter configuration. If the current optimal key parameter configuration is better than the known optimal parameter configuration, replace it with the new optimal parameter configuration as the anomaly detection model parameter with the best performance.

[0183] Specifically, the parameter selection algorithm is the honey bee algorithm, a new swarm intelligence search algorithm that simulates the foraging behavior of honey bees in nature to find the optimal solution to a problem. This proposed optimization algorithm uses the location of a food source as a set of solutions to the optimization problem. The location of a food source and its immediate neighborhood form a small search area, called a flower cluster. During the food search, a honey bee swarm first sends out scout bees. These scout bees continuously fly from one food source to another, searching for all possible food sources. When the profitability of a food source (such as its sugar content) exceeds a certain level, the scout bees return to the hive, unload honey, and perform a magical "waggle dance" on a "dance board." This "waggle dance" is an essential tool for honey bee communication, enabling the entire swarm to understand the location, distance to the hive, and profitability of the food source. After the dance, the scout bees lead the follower bees waiting outside the hive back to the corresponding food source to gather honey. Highly profitable food sources attract more follower bees to gather honey. During the honey-gathering process, bees calculate the profitability of the food source they are currently collecting honey from and return to the hive to perform a "waggle dance" to convey this information to the colony. If the food source's profitability remains high, more follower bees will be recruited to collect honey; if the food source's profitability is low, the current food source will be abandoned.

[0184] Preferably, performing a neighborhood search near the key parameter configuration with the highest current fitness in the candidate parameter configuration set, and updating and recording the key parameter configuration and feature subset with the highest fitness comprises the following steps:

[0185] S2641. Identify the key parameter configuration with the highest fitness value in the current candidate parameter configuration set as the center of the neighborhood search;

[0186] S2642. Define a neighborhood range (i.e., nearest neighbor search) according to preset rules, set a perturbation amplitude for each dimension of the key parameter with the current optimal parameter configuration as the center, and calculate the neighborhood separately for each key parameter configuration to form a multi-dimensional search space;

[0187] S2643. Randomly generate several new key parameter configurations within the neighborhood, retrain the anomaly detection model using the new parameter configurations, and calculate the fitness value of each new key parameter configuration;

[0188] S2644. Compare the fitness values ​​of all new key parameter configurations within the domain with the fitness value of the current optimal parameter configuration. If the fitness value of the new key parameter configuration is greater than the fitness value of the current optimal parameter configuration, the new key parameter configuration is used as the key parameter configuration with the highest fitness. At the same time, the feature subset corresponding to the key parameter configuration with the highest fitness is updated and recorded. Otherwise, the current optimal parameter configuration is retained unchanged.

[0189] Preferably, randomly perturbing the current optimal anomaly detection model parameters to generate new candidate anomaly detection model parameter combinations, optimizing the perturbed anomaly detection model parameters using a model optimization algorithm, and obtaining the local optimal anomaly detection model parameters in the perturbed area include the following steps:

[0190] S271. Initialize the parameters of the model optimization algorithm and randomly select a number of candidate anomaly detection model parameter configurations within the disturbance region;

[0191] S272. Input the candidate anomaly detection model parameters (i.e., each star in the black hole optimization algorithm) into the anomaly detection model, and calculate the fitness value of each candidate anomaly detection model parameter configuration through the objective function;

[0192] S273. Select the anomaly detection model parameter configuration with the highest fitness value from all candidate anomaly detection model parameter configurations as the current optimal anomaly detection model parameter solution (i.e., the black hole in the black hole optimization algorithm);

[0193] S274. Update the position of each anomaly detection model parameter configuration according to the position update formula;

[0194] S275. If the fitness value of any candidate parameter is greater than the fitness value of the current optimal anomaly detection model parameter, then update the optimal anomaly detection model parameter solution, remove the anomaly detection model parameter configuration with the lowest fitness value from the current candidate anomaly detection model parameter set, and randomly generate a new anomaly detection model parameter configuration within the disturbance region of the current optimal anomaly detection model parameter.

[0195] S276. Check whether the maximum number of iterations is reached. If so, output the current optimal anomaly detection model parameter configuration as the local optimal anomaly detection model parameter of the disturbance area.

[0196] Specifically, the model optimization algorithm is the black hole optimization algorithm, and the idea of ​​the algorithm relatively completely describes the general characteristics of the black hole phenomenon in nature; based on these characteristics of the black hole optimization algorithm, the optimization search principle of the black hole optimization algorithm is relatively simple and easy to implement. The black hole optimization algorithm mainly simulates the actual black hole phenomenon, randomly arranges a certain number of stars in a certain search space, and determines and evaluates the fitness function of each star in the search space through statistical means, and selects a star with the best fitness value as the black hole. The boundary of this black hole is regarded as the area where the current global optimal solution is located, and the black hole itself is regarded as the current global optimal solution.

[0197] Specifically, the anomaly detection model is the Kernel Extreme Learning Machine (KELM), a new artificial intelligence prediction model. It builds on the Extreme Learning Machine (ELM) by incorporating the kernel function concept. This effectively overcomes the ELM's problem of low output stability due to the random generation of initial weights and thresholds, improving output stability. However, its prediction performance is often affected by parameter selection. Therefore, the present invention uses variational mode decomposition to decompose the actual time series signal. It then incorporates grey correlation analysis to analyze the correlation between the input and output variables of the pitch-coordinated control model, increasing the data correlation. Based on the KLEM model, a model optimization algorithm (i.e., the black hole optimization algorithm) is used to optimize its parameters. Variational Mode Decomposition (VMD) is a signal processing technique widely used to analyze nonstationary and nonlinear signals. It uses a variational method to decompose the signal into a set of intrinsic mode functions (IMFs) with distinct center frequencies. These IMFs represent the different frequency components in the signal and can be used for feature extraction, predictive modeling, and pattern recognition.

[0198] Preferably, the location update formula is:

[0199] ;

[0200] Where, Y i ( s +1) indicates the i The anomaly detection model parameters are configured in s +1 position at iteration time; rand Represents a random number between [0, 1]; Y Pc represents the location of the optimal anomaly detection model parameter scheme within the disturbance region; PC Represents the optimal anomaly detection model parameter scheme.

[0201] Preferably, comparing the generated set of abnormal editing segments with known high-quality video material standards to evaluate the degree of influence of the abnormal editing segments on the quality of the finished editing includes the following steps:

[0202] S31, importing the generated abnormal editing segment set and known high-quality video material standards, setting evaluation parameters, and pre-processing the abnormal editing segment set and the high-quality video material;

[0203] S32. Screening out the editing features that have the greatest impact on the quality of the finished editing through a feature screening algorithm;

[0204] S33, dividing the abnormal clip set into a training set and a test set;

[0205] S34, randomly initializing the parameter configuration of the clip evaluation model, defining the parameter search space as multi-dimensional, and using each set of parameter configurations as a combination of evaluation criteria;

[0206] S35, using a cross-validation method to evaluate the fitness value of each parameter configuration in the training set, recording the fitness value of the current parameter configuration, and marking the optimal fitness value among all parameter configurations;

[0207] S36, using a heuristic algorithm to adjust the evaluation model parameter configuration and generate a new parameter configuration;

[0208] S37, using the new parameter configuration to evaluate the training set, calculate the fitness value of each abnormal clip set under the new parameter configuration; and update the optimal fitness value and its corresponding parameter configuration among all parameter configurations;

[0209] S38, check whether the maximum number of iterations has been reached, if so, proceed to step S39, otherwise return to step S36, continue to adjust the evaluation model parameter configuration; output the evaluation model parameter configuration with the highest fitness value;

[0210] S39. Use the optimal parameter configuration to evaluate the abnormal editing segments in the test set, quantify the matching degree between the abnormal editing segments and the high-quality standard, calculate the impact score of the editing abnormality segment set on the quality of the finished editing, and output the evaluation results.

[0211] It needs to be explained that relevant feature data of abnormal clips, including visual, audio and timing features, such as picture clarity, background noise intensity, shot switching smoothness, etc., are collected; relevant feature data of abnormal clips, including visual, audio and timing features, such as picture clarity, background noise intensity, shot switching smoothness, etc., are collected; the weight distribution and evaluation range of key features are defined; the set of abnormal editing clips and high-quality video materials are normalized and feature extracted so as to be analyzed in a unified format; feature screening algorithms are used to extract the core features that have the greatest impact on the quality of the finished editing; based on the algorithm results, the most important features in visual, audio and timing features are screened out as the basis for subsequent analysis; the set of abnormal editing clips is divided into training sets and test sets; the training set is used for model optimization, and the test set is used to verify the effectiveness of the final evaluation model; a suitable editing evaluation model is selected, such as a model based on machine learning or statistical analysis; the search space of model parameters is defined, including the range of feature weights and the specific hyperparameters of the model; randomly generate initial parameter configurations to provide a basis for subsequent optimization; evaluate the fitness value of each set of parameter configurations on the training set to measure the performance of the editing evaluation model for abnormal fragment features; record the current fitness value and select the parameter configuration with the best performance; use heuristic algorithms (such as genetic algorithms, particle swarm optimization, crow search algorithms, etc.) to generate new parameter combinations based on the current parameter configuration; the parameter configuration will be further evaluated in subsequent steps; use the newly generated parameter configuration to evaluate the training set and calculate the fitness value of each set of parameters; if the fitness value of the new parameter configuration is better than the previous optimal value, update the optimal configuration; check whether the number of iterations has reached the maximum value; if the maximum value has not been reached, return to the parameter adjustment step to continue optimization; if it has been reached, end the iteration and lock the current optimal parameter configuration; use the optimal parameter configuration to evaluate the abnormal editing fragments in the test set; quantify the matching degree of the abnormal fragments with the high-quality video material standards, and output the score of the finished editing quality and the final evaluation report.

[0212] Specifically, the present invention compares the generated abnormal editing segments with known high-quality video material standards by improving the longhorn beetle swarm algorithm to evaluate the degree of impact of the abnormal editing segments on the quality of the finished editing. The traditional longhorn beetle swarm algorithm only considers the behavior of individuals learning towards global and local extremes when updating speed, while ignoring the judgment made by individual longhorn beetles on the surrounding environment. The present invention comprehensively considers various influencing factors when updating the longhorn beetle speed. In order to better coordinate global and local searches, a feature screening algorithm is introduced for feature screening.

[0213] Preferably, screening out the editing features that have the greatest impact on the quality of the finished editing by a feature screening algorithm includes the following steps:

[0214] S321, initializing all parameters in the feature screening algorithm;

[0215] S322, randomly select a feature combination as an initial feature candidate set, and initialize the feature weights using chaos thinking;

[0216] S323, evaluating the impact of each candidate feature combination on the editing quality, calculating the fitness value of the candidate feature combination, and finding the feature combination with the best performance at present;

[0217] S324: Update the weight of the current feature combination to generate a new candidate feature combination;

[0218] S325. Check whether the updated feature combination meets the feature quantity threshold. If so, update the feature combination according to the update rule. Otherwise, keep the current feature combination unchanged.

[0219] S326, recalculate the fitness value of the updated feature combination, and compare it with the previous feature combination to find the feature combination with the highest fitness value;

[0220] S327, comparing the fitness values ​​of the current feature combination and the historical optimal feature combination, and recording the feature combination with the highest fitness as the new optimal feature combination;

[0221] S328. Check whether the maximum number of iterations has been reached. If so, output the optimal feature combination as the editing feature that has the greatest impact on the quality of the finished editing. Otherwise, return to step S324 and continue iterative optimization.

[0222] Specifically, the feature screening algorithm is an improved crow search algorithm, which is a swarm intelligence optimization algorithm inspired by the memory behavior and food hiding strategy of crows. Crows search for food in groups, memorize the location of food and prevent other crows from stealing food. CSA solves complex optimization problems by simulating these behaviors of crows. When initializing the crow population, the present invention selects cubic chaotic mapping to produce a more evenly distributed population, thereby increasing population diversity. Compared with the logistic mapping structure, the cubic mapping is simpler, and most importantly, it can obtain a more evenly distributed population. The crows' activity range will also become larger during the entire iterative optimization process.

[0223] Preferably, adjusting the evaluation model parameter configuration using a heuristic algorithm to generate a new parameter configuration includes the following steps:

[0224] S361. Set the parameters of the heuristic algorithm and randomly generate a parameter configuration population matrix, where each row represents a candidate parameter configuration and each column represents an evaluation model parameter;

[0225] S362, sequentially evaluating the fitness value of each parameter configuration to check whether the parameter configuration exceeds the upper and lower limits. If so, adjusting the parameter configuration to the boundary value; otherwise, retaining the current parameter configuration and calculating the fitness value of each parameter configuration using the fitness function;

[0226] S363. During the initial iteration, the parameter configurations are sorted from small to large according to the fitness values. The sorted parameter configuration population is the optimal parameter configuration population. The first parameter configuration in the optimal parameter configuration population is selected as the optimal parameter configuration. During subsequent iterations, the current optimal parameter configuration population is merged with the previous optimal parameter configuration population, and the population is re-sorted according to the fitness values ​​to select a new optimal parameter configuration population. The first parameter configuration in the new optimal parameter configuration population is selected as the optimal parameter configuration.

[0227] S364. Update the parameter configurations in the parameter configuration population one by one according to the update rule of the heuristic algorithm;

[0228] S365. Check whether the maximum number of iterations is reached. If so, output the optimal parameter configuration and the corresponding fitness value as the new parameter configuration. Otherwise, return to step S362 and continue executing until the maximum number of iterations is reached.

[0229] Specifically, the heuristic algorithm is an improved moth-to-flame algorithm. The main inspiration for the traditional moth-to-flame algorithm comes from the lateral navigation flight mechanism of moths. When flying at night, moths maintain a fixed angle to the moon. This behavior enables them to always fly in a straight line when flying long distances. In daily life, moths mistake artificial light for moonlight, so moths move in a spiral curve around the artificial light. In order to improve the performance of the traditional moth-to-flame algorithm, the present invention proposes some improvement measures. After one iteration of the moth-to-flame algorithm is completed, a simplex search strategy is used to select several moth positions with poor positions for optimization. The simplex search method is an early developed unconstrained optimization method. The basic idea of ​​this algorithm is similar to that of a blind man climbing a mountain. Each time he moves forward, he first moves the stick forward to test a few points. When he finds the highest point, he takes a step forward.

[0230] Preferably, the formula for the update rule is:

[0231] ;

[0232] Where, A e Indicates the e parameter configurations (i.e., the moth in the moth-to-flame algorithm); M n Indicates the n target parameter configuration (i.e. the flame in the moth-to-flame algorithm); G e Indicates thee Parameter configuration and n The difference between the target parameter configurations (i.e., the distance from the moth to the flame in the moth-to-flame algorithm); b represents a constant that defines the shape of a logarithmic spiral; t Represents a random number between [0, 1]; B represents the spiral function.

[0233] According to another embodiment of the present invention, Figure 2 As shown, a judgment system based on AI automatic editing of video materials is also provided, and the judgment system includes:

[0234] Data acquisition module 1, used to acquire video material data to be edited and pre-process the video material data to obtain multimodal video data;

[0235] Anomaly recognition module 2 is used to analyze the obtained multimodal video data based on anomaly recognition algorithm, mark abnormal material data, and generate a set of abnormal clips;

[0236] The quality comparison module 3 is used to compare the generated set of abnormal editing segments with known high-quality video material standards to evaluate the degree of influence of the abnormal editing segments on the quality of the finished editing;

[0237] Evaluation and judgment module 4 is used to judge whether the set of abnormal editing segments meets the video editing standards based on the evaluation results, screen out the materials that meet the video editing standards, and generate a high-quality material library for editing for storage;

[0238] The data acquisition module 1 is connected to the quality comparison module 3 via the abnormality recognition module 2 , and the quality comparison module 3 is connected to the evaluation and judgment module 4 .

[0239] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A judgment method for automatically editing video materials based on AI, characterized in that: The determination method includes the following steps: S1. Obtain video material data to be edited, and pre-process the video material data to obtain multimodal video data; S2. Analyze the obtained multimodal video data based on an anomaly recognition algorithm, mark abnormal material data, and generate a set of abnormal clips; S3. Compare the generated set of abnormal editing segments with known high-quality video material standards to evaluate the degree of influence of the abnormal editing segments on the quality of the finished editing; S4. Based on the evaluation results, determine whether the set of abnormal clips meets the video editing standards, screen out materials that meet the video editing standards, and generate a high-quality material library for editing and store it; The method of analyzing the obtained multimodal video data based on the anomaly recognition algorithm, marking the abnormal material data, and generating a set of abnormal clips includes the following steps: S21, quantizing the obtained multimodal video data, dividing the quantized multimodal video data into a training set and a test set, and setting a maximum number of iterations and a size of a candidate parameter set for the anomaly recognition algorithm; S22. Generate an initial parameter combination for the anomaly recognition algorithm using a random initialization strategy combined with the distribution characteristics of the feature space, and configure key parameters of the anomaly detection model; S23. Dynamically adjust the weight of the objective function of the anomaly detection model according to the changing pattern of the abnormal features of the video clips, combine the temporal characteristics of the abnormal clips, and set the feature priority; S24. Using the anomaly detection model objective function, evaluate the recognition accuracy of the anomaly detection model for video data anomalies under each parameter combination to obtain a fitness value, and sort the fitness values ​​of all parameter combinations to select the currently optimal anomaly detection model parameters; S25. Based on the distribution of abnormal features, calculate the nonlinear deviation of the anomaly detection model. If the nonlinear deviation is less than the threshold, adopt a local optimization strategy to adjust the parameter combination of the anomaly detection model. Otherwise, adopt a full local optimization strategy to adjust the direction of all parameter combinations of the anomaly detection model and adjust the weights of the auxiliary features. S26. Based on the parameter selection algorithm, compare the optimization result of the current anomaly detection model with the known optimal parameter combination, and select the anomaly detection model parameters with the best performance; S27. Randomly perturb the current optimal anomaly detection model parameters to generate new candidate anomaly detection model parameter combinations, and optimize the perturbed anomaly detection model parameters using a model optimization algorithm to obtain the local optimal anomaly detection model parameters in the perturbed area. S28. If the local optimal anomaly detection model parameters are better than the current optimal anomaly detection model parameters, then record the current optimal anomaly detection model and update the anomaly segment set; otherwise, return to step S22 to continue optimization; S29, outputting the optimal anomaly detection model parameters and the anomaly labeling results of the anomaly detection model, and generating a set of clip anomaly segments according to the anomaly labeling results of the anomaly detection model; The parameter selection algorithm is based on comparing the optimization result of the current anomaly detection model with the known optimal parameter combination, and selecting the anomaly detection model parameters with the best performance includes the following steps: S261. Define the key parameter search range and feature selection range of the anomaly detection model; S262. Randomly generate several key parameter configurations and feature subsets, each combination of a key parameter configuration and a feature subset represents a candidate parameter configuration, and all combinations constitute a candidate parameter configuration group; S263. Train the anomaly detection model using each candidate parameter configuration, calculate its fitness value, and record the performance results of the current candidate parameter configuration; S264: Perform a neighborhood search near the key parameter configuration with the highest fitness in the candidate parameter configuration set, and update and record the key parameter configuration and feature subset with the highest fitness; S265: Distribute the remaining candidate parameter configurations to random key parameter configurations and feature subsets in the global range, calculate their fitness values, and update the global optimal key parameter configuration; S266. Repeat steps S263 to S265 until the maximum number of iterations is reached, and compare the current optimal key parameter configuration with the known optimal parameter configuration. If the current optimal key parameter configuration is better than the known optimal parameter configuration, replace it with the new optimal parameter configuration as the anomaly detection model parameter with the best performance.

2. The method for determining the AI-based automatic editing of video materials according to claim 1, characterized in that: The performing of neighborhood search near the key parameter configuration with the highest fitness in the candidate parameter configuration set, and updating and recording the key parameter configuration and feature subset with the highest fitness comprises the following steps: S2641. Identify the key parameter configuration with the highest fitness value in the current candidate parameter configuration set as the center of the neighborhood search; S2642. Define a neighborhood range according to preset rules, set a perturbation amplitude for each dimension of the key parameter with the current optimal parameter configuration as the center, and calculate a neighborhood separately for each key parameter configuration to form a multi-dimensional search space; S2643. Randomly generate several new key parameter configurations within the neighborhood, retrain the anomaly detection model using the new parameter configurations, and calculate the fitness value of each new key parameter configuration; S2644. Compare the fitness values ​​of all new key parameter configurations within the domain with the fitness value of the current optimal parameter configuration. If the fitness value of the new key parameter configuration is greater than the fitness value of the current optimal parameter configuration, the new key parameter configuration is used as the key parameter configuration with the highest fitness. At the same time, the feature subset corresponding to the key parameter configuration with the highest fitness is updated and recorded. Otherwise, the current optimal parameter configuration is retained unchanged.

3. The method for determining the automatic editing of video materials based on AI according to claim 1, characterized in that: The method of randomly perturbing the current optimal anomaly detection model parameters to generate a new candidate anomaly detection model parameter combination, optimizing the perturbed anomaly detection model parameters using a model optimization algorithm, and obtaining the local optimal anomaly detection model parameters in the perturbed area includes the following steps: S271. Initialize the parameters of the model optimization algorithm and randomly select a number of candidate anomaly detection model parameter configurations within the disturbance region; S272: Input the candidate anomaly detection model parameter configuration into the anomaly detection model, and calculate the fitness value of each candidate anomaly detection model parameter configuration through the objective function; S273. Select the anomaly detection model parameter configuration with the highest fitness value from all candidate anomaly detection model parameter configurations as the current optimal anomaly detection model parameter solution; S274. Update the position of each anomaly detection model parameter configuration according to the position update formula; S275. If the fitness value of any candidate parameter is greater than the fitness value of the current optimal anomaly detection model parameter, then update the optimal anomaly detection model parameter solution, remove the anomaly detection model parameter configuration with the lowest fitness value from the current candidate anomaly detection model parameter set, and randomly generate a new anomaly detection model parameter configuration within the disturbance region of the current optimal anomaly detection model parameter. S276. Check whether the maximum number of iterations is reached. If so, output the current optimal anomaly detection model parameter configuration as the local optimal anomaly detection model parameter of the disturbance area.

4. The method for determining the AI-based automatic editing of video materials according to claim 3, characterized in that: The position update formula is: AND i (s+1)=Y i (s)+rand·[Y Pc -AND i (s)]; Where Y i (s+1) represents the position of the i-th anomaly detection model parameter configuration at the s+1 iteration; rand represents a random number; Y Pc It represents the location of the optimal anomaly detection model parameter scheme in the disturbance area; Pc represents the optimal anomaly detection model parameter scheme.

5. The method for determining the automatic editing of video materials based on AI according to claim 1, characterized in that: Comparing the generated set of abnormal editing segments with known high-quality video material standards to evaluate the impact of the abnormal editing segments on the quality of the finished editing includes the following steps: S31, importing the generated abnormal editing segment set and known high-quality video material standards, setting evaluation parameters, and pre-processing the abnormal editing segment set and the high-quality video material; S32. Screening out the editing features that have the greatest impact on the quality of the finished editing through a feature screening algorithm; S33, dividing the abnormal clip set into a training set and a test set; S34, randomly initializing the parameter configuration of the clip evaluation model, defining the parameter search space as multi-dimensional, and using each set of parameter configurations as a combination of evaluation criteria; S35, using a cross-validation method to evaluate the fitness value of each parameter configuration in the training set, recording the fitness value of the current parameter configuration, and marking the optimal fitness value among all parameter configurations; S36. Using a heuristic algorithm to adjust the evaluation model parameter configuration to generate a new parameter configuration; S37, using the new parameter configuration to evaluate the training set, calculate the fitness value of each abnormal clip set under the new parameter configuration; and update the optimal fitness value and its corresponding parameter configuration among all parameter configurations; S38, check whether the maximum number of iterations has been reached, if so, proceed to step S39, otherwise return to step S36, continue to adjust the evaluation model parameter configuration; output the evaluation model parameter configuration with the highest fitness value; S39. Use the optimal parameter configuration to evaluate the abnormal editing segments in the test set, quantify the matching degree between the abnormal editing segments and the high-quality standard, calculate the impact score of the editing abnormality segment set on the quality of the finished editing, and output the evaluation results.

6. The method for determining the AI-based automatic editing of video materials according to claim 5, characterized in that: The method of screening out the editing features that have the greatest impact on the quality of the finished editing by using the feature screening algorithm includes the following steps: S321, initializing all parameters in the feature screening algorithm; S322, randomly select a feature combination as an initial feature candidate set, and initialize the feature weights using chaos thinking; S323, evaluating the impact of each candidate feature combination on the editing quality, calculating the fitness value of the candidate feature combination, and finding the feature combination with the best performance at present; S324: Update the weight of the current feature combination to generate a new candidate feature combination; S325. Check whether the updated feature combination meets the feature quantity threshold. If so, update the feature combination according to the update rule. Otherwise, keep the current feature combination unchanged. S326, recalculate the fitness value of the updated feature combination, and compare it with the previous feature combination to find the feature combination with the highest fitness value; S327, comparing the fitness values ​​of the current feature combination and the historical optimal feature combination, and recording the feature combination with the highest fitness as the new optimal feature combination; S328. Check whether the maximum number of iterations has been reached. If so, output the optimal feature combination as the editing feature that has the greatest impact on the quality of the finished editing. Otherwise, return to step S324 and continue iterative optimization.

7. The method for determining the automatic editing of video materials based on AI according to claim 5, characterized in that: The method of adjusting the evaluation model parameter configuration using a heuristic algorithm to generate a new parameter configuration includes the following steps: S361. Set the parameters of the heuristic algorithm and randomly generate a parameter configuration population matrix, where each row represents a candidate parameter configuration and each column represents an evaluation model parameter; S362, sequentially evaluating the fitness value of each parameter configuration to check whether the parameter configuration exceeds the upper and lower limits. If so, adjusting the parameter configuration to the boundary value; otherwise, retaining the current parameter configuration and calculating the fitness value of each parameter configuration using the fitness function; S363. During the initial iteration, the parameter configurations are sorted from small to large according to the fitness values. The sorted parameter configuration population is the optimal parameter configuration population. The first parameter configuration in the optimal parameter configuration population is selected as the optimal parameter configuration. During subsequent iterations, the current optimal parameter configuration population is merged with the previous optimal parameter configuration population, and the population is re-sorted according to the fitness values ​​to select a new optimal parameter configuration population. The first parameter configuration in the new optimal parameter configuration population is selected as the optimal parameter configuration. S364. Update the parameter configurations in the parameter configuration population one by one according to the update rule of the heuristic algorithm; S365. Check whether the maximum number of iterations is reached. If so, output the optimal parameter configuration and the corresponding fitness value as the new parameter configuration. Otherwise, return to step S362 and continue executing until the maximum number of iterations is reached.

8. The method for determining the automatic editing of video materials based on AI according to claim 7, characterized in that: The formula of the update rule is: Where A e Indicates the e-th parameter configuration; M n Indicates the nth target parameter configuration; G e represents the difference between the e-th parameter configuration and the n-th target parameter configuration; b represents a constant defining the shape of the logarithmic spiral; t represents a random number; and B represents the spiral function.

9. A judgment system based on AI automatic editing of video materials, used to implement the judgment method based on AI automatic editing of video materials according to any one of claims 1 to 8, characterized in that: The judgment system includes: The data acquisition module is used to obtain the video material data to be edited and pre-process the video material data to obtain multimodal video data; Anomaly recognition module, which is used to analyze the obtained multimodal video data based on the anomaly recognition algorithm, mark the abnormal material data, and generate a set of abnormal clips; A quality comparison module is used to compare the generated set of abnormal editing clips with known high-quality video material standards to evaluate the impact of the abnormal editing clips on the quality of the finished editing; An evaluation and judgment module is used to determine whether the set of abnormal editing clips meets the video editing standards based on the evaluation results, screen out materials that meet the video editing standards, and generate a high-quality material library for editing for storage; Among them, the data acquisition module is connected through the abnormality recognition module and the quality comparison module, and the quality comparison module is connected with the evaluation and judgment module.

Images