A mixed video clip homologous detection method, device and equipment and storage medium

Abstract

This application discloses a method, apparatus, device, and storage medium for detecting homogeneity in multi-source fragmented mixed-edit videos, relating to the field of video editing. The method includes: performing temporal slicing of the video to be detected and extracting features from each slice to obtain corresponding target feature vectors; determining target parent slices based on the target similarity between each parent feature vector and the target feature vectors; and constructing a target hit relationship between the slice to be detected and the target parent slices; grouping the target hit relationships based on the parent identifiers of the target parent slices to construct a target attribution matrix between the video to be detected and each target parent video; determining the continuous reference duration, relative concentration ratio, and core plot coverage rate of each target parent video based on the target attribution matrix to determine a target coverage intensity coefficient; and generating a homogeneity detection result evidence package based on the target coverage intensity coefficient. This application achieves structured homogeneity detection in multi-source fragmented mixed-edit scenarios.

Description

Technical Field

[0001] This invention relates to the field of video editing, and in particular to a method, apparatus, device, and storage medium for detecting the same source in mixed-edit videos. Background Technology

[0002] With the rapid development of short video platforms and the self-media content ecosystem, a large number of mashup videos have emerged online. These mashup videos typically splice together clips from multiple copyrighted films and television series for secondary dissemination, deliberately controlling the continuous usage duration of a single copyrighted material, or distributing the same copyrighted material across multiple locations within the video to circumvent existing detection mechanisms based on duration thresholds or hit frequency. Existing detection solutions struggle to effectively quantify the usage ratio of each source material in these scenarios, and cannot distinguish whether the referenced content is concentrated in the core, high-value segments of the original film, leading to missed detections. Currently, the field of video copyright content same-source detection typically employs the following technical approaches: frame-level comparison based on hash fingerprints (such as perceptual hashing, pHash), which performs global matching by extracting the visual hash values ​​of video frames; sound comparison based on audio frequency domain features (such as Mel spectrum, acoustic fingerprinting); and video feature embedding and vector retrieval based on deep learning, mapping video clips to a high-dimensional semantic space for similarity calculation. In large-scale retrieval scenarios, Locality Sensitive Hashing (LSH) or Approximate Nearest Neighbor (ANN) indexes are typically introduced as coarse-grained screening methods to reduce the computational overhead of full-database fine-grained ranking. Existing video homology detection schemes typically use the number of hits of content features per unit time (i.e., cumulative hit frequency count) as the core criterion. The main steps are: extracting frame-level features from the video to be tested and matching them frame-by-frame with the parent feature library; counting the number of hits within a unit time window; and determining and reporting homology risk when the number of hits exceeds a preset threshold. Since existing schemes use cumulative hit frequency count as the core criterion, their underlying logic is that the number of hit segments is used as a proxy indicator for homology risk. However, there is no linear relationship between the number of hit segments and the actual impact of content reuse. For example, a 5-minute core plot segment condensed into a 30-second reference may have very few hit frames, but it occupies a large portion of the original core plot; conversely, a 30-second black screen or transition may generate a large number of hash hits, but its actual value is extremely low. This design flaw makes the solution insufficient for handling "condensed highlight" type editing strategies. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a method, apparatus, device, and storage medium for detecting the same source in multi-source fragmented mixed-edit videos, capable of achieving structured same-source detection in such scenarios. The specific solution is as follows: Firstly, this application discloses a method for detecting the same source in mixed-edit videos, including: The video to be detected is divided into temporal slices, and the target feature vectors corresponding to each slice are extracted. The target parent slice is determined based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, and the target hit relationship between the slice to be detected and the target parent slice is constructed. Based on the parent identifier corresponding to the target parent video slice, the target hit relationships are grouped, and a target attribution matrix is ​​constructed between the video to be detected and each target parent video based on the target grouping results; different target attribution matrices are independent of each other; the target attribution matrix is ​​used to record the time interval of the video to be detected, the time interval of the target parent video, and the target saliency weight corresponding to the target parent video; the target saliency weight characterizes the degree to which the target parent video slice carries the core plot content of the corresponding parent video; Based on the target attribution matrix, the continuous reference duration, relative concentration ratio, and core plot coverage of each target parent video in the video to be detected are determined. The target coverage intensity coefficient corresponding to the target parent video is determined based on the continuous reference duration, relative concentration ratio, and core plot coverage, and a homology detection result evidence package is generated based on the target coverage intensity coefficient. The relative concentration ratio represents the proportion of the video content of the target parent video in the video to be detected, and the core plot coverage represents the degree to which the video to be detected covers the target core plot of the target parent video.

[0004] Optionally, before determining the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, the method further includes: Each source video is acquired, and the source videos are time-series sliced ​​to obtain corresponding source slices. Feature extraction is performed on the source slices to obtain source feature vectors corresponding to the source slices. The source feature vectors include visual dimension feature vectors, acoustic dimension feature vectors, and text dimension feature vectors. The visual dimension feature vector, the acoustic dimension feature vector, and the text dimension feature vector are weighted and fused based on the first preset weighting strategy, so as to determine the target saliency weight corresponding to the parent slice based on the parent feature vector, and the parent feature vector corresponding to each parent video is written into the target parent feature library.

[0005] Optionally, determining the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector includes: Using the Locality Sensitive Hash algorithm, based on the Euclidean distance between the target feature vector and each parent feature vector in the target parent feature library, the candidate parent feature vector corresponding to the slice to be detected is determined, and the target parent feature vector is determined based on the target similarity between the candidate parent feature vector and the target feature vector. The parent slice corresponding to the target parent feature vector is then determined as the target parent slice.

[0006] Optionally, after determining the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, the method further includes: If the same slice to be detected corresponds to at least two target parent slices, then the target parent slice with the highest target similarity is determined as the unique parent slice corresponding to the slice to be detected. Accordingly, constructing the target hit relationship between the slice to be detected and the target parent slice includes: Construct the target hit relationship between the slice to be detected and the unique parent slice.

[0007] Optionally, determining the continuous reference duration, relative concentration ratio, and core plot coverage of each target parent video in the video to be detected based on the target attribution matrix includes: Based on the target attribution matrix, the target parent video is determined to have a set of hit segments to be detected. Then, based on the time axis of the target video, the time intervals of each target segment in the set of hit segments to be detected are deduplicated and merged to determine the first hit interval corresponding to each target parent video. The first hit interval is a continuous interval. Based on the interval duration corresponding to each of the first hit intervals, the longest consecutive hit interval and the total hit duration corresponding to the target parent video are determined, and the continuous reference duration corresponding to the target parent video is determined based on the interval duration corresponding to the longest consecutive hit interval. Based on the total hit duration and the total video duration of the video to be detected, the relative concentration ratio corresponding to the target parent video is determined. Based on the target attribution matrix, the target parent video is determined to have a target parent slice hit set. Then, based on the time axis of the target parent video, the time intervals of each target parent slice in the target parent slice hit set are deduplicated and merged to obtain the corresponding second hit interval. The target coverage score of the target parent video is determined based on the target salience weight corresponding to each of the second hit intervals and the target parent slice, and the core plot coverage rate of the target parent video is determined based on the ratio of the target coverage score to the full coverage score of the target parent video; wherein, the full coverage score is the score determined based on the duration of each parent slice corresponding to the target parent video and the corresponding target salience weight.

[0008] Optionally, determining the target coverage intensity coefficient corresponding to the target source video based on continuous citation duration, relative concentration ratio, and core plot coverage includes: The continuous reference duration is normalized based on a preset baseline duration to obtain a corresponding normalized reference duration. The normalized reference duration, the relative concentration ratio, and the core plot coverage are weighted and summed based on a second preset weighting strategy to determine the target coverage intensity coefficient corresponding to the target parent video.

[0009] Optionally, the method for detecting the same source in mixed-edit videos further includes: If the target parent feature vector corresponding to the slice to be detected does not exist in the target parent feature library, or the target similarity between the target parent feature vector and the target feature vector does not meet the preset similarity condition, then the slice to be detected is determined as an unattributed slice, and the proportion of proprietary content corresponding to the video to be detected is determined based on the unattributed slice.

[0010] Secondly, this application discloses a device for detecting the same source in mixed-edit videos, comprising: The detection feature extraction module is used to perform temporal slicing of the video to be detected and extract the target feature vector corresponding to each slice. The hit relationship construction module is used to determine the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, and to construct the target hit relationship between the slice to be detected and the target parent slice; The attribution matrix determination module is used to group the target hit relationships based on the parent identifier corresponding to the target parent slice, and construct a target attribution matrix between the video to be detected and each target parent video based on the target grouping results; different target attribution matrices are independent of each other; the target attribution matrix is ​​used to record the time interval of the video to be detected, the time interval of the target parent slice, and the target saliency weight corresponding to the target parent video attributed to the corresponding target parent video; the target saliency weight characterizes the degree to which the target parent slice carries the core plot content of the corresponding parent video; The detection result acquisition module is used to determine the continuous reference duration, relative concentration ratio, and core plot coverage of each target parent video in the video to be detected based on the target attribution matrix, and to determine the target coverage intensity coefficient corresponding to the target parent video based on the continuous reference duration, relative concentration ratio, and core plot coverage, and to generate a homogeneous detection result evidence package based on the target coverage intensity coefficient; wherein, the relative concentration ratio represents the proportion of the video content of the target parent video in the video to be detected, and the core plot coverage represents the degree to which the video to be detected covers the target core plot of the target parent video.

[0011] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned method for detecting the same source in mixed-edit videos.

[0012] Fourthly, this application discloses a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned method for detecting the same source of mixed-edit videos.

[0013] In this application, when performing source detection on a mixed video, the video to be detected is divided into temporal slices, and the target feature vectors corresponding to each slice are extracted; the target parent slices are determined based on the target similarity between each parent feature vector in the target parent feature library and the target feature vectors, and a target hit relationship is constructed between the slice to be detected and the target parent slices; the target hit relationship is grouped based on the parent identifier corresponding to the target parent slice, and a target attribution matrix is ​​constructed between the video to be detected and each target parent video based on the target grouping results; different target attribution matrices are independent of each other; the target attribution matrix is ​​used to record the time interval of the slice to be detected attributed to the corresponding target parent video, the target parent slice's... The time interval and target saliency weights corresponding to the target parent video slices are defined. The target saliency weights characterize the degree to which the target parent video slices carry the core plot content of the corresponding parent video. Based on the target attribution matrix, the continuous reference duration, relative concentration ratio, and core plot coverage of each target parent video in the video to be detected are determined. The target coverage intensity coefficient corresponding to the target parent video is determined based on the continuous reference duration, relative concentration ratio, and core plot coverage, and a homology detection result evidence package is generated based on the target coverage intensity coefficient. The relative concentration ratio characterizes the proportion of the video content of the target parent video in the video to be detected, and the core plot coverage represents the degree to which the video to be detected covers the target core plot of the target parent video. As can be seen, when performing source detection on multi-source mixed videos, this application first performs temporal slicing and feature extraction on the video to be detected, determines the target parent slices based on the target parent feature library, and constructs target hit relationships; then, it groups the target hit relationships based on the parent identifier, constructing independent target attribution matrices for each target parent video; subsequently, it jointly calculates three indicators—continuous citation duration, relative concentration ratio, and core plot coverage rate—based on the target attribution matrices, weights them to obtain the target coverage intensity coefficient, and generates a source detection result evidence package when preset output conditions are met. Thus, this application provides structured modeling of the temporal attribution relationships of multi-source mixed videos at the shot segment granularity, and the output detection results simultaneously present multi-dimensional temporal mapping information of absolute duration and relative weight, providing a structured and quantitative basis for manual review. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0015] Figure 1This is a flowchart of a method for detecting the same source in mixed-edit videos disclosed in this application; Figure 2 This is a schematic diagram of a specific source detection system architecture for mixed-edit videos disclosed in this application; Figure 3 This is a schematic diagram of a specific attribution matrix construction process disclosed in this application; Figure 4 This is a schematic diagram of a specific process for determining the homogeneous coverage strength coefficient disclosed in this application; Figure 5 This is a schematic diagram of the structure of a homogeneous video detection device disclosed in this application; Figure 6 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] In the short video and self-media content ecosystem, many mashup videos densely splice together clips from multiple copyrighted film and television works, deliberately controlling the continuous usage duration of individual copyrighted materials, or distributing them throughout the video to circumvent existing detection mechanisms based on duration thresholds or hit frequency. Existing detection schemes struggle to effectively quantify the usage ratio of each source material in these scenarios, and cannot distinguish whether the referenced content is concentrated in the core, high-value segments of the original film, leading to missed detections. To address these technical issues, this application discloses a method for detecting the same source in mashup videos, enabling structured same-source detection in multi-source fragmented mashup scenarios.

[0018] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for detecting the same source in mixed-edit videos, including: Step S11: Perform temporal slicing on the video to be detected and extract the target feature vector corresponding to each slice.

[0019] In this embodiment, a mashup video refers to a fragmented derivative video that extracts and splices together multiple copyrighted video materials from different sources, commonly found in film and television commentary, compilations, and highlight reels. The source video library refers to a set of copyrighted benchmark videos that have completed copyright registration or authorization filing and feature entry, serving as a reference for source comparison. The plot salience score is the salience weight calculated by comprehensively considering visual, acoustic, and textual three-dimensional signals for each temporal segment of the source video, used to identify high-value plot segments. The core plot coverage ratio is the proportion of content extracted from the mashup video to the total core plot of the original source video, based on a weighted salience integral. The attribution matrix maps temporal slices of the mashup video to a hit table of candidate source videos, supporting parallel and independent calculation across multiple source videos. The homology coverage strength... Intensity refers to a comprehensive risk coefficient calculated by weighting three indicators: continuous citation duration, relative concentration ratio of the entire film, and coverage of core plot points. Local Sensitive Hash (LSH) is a coarse screening algorithm used for high-dimensional vector approximate nearest neighbor retrieval, which can quickly recall candidate segments in a large-scale parent feature library. Evidence Package refers to the structured comparison results finally output by the system, including the parent source matrix, detailed distribution of proportions, and a time-series mapping table for manual review. It should also be noted that the specific values ​​appearing in this embodiment, such as slice duration, recall quantity, similarity threshold, confidence difference, time tolerance, normalized baseline duration, and various weight values, are all exemplary parameters and are only used to help understand the technical solution of this embodiment. Those skilled in the art can make adaptive adjustments to the above parameters according to actual application scenarios.

[0020] In one specific implementation, the video to be detected can be a mixed video from a short video platform or self-media platform that needs to be detected for source homology detection. When performing temporal slicing on the mixed video to be detected, an adaptive temporal slicing can be performed using a shot boundary detection algorithm, that is, using the shot switching point as the slice boundary, so that each slice to be detected corresponds to a relatively complete shot unit; alternatively, a fixed time window method can be used for slicing. After obtaining each slice to be detected, a multimodal feature vector consistent with the source code is extracted for each slice, i.e., the target feature vector; the multimodal feature vector can include visual dimension features, acoustic dimension features, and text dimension features to ensure that the target code and the source code are compared in the same feature space.

[0021] Step S12: Determine the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, and construct the target hit relationship between the slice to be detected and the target parent slice.

[0022] In this embodiment, as Figure 2 As shown, from a system architecture perspective, the technical solution of this embodiment can be composed of four processing modules: the input side, the detection side, the comparison side, and the output side. The input side is used for multimodal saliency calculation and feature indexing of the parent video; the detection side is used for deconstructing the mixed video and generating the attribution matrix; the comparison side is used for calculating the three-dimensional homogeneous coverage intensity; and the output side is used for generating structured evidence packages.

[0023] In this embodiment, the target master feature library stores the master feature vectors of the master segments corresponding to each master video. The target salience weights corresponding to the master feature vectors are used to characterize the degree to which the master segment carries the core plot content of the corresponding master video. That is, each master segment record in the target master feature library, in addition to the feature vector itself, is also associated with a pre-calculated core plot salience. The higher the salience, the more likely the master segment is to carry the high-value core plot content of the master video (such as action climax, plot twist, etc.). The lower the salience, the more likely the master segment is to be low-value content such as transitional scenes, black screen frames, opening and closing credits.

[0024] In this embodiment, before determining the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, the method further includes: acquiring each parent video, performing time-series slicing on the parent video to obtain the corresponding parent slices, and extracting features from the parent slices to obtain the parent feature vectors corresponding to the parent slices; the parent feature vectors include visual dimension feature vectors, acoustic dimension feature vectors, and text dimension feature vectors; performing weighted fusion of the visual dimension feature vectors, acoustic dimension feature vectors, and text dimension feature vectors based on a first preset weighting strategy, so as to determine the target saliency weights corresponding to the parent slices based on the parent feature vectors, and writing the parent feature vectors corresponding to each parent video into the target parent feature library.

[0025] In one specific implementation, the copyrighted master video can be sequentially sliced ​​according to a fixed time window, with an exemplary parameter of 5 seconds, meaning each 5-second segment is a master video slice. For each master video slice, feature signals are extracted from three dimensions: visual, acoustic, and textual, to comprehensively calculate the core plot saliency score. In the visual dimension, the rate of change of optical flow between video frames is extracted to quantify the intensity of camera motion; keypoint pose estimation algorithms are used to assess the amplitude of character movements; and the scene change rate per unit time is statistically analyzed. High motion amplitude and high change rate are usually associated with intense action scenes; therefore, the aforementioned visual signals can be used to identify action-intensive segments in the master video. In the acoustic dimension, short-time energy and spectral fluid features are extracted; points of dramatic change in speech intensity are detected to capture climactic moments such as collisions, outbursts, or high-decibel emotional releases. The aforementioned acoustic signals can be used to identify climax segments in the master video at the audio level. At the text level, the Natural Language Processing (NLP) module is invoked to process the Automatic Speech Recognition (ASR) and Optical Character Recognition (OCR) results, calculating the word frequency of terms in a pre-defined conflict dictionary, such as "counterattack," "betrayal," and "final battle." Emotional state transition points are detected using a sequence labeling model, and keyword tags with plot-specific implications are identified. These text signals can be used to semantically identify segments in the source video that carry plot conflicts and transitions.

[0026] Then, the visual, acoustic, and textual feature signals can be normalized to the [0,1] interval, and then weighted and fused using configurable dynamic weights (i.e., the first preset weighting strategy). The final core plot saliency_score of the parent slice is output as the target saliency weight for the parent slice. To reduce false positive interference, the following masking rules can be introduced for noise filtering: filtering the opening and closing segments based on their temporal positions to avoid fixed content such as opening theme songs and closing subtitles being misjudged as high-saliency segments; filtering high-volume pure noise segments through voiceprint detection to avoid simply raising acoustic saliency through high-decibel noise; and using text area masks to exclude visual abrupt changes caused by large-area sudden subtitle effects. The normalized target saliency weight will be used as the integral weight in the subsequent calculation of core plot coverage, so that high-value segments of the original film can obtain a higher impact factor in the coverage integral, thereby achieving differentiated modeling of the carrying capacity of the core plot content of the parent film.

[0027] For example, the parent work segment feature data structure may include the following fields (the values ​​are all exemplary parameters): parent work identifier source_work_id, for example, BASE_MOV_007; segment set, where each segment record includes: time interval time_window, for example, 00:45:10-00:45:15; saliency weight saliency_score, for example, 0.85; saliency type saliency_type, for example, ACTION_CLIMAX; visual feature vector visual_embedding, for example, [0.15, -0.09, …, 0.54].

[0028] In terms of vector index management, the multimodal feature vectors of each source video slice are written into the target source video feature library (i.e., the source video feature index library). This feature library supports incremental addition and version updates of copyrighted materials. That is, when a new copyrighted source video is added to the library or existing source video content is revised, only the newly added or changed source video slices need to be feature extracted and the index updated, avoiding a full library recalculation when source video content is revised. Accordingly, each source video slice record can include at least the following fields: source video identifier, source video time interval, feature vector, core plot saliency, and version identifier.

[0029] In this embodiment, the target parent slice is determined based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector. This includes: using a locality-sensitive hashing algorithm, determining the candidate parent feature vector corresponding to the slice to be detected based on the Euclidean distance between the target feature vector and each parent feature vector in the target parent feature library, and determining the target parent feature vector based on the target similarity between the candidate parent feature vector and the target feature vector, thus determining the parent slice corresponding to the target parent feature vector as the target parent slice. It should be noted that the locality-sensitive hashing in this application is not only used for conventional candidate recall, but also serves as a pre-indexing mechanism for attribution matrix construction. Based on this, the hit relationship acquisition process in this embodiment may specifically include the following steps: Step 1: Perform temporal slicing on the master video and extract multimodal features such as visual, acoustic, and textual features from each master slice to form a master slice feature vector; each master slice record includes at least the following fields: master identifier, master time interval, feature vector, core plot saliency, and version identifier.

[0030] Step 2: Perform local sensitive hashing on the feature vector of the parent slice to map the high-dimensional features to one or more hash buckets and establish a "hash bucket - parent slice" index. This index is used to quickly recall candidate parent slices in a large-scale parent database, avoiding the need to sort all slices in the database one by one, thereby reducing the retrieval computation overhead in the scenario of massive parent database.

[0031] Step 3: Perform shot boundary detection or fixed time window slicing on the mixed video to be detected, and extract the multimodal feature vector, i.e. target feature vector, for each slice to be detected, which is consistent with the original data.

[0032] Step 4: Input the target feature vector of the slice to be detected into the Local Sensitive Hash Index (LSI) to recall the set of candidate parent slices that fall into the same or similar hash buckets. That is, based on the Euclidean distance between the target feature vector and each parent feature vector, recall the top K candidate parent slices with the closest distance. An example parameter is Top 50. It should be noted that this set is only used as a candidate set and is not directly used as the final hit result.

[0033] Step 5: Perform fine-tuning confirmation and threshold truncation on candidate parent slices. Fine-tuning can use cosine similarity, Euclidean distance, weighted multimodal similarity, or other similarity calculation methods to determine the target similarity. Only when the target similarity meets the preset similarity conditions, such as calculating the cosine similarity between the target feature vector of the slice to be detected and the feature vector of the candidate parent slice, and the similarity reaches the preset fine-tuning threshold (example parameter: 0.85), is the slice to be detected and the corresponding candidate parent slice recorded as a valid hit mapping, that is, the candidate parent slice is determined as the target parent slice, and the target hit relationship is constructed; slices to be detected that do not meet the similarity standard are marked as unattributed.

[0034] Step 6: For cases where the same target slice simultaneously hits multiple parent slices (i.e., multiple parent slice conflicts), the system performs conflict decision processing based on factors such as confidence level, similarity difference, modal consistency, and parent slice temporal continuity. Specifically, if the same target slice corresponds to at least two target parent slices, for example, if the same target slice simultaneously hits different segments of parent slice A and parent slice B, the system compares their confidence levels and determines the target parent slice with the highest target similarity as the unique parent slice corresponding to the target slice, prioritizing the retention of mapping records with higher confidence levels; correspondingly, a target hit relationship is constructed between the target slice and the unique parent slice. Furthermore, if the difference between the highest confidence result and the second highest confidence result is extremely small (example parameter: less than 0.02), bidirectional mapping records can be retained simultaneously and submitted to subsequent noise reduction processing, i.e., multiple candidate mappings are retained and participate in the coverage intensity evaluation of the corresponding parent slices in subsequent calculations.

[0035] Step S13: Based on the parent identifier corresponding to the target parent slice, the target hit relationship is grouped, and a target attribution matrix is ​​constructed between the video to be detected and each target parent video based on the target grouping results; different target attribution matrices are independent of each other; the target attribution matrix is ​​used to record the time interval of the video to be detected, the time interval of the target parent slice, and the target saliency weight corresponding to the target parent video attributed to the corresponding target parent video; the target saliency weight characterizes the degree to which the target parent slice carries the core plot content of the corresponding parent video.

[0036] In this embodiment, the target parent video is the parent video to which the target parent slice belongs, and the target attribution matrices corresponding to different target parent videos are independent of each other; the target attribution matrix is ​​used to record the time interval of the detected slice attributed to the corresponding target parent video, the time interval of the target parent slice, and the target saliency weight corresponding to the target parent slice. See also Figure 3 As shown, a temporal slice of a mixed video can be mapped to three categories: parent A, parent B, and unattributed, thus forming a multi-parent parallel attribution matrix structure. Slice records attributed to parent A are merged into the target attribution matrix corresponding to parent A, and slice records attributed to parent B are merged into the target attribution matrix corresponding to parent B. The two do not interfere with each other, and the subsequent coverage coefficient calculations of each parent are independent of each other.

[0037] In this embodiment, after obtaining the target hit relationship, the detection side uses a locality-sensitive hashing-based attribution matrix generation mechanism, which specifically includes the following steps: generating an attribution matrix based on the aforementioned effective hit mapping. It should be noted that the attribution matrix is ​​not a simple hit count count, but a structured record with the segment to be detected as rows and candidate parent videos and their corresponding time intervals as attribution objects. It records fields such as the time interval to be detected, parent video identifier, parent video time interval, similarity or confidence score, saliency weight, and hit status. Exemplary attribution matrix fields include: segment identifier test_segment_id; video time interval test_time_window; candidate parent video identifier source_work_id; parent video time interval source_time_window; matching confidence score between the segment to be detected and the parent video segment confidence_score; core plot saliency score corresponding to the parent video segment saliency_score; and attribution status attribute_status, such as attributed, not attributed, multiple candidates pending review, etc. Then, the system groups the attribution matrices according to the parent video identifier (i.e., multi-source independent aggregation), generating an independent attribution factor matrix for each candidate parent video, which is the target attribution matrix corresponding to each target parent video. Subsequent continuous citation duration, relative concentration percentage of the whole video, and core plot coverage are all calculated independently based on the target attribution matrix of the corresponding parent video. The coverage coefficient calculations between parent videos do not interfere with each other, avoiding the superposition or dilution of hit results from multiple parent videos.

[0038] For example, the attribution record data structure for the slice to be detected may include the following fields (all values ​​are exemplary parameters): the slice identifier test_segment_id, for example, seg_005; the time interval to be detected test_time, for example, 00:15-00:20; the list of matched sources matched_sources, where each matched source record includes: the source identifier source_id, for example, BASE_MOV_007; the source time interval source_time, for example, 00:45:10-00:45:15; the confidence score, for example, 0.92; the saliency score, for example, 0.85; in addition, the attribution record also includes the attribution status status, for example, ATTRIBUTED.

[0039] As can be seen from the above processing, this embodiment differs from the scheme that uses the number of hits per unit time as the core indicator: this embodiment does not directly use the number of hits as the basis for risk judgment, but first recalls candidates through local sensitive hashing, then generates a structured attribution matrix through fine ranking confirmation, conflict handling and parent book grouping, and then calculates the coverage strength of different parent books based on the attribution matrix.

[0040] Step S14: Based on the target attribution matrix, determine the continuous reference duration, relative concentration ratio, and core plot coverage of each target parent video in the video to be detected. Determine the target coverage intensity coefficient corresponding to the target parent video based on the continuous reference duration, relative concentration ratio, and core plot coverage, and generate a homology detection result evidence package based on the target coverage intensity coefficient. The relative concentration ratio represents the proportion of the video content of the target parent video in the video to be detected, and the core plot coverage represents the degree to which the video to be detected covers the target core plot of the target parent video.

[0041] In this embodiment, the relative concentration ratio is used to characterize the proportion of the target parent video's video content in the video to be detected, and the core plot coverage rate is used to characterize the degree to which the video to be detected covers the target parent video's target core plot. Preset output conditions can be, for example, that the target coverage intensity coefficient reaches a preset output threshold, or that the target coverage intensity coefficient ranks high among all candidate parent videos; this embodiment does not limit this. In this embodiment, such as... Figure 4 As shown, the continuous reference duration, relative concentration ratio, and core plot coverage of each target parent video in the video to be detected are determined based on the target attribution matrix. Specifically, this can be achieved based on a dual-track deduplication mechanism on the comparison side. Dual-track deduplication refers to performing deduplication and merging on the timeline of the video to be detected and the timeline of the parent video separately. The calculation objectives of the two tracks are different, and the problems they solve are also different. The processing procedures for the two tracks are explained below.

[0042] Track 1: Deduplication based on the timeline of the video to be detected. This track is used to calculate the actual proportion of material from a target parent video in the montage video to be detected, as well as the longest continuous reference duration. Specifically, it includes: determining the hit set of the target segments corresponding to each target parent video based on the target attribution matrix, and deduplicating and merging the time intervals of each target segment in the hit set based on the timeline of the video to be detected to determine the first hit interval corresponding to each target parent video; the first hit interval is a continuous interval; determining the longest continuous hit interval and total hit duration corresponding to the target parent video based on the interval duration corresponding to each first hit interval, determining the continuous reference duration corresponding to the target parent video based on the interval duration corresponding to the longest continuous hit interval, and determining the relative concentration ratio corresponding to the target parent video based on the total hit duration and the total video duration of the video to be detected.

[0043] In one specific implementation, all hit records attributing to a target parent video are extracted from the target attribution matrix corresponding to that target parent video. These hit records are then sorted in ascending order of the start time of the time interval of the video to be detected. Time intervals that overlap or are adjacent on the time axis to be detected with a distance less than a preset tolerance are merged (exemplary parameter: a tolerance of 2 seconds for a time window considered continuous), forming a set of deduplicated hit intervals on the video to be detected side, i.e., the first hit interval corresponding to each target parent video. The first hit interval is a continuous interval. Based on the interval duration corresponding to each merged first hit interval, the total hit duration (i.e., total mapping duration) of the target parent video in the video to be detected is calculated. Furthermore, the ratio of the total hit duration to the total duration of the video to be detected is calculated to obtain the relative concentration percentage (relative concentration percentage of the entire video) corresponding to the target parent video. In the merged set of first hit intervals, the first hit interval with the longest interval duration is determined as the longest continuous hit interval corresponding to the target parent video, and the continuous reference duration corresponding to the target parent video is determined based on the interval duration corresponding to the longest continuous hit interval. The purpose of this track is to prevent the reference ratio of the target side from being calculated repeatedly due to multiple hits, duplicate candidates, or multiple modal scoring during the same time period to be detected, thereby ensuring that the relative concentration ratio and continuous reference duration reflect the true distribution of the parent material on the time axis of the video to be detected.

[0044] The second track: Deduplication based on the parent video timeline. This track is used to calculate which content in the original target parent video is actually covered by the mashup video to be detected, especially for calculating the core plot coverage rate. Specifically, it includes: determining the target parent video's corresponding target parent slice hit set based on the target attribution matrix, and deduplicating and merging the time intervals of each target parent slice in the target parent slice hit set based on the target parent video's timeline to obtain the corresponding second hit interval; determining the target coverage integral corresponding to the target parent video based on each second hit interval and the target saliency weight corresponding to the target parent slice, and determining the core plot coverage rate corresponding to the target parent video based on the ratio of the target coverage integral to the total coverage integral corresponding to the target parent video; wherein, the total coverage integral is the integral determined based on the duration of each parent slice corresponding to the target parent video and the corresponding target saliency weight.

[0045] In one specific implementation, all hit records attributing to a target parent video are extracted from the target attribution matrix corresponding to that target parent video. These hit records are then sorted in ascending order of the start time of the parent video's time interval. Time intervals that overlap or are adjacent on the parent timeline with a distance less than a preset tolerance are merged to form a set of deduplicated coverage intervals on the parent side, i.e., the corresponding second hit intervals. During the merging process, for intervals with overlap, the saliency weights can be weighted and merged according to the newly added duration: the saliency weight of the merged interval is equal to the product of the original interval duration and the original interval saliency weight, plus the sum of the product of the newly added non-overlapping duration and the newly merged interval saliency weight, and then divided by the total duration after merging. This ensures that the saliency score is not repeatedly accumulated due to interval overlap. For each merged second hit interval, the target saliency weight of the corresponding target parent slice is read, and the saliency-weighted coverage score is calculated according to the coverage duration and saliency weight, i.e., the target coverage score corresponding to that target parent video. The target coverage integral is compared with the full coverage integral corresponding to the target source video. The core plot coverage rate of the target source video is determined based on the ratio of the target coverage integral to the full coverage integral. The full coverage integral is determined based on the duration of each source video segment and its corresponding target saliency weight, i.e., the sum of the saliency integrals of all segments in the source video. This track prevents the core plot coverage rate from being repeatedly inflated due to the repeated looping of the same source video segment. In other words, even if the same source video climax segment is used repeatedly, after deduplication along the source timeline, the segment is only counted once in the core plot coverage rate calculation based on its actual coverage within the source video, or only as the newly added coverage portion.

[0046] For example, the core plot coverage C_core can be calculated using the following formula: ; in, Let k be the coverage duration of the second hit interval. The significance weight is the value corresponding to the k-th second hit interval. This represents the summation over all second-hit intervals. The full coverage integral corresponding to the target parent video (i.e., the sum of the saliency integrals of all slices of the original parent video) is given. This means that the calculation result will be truncated to no more than 1.0.

[0047] In one specific implementation, the target coverage intensity coefficient corresponding to the target parent video is determined based on the continuous citation duration, relative concentration ratio, and core plot coverage rate. Specifically, this may include: normalizing the continuous citation duration based on a preset baseline duration to obtain the corresponding normalized citation duration, and weighting and summing the normalized citation duration, relative concentration ratio, and core plot coverage rate based on a second preset weighting strategy to determine the target coverage intensity coefficient corresponding to the target parent video.

[0048] For example, a normalized continuous duration reference base can be set (example parameter: preset base duration is 60 seconds), then the normalized reference duration is: ; in This refers to the duration of continuous citation. Correspondingly, the target coverage intensity coefficient I can be calculated using the following formula: ; in, This represents the relative concentration percentage. For the coverage of core plot points, , , Configurable weights (i.e., the second preset weighting strategy, example parameter: continuous duration weights) The concentration weight is 0.2. The weight of the coverage rate of the core plot is 0.3. (0.5) This means truncating the weighted sum to the interval [0,1].

[0049] To facilitate understanding, the following is a simplified implementation of a complete example of the logic for calculating the 3D homogeneous coverage intensity of the comparison side (all values ​​are exemplary parameters). First, the configuration parameters are read: the continuous duration weight w_continuous is 0.2, the concentration percentage weight w_concentration is 0.3, the core plot coverage weight w_core_plot is 0.5, the normalized continuous duration reference base_time_thresh is 60.0 seconds, and the tolerance for the continuous time window is 2.0 seconds. The calculation process takes the hit slice set segments_mapped attributed to a certain parent video, the total duration of the video to be detected test_total_time, and the saliency integral of the entire parent video source_total_saliency as input.

[0050] Track 1 Processing: Sort `segments_mapped` by the start time of the segment to be detected and iterate through each segment. If the start time of the current segment is not greater than the sum of the end time of the merged interval and the tolerance, then merge the current segment into the merged interval and update the end time of the interval; otherwise, create a new merged interval. After the iteration is complete, sum the durations of each merged interval to obtain the total mapping duration `total_mapped_time`, and take the maximum value of the durations of each merged interval to obtain the longest continuous duration `max_continuous`. Track 2 Processing: Re-sort `segments_mapped` by the start time of the parent dataset and iterate through each segment. Merge intervals on the parent timeline according to the same tolerance rules. When merging overlapping intervals, use the aforementioned newly added duration weighting method to fuse the saliency weights. After the iteration is complete, sum the products of the coverage duration and the saliency weights of each merged interval to obtain the saliency weighted hit integral `weighted_hit`.

[0051] Subsequently, three normalized sub-indices are calculated: the continuous duration sub-indicator `score_cont = min(1.0, max_continuous / base_time_thresh)`; the concentration ratio sub-indicator `score_conc = min(1.0, total_mapped_time / test_total_time)`; and the core plot coverage sub-indicator `score_core = min(1.0, weighted_hit / source_total_saliency)`. Finally, the homogeneous coverage intensity `intensity = min(1.0, score_cont × w_continuous + score_conc × w_concentration + score_core × w_core_plot)` is calculated, and detailed breakdowns of each sub-indicator are output simultaneously, including the longest continuous hit duration `max_continuous_sec`, the concentration ratio `concentration_ratio`, and the core plot coverage `core_plot_coverage`, for use in subsequent evidence package assembly.

[0052] In this embodiment, regarding anomaly handling, if the cause is attributed to an empty set of hit slices for a target parent video, a total duration of the video to be detected being zero, or a saliency integral of the entire parent video being zero, the system records the anomaly (e.g., returns the error flag ERR_INVALID_INPUT), outputs a coverage strength of 0 for the target parent video, and does not interrupt the parallel computation of other target parent videos, so as to ensure the overall robustness of the detection process in multi-parent scenarios.

[0053] Specifically, this embodiment may further include: if there is no candidate parent feature vector corresponding to the slice to be detected in the target parent feature library, or if the target similarity between the candidate parent feature vector and the target feature vector does not meet the preset similarity condition, then the slice to be detected is determined as an unattributed slice, and the proportion of proprietary content corresponding to the video to be detected is determined based on the unattributed slice. Specifically, the proportion of the sum of the durations of all unattributed slices to the total duration of the video to be detected, or the proportion of the number of unattributed slices to the total number of slices to be detected, can be used as the proportion of proprietary content; the proportion of unattributed categories is used to quantify the proportion of proprietary content in the video to be detected, providing auxiliary reference information for manual review.

[0054] It is understood that the dual-track deduplication in this embodiment serves different indicators: the deduplication of the time axis to be detected is used to calculate the relative concentration ratio of the entire video and the continuous citation duration; the deduplication of the parent video time axis is used to calculate the core plot coverage rate; the three indicators are normalized and then fused according to configurable weights to obtain the homogeneous coverage strength coefficient. Therefore, the coverage strength of this application is not simply determined by the number of hits, the duration of a single segment, or the total citation duration, but rather by comprehensively considering the following three aspects: whether the parent video material appears in a concentrated manner in the video to be detected; whether the parent video material accounts for a high proportion in the video to be detected; and whether the cited content covers the highly significant core plot area in the parent video.

[0055] Understandably, regarding anti-circumvention mechanisms, if the content processor repeatedly plays segments from the video to be detected, the timeline of the video to be detected is lengthened, but the parent timeline does not advance. This application addresses this situation through a dual-track deduplication mechanism: deduplication is performed based on the time of the video to be detected when calculating the concentration percentage; deduplication is forcibly performed based on the parent timeline when calculating the coverage rate of the core plot, and the significance is weighted by the newly added duration, making it difficult for strategies to inflate coverage rates by repeatedly playing the same climax segment to be effective. Combined with the upper limit truncation design (min(1.0, value)) for each sub-item, it helps to improve the objective reflection of the coverage coefficient on the actual citation ratio.

[0056] On the output side, when the target coverage intensity coefficient of any target parent video meets the preset output conditions, the system generates a homology detection result evidence package corresponding to the video to be detected and the target parent video. The evidence package is structured data, and its contents include: the coverage intensity coefficient and three-dimensional sub-item details of each target parent video; a slice-level mapping table between the time axis of the video to be detected and the time axis of each target parent video; the proportion of unattributed segments; and manual review suggestions.

[0057] For example, the evidence package data structure may include the following fields (all values ​​are exemplary parameters): report_id (e.g., MASHUP_2026_112); candidate_work_id (e.g., TEST_MIX_005); traced_source_works (a list of source works), where each traced record includes: source_work_id (e.g., BASE_MOV_007); intensity_score (e.g., 0.88); and breakdown details, including the longest consecutive hit duration max_c. The evidence package includes `ontinuous_sec` (e.g., 45 seconds), `concentration_ratio` (e.g., 0.65), and `core_plot_coverage` (e.g., 0.92); a slice-level timeline mapping table `mapped_timeline`, where each mapping record includes the test time interval `test_time` (e.g., 01:10-01:15), the source time interval `source_time` (e.g., 00:45:10-00:45:15), and the saliency type `saliency_type` (e.g., `ACTION_CLIMAX`). Additionally, the evidence package includes `unattributed_ratio` (e.g., 0.15), and a manual review suggestion, such as "Suggestion to enter the manual review queue: A high proportion of concentrated source core plot features was detected; it is recommended to verify the content reuse ratio, contextual purpose, and source correspondence of the target segment." By outputting the correspondence and significance type between the time axis to be detected and the time axes of each parent sample at the slice granularity, manual reviewers can quickly locate the target segment and its contextual correspondence, thereby improving review efficiency and the interpretability of the conclusions.

[0058] To facilitate understanding of the differences in performance between this embodiment and existing solutions, the following descriptions are based on typical scenarios (all values ​​are exemplary parameters). Scenario 1 is a five-minute highlight condensation scenario, where the video to be detected condenses the core highlight content of the source material into a short duration for intensive citation: Under existing hit frequency schemes, due to the short total duration, false negatives are easily generated; however, under the three-dimensional evaluation mechanism of this embodiment, the core plot coverage of such videos is high, and the comprehensive coverage intensity coefficient can reach a high level, thereby triggering manual review recommendations. Scenario 2 is a multi-source film and television review scenario, where the video to be detected interweaves and cites fragments from multiple source works: Existing solutions cannot distinguish the sources, making judgment difficult; this embodiment generates independent attribution matrices for each source, quantifying their respective proportions. Scenario 3 is a low-frequency, scattered fragment citation scenario: Existing solutions have few hits, easily leading to false positives or false negatives; in this embodiment, such videos have low concentration and low core plot coverage, resulting in a low comprehensive coefficient, and can be classified into a low-priority queue, thereby reasonably allocating review resources.

[0059] It should be noted that, without departing from the concept of this application, some processing modules in the embodiments of this application can also adopt alternative implementation methods to achieve the same inventive purpose. Several exemplary alternative implementation methods are given below.

[0060] Alternative implementation of the saliency calculation module: In addition to the combination of visual optical flow, acoustic short-time energy and text conflict dictionary, it can also be replaced by calling the Visual-Language Large Model (VLM) to perform plot summary description on the video frame sequence and directly extract high-level saliency labels from the semantic level. This alternative method is suitable for scenarios with higher requirements for semantic understanding accuracy.

[0061] Alternative implementation of time-series slice aggregation: The method of merging adjacent slices with a fixed tolerance during the attribution matrix generation stage can be replaced by a time-axis dynamic clustering scheme based on density clustering (such as the DBSCAN algorithm). This alternative is suitable for scenarios with irregular slice interval distribution.

[0062] An alternative implementation of the coverage strength calculation model: The three-dimensional index weighting formula can be replaced by a machine learning binary classification model trained with labeled samples. The optimal mapping relationship of the three indicators is learned adaptively through data. This alternative method is suitable for scenarios with sufficient labeled samples.

[0063] Alternative implementations of feature retrieval indexes: The Local Sensitive Hash (LSH) coarse screening scheme can be replaced by a graph-based Approximate Nearest Neighbor (HNSW) index, offering different trade-offs between recall and retrieval speed.

[0064] Alternative implementation of the output format: The structured evidence package can be replaced by a timeline visualization chart (such as a percentage distribution chart in the form of a Gantt chart) as the main presentation method. This alternative is suitable for the visualization display needs of the manual review interface.

[0065] As can be seen, when performing source detection on multi-source mixed videos, this application first performs temporal slicing and feature extraction on the video to be detected, determines the target parent slices based on the target parent feature library, and constructs target hit relationships; then, it groups the target hit relationships based on the parent identifier, constructing independent target attribution matrices for each target parent video; subsequently, it jointly calculates three indicators—continuous citation duration, relative concentration ratio, and core plot coverage rate—based on the target attribution matrices, weights them to obtain the target coverage intensity coefficient, and generates a source detection result evidence package when preset output conditions are met. Thus, this application provides structured modeling of the temporal attribution relationships of multi-source mixed videos at the shot segment granularity, and the output detection results simultaneously present multi-dimensional temporal mapping information of absolute duration and relative weight, providing a structured and quantitative basis for manual review.

[0066] See Figure 5 As shown, this application discloses a device for detecting the same source in mixed-edit videos, comprising: The detection feature extraction module 11 is used to perform temporal slicing of the video to be detected and extract the target feature vector corresponding to each slice to be detected. The hit relationship construction module 12 is used to determine the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, and to construct the target hit relationship between the slice to be detected and the target parent slice; The attribution matrix determination module 13 is used to group the target hit relationships based on the parent identifier corresponding to the target parent slice, and construct a target attribution matrix between the video to be detected and each target parent video based on the target grouping results; different target attribution matrices are independent of each other; the target attribution matrix is ​​used to record the time interval of the video to be detected, the time interval of the target parent slice, and the target saliency weight corresponding to the target parent slice, which are attributed to the corresponding target parent video; the target saliency weight characterizes the degree to which the target parent slice carries the core plot content of the corresponding parent video; The detection result acquisition module 14 is used to determine the continuous reference duration, relative concentration ratio, and core plot coverage of each target parent video in the video to be detected based on the target attribution matrix, to determine the target coverage intensity coefficient corresponding to the target parent video based on the continuous reference duration, relative concentration ratio, and core plot coverage, and to generate a homogeneous detection result evidence package based on the target coverage intensity coefficient; wherein, the relative concentration ratio represents the proportion of the video content of the target parent video in the video to be detected, and the core plot coverage represents the degree to which the video to be detected covers the target core plot of the target parent video.

[0067] As can be seen, when performing source detection on multi-source mixed videos, this application first performs temporal slicing and feature extraction on the video to be detected, determines the target parent slices based on the target parent feature library, and constructs target hit relationships; then, it groups the target hit relationships based on the parent identifier, constructing independent target attribution matrices for each target parent video; subsequently, it jointly calculates three indicators—continuous citation duration, relative concentration ratio, and core plot coverage rate—based on the target attribution matrices, weights them to obtain the target coverage intensity coefficient, and generates a source detection result evidence package when preset output conditions are met. Thus, this application provides structured modeling of the temporal attribution relationships of multi-source mixed videos at the shot segment granularity, and the output detection results simultaneously present multi-dimensional temporal mapping information of absolute duration and relative weight, providing a structured and quantitative basis for manual review.

[0068] In one specific embodiment, the device may further include: The parent feature extraction module is used to acquire each parent video, perform time-series slicing on the parent video to obtain corresponding parent slices, and perform feature extraction on the parent slices to obtain the parent feature vectors corresponding to the parent slices; the parent feature vectors include visual dimension feature vectors, acoustic dimension feature vectors and text dimension feature vectors; The parent feature storage module is used to perform weighted fusion of the visual dimension feature vector, the acoustic dimension feature vector and the text dimension feature vector based on a first preset weighting strategy, so as to determine the target saliency weight corresponding to the parent slice based on the parent feature vector, and write the parent feature vector corresponding to each parent video into the target parent feature library.

[0069] In one specific implementation, the hit relationship construction module 12 may include: The candidate vector determination module is used to determine the candidate parent feature vector corresponding to the slice to be detected based on the Euclidean distance between the target feature vector and each parent feature vector in the target parent feature library using the locality-sensitive hash algorithm, and to determine the target parent feature vector based on the target similarity between the candidate parent feature vector and the target feature vector, and to determine the parent slice corresponding to the target parent feature vector as the target parent slice.

[0070] In one specific embodiment, the device may further include: The unique slice determination module is used to determine the target parent slice with the highest target similarity as the unique parent slice corresponding to the slice to be detected if the same slice to be detected corresponds to at least two target parent slices. Accordingly, the hit relationship construction module 12 may specifically include: The hit relationship construction unit is used to construct the target hit relationship between the slice to be detected and the unique parent slice.

[0071] In one specific implementation, the detection result acquisition module 14 may include: The first interval determination unit is used to determine the target slice hit set corresponding to each target parent video based on the target attribution matrix, and to deduplicate and merge the time intervals of each target slice in the target slice hit set based on the time axis of the target video to determine the first hit interval corresponding to each target parent video; the first hit interval is a continuous interval. The first interval analysis unit is used to determine the longest consecutive hit interval and total hit duration corresponding to the target parent video based on the interval duration corresponding to each first hit interval, and to determine the continuous reference duration corresponding to the target parent video based on the interval duration corresponding to the longest consecutive hit interval, and to determine the relative concentration ratio corresponding to the target parent video based on the total hit duration and the total video duration of the video to be detected; The second interval determination unit is used to determine the target parent video corresponding to each target parent video hit set based on the target attribution matrix, and to deduplicate and merge the time intervals of each target parent video in the target parent video hit set based on the time axis of the target parent video to obtain the corresponding second hit interval. The second interval analysis unit is used to determine the target coverage score of the target parent video based on the target salience weight corresponding to each of the second hit intervals and the target parent slice, and to determine the core plot coverage rate of the target parent video based on the ratio of the target coverage score to the full coverage score of the target parent video; wherein, the full coverage score is the score determined based on the duration of each parent slice corresponding to the target parent video and the corresponding target salience weight.

[0072] In one specific implementation, the detection result acquisition module 14 may include: The coverage intensity coefficient determination unit is used to normalize the continuous reference duration based on a preset reference duration to obtain the corresponding normalized reference duration, and to perform a weighted summation of the normalized reference duration, the relative concentration ratio, and the core plot coverage rate based on a second preset weighting strategy to determine the target coverage intensity coefficient corresponding to the target parent video.

[0073] In one specific embodiment, the device may further include: The unattributed slice analysis module is used to determine the slice to be detected as an unattributed slice if the target parent feature vector corresponding to the slice to be detected does not exist in the target parent feature library, or the target similarity between the target parent feature vector and the target feature vector does not meet the preset similarity conditions, and to determine the proportion of proprietary content corresponding to the video to be detected based on the unattributed slice.

[0074] Furthermore, embodiments of this application also disclose an electronic device, Figure 6 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0075] Figure 6 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the video homology detection method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be a computer.

[0076] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0077] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0078] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the video montage source detection method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0079] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for detecting the same source in mixed-edit videos. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0080] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0081] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0082] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0083] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0084] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for detecting the same source in mixed-edit videos, characterized in that, include: The video to be detected is divided into temporal slices, and the target feature vectors corresponding to each slice are extracted. The target parent slice is determined based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, and the target hit relationship between the slice to be detected and the target parent slice is constructed. Based on the parent identifier corresponding to the target parent slice, the target hit relationship is grouped, and a target attribution matrix is ​​constructed between the video to be detected and each target parent video based on the target grouping results; different target attribution matrices are independent of each other; the target attribution matrix is ​​used to record the time interval of the video to be detected, the time interval of the target parent slice, and the target saliency weight corresponding to the target parent slice, which are attributed to the corresponding target parent video. The target saliency weight represents the degree to which the target parent slice carries the core plot content of the corresponding parent video; Based on the target attribution matrix, the continuous reference duration, relative concentration ratio, and core plot coverage of each target parent video in the video to be detected are determined. The target coverage intensity coefficient corresponding to the target parent video is determined based on the continuous reference duration, relative concentration ratio, and core plot coverage, and a homology detection result evidence package is generated based on the target coverage intensity coefficient. The relative concentration ratio represents the proportion of the video content of the target parent video in the video to be detected, and the core plot coverage represents the degree to which the video to be detected covers the target core plot of the target parent video.

2. The method for detecting the same source in mixed-edit videos according to claim 1, characterized in that, Before determining the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, the method further includes: Each source video is acquired, and the source videos are time-series sliced ​​to obtain corresponding source slices. Feature extraction is performed on the source slices to obtain source feature vectors corresponding to the source slices. The source feature vectors include visual dimension feature vectors, acoustic dimension feature vectors, and text dimension feature vectors. The visual dimension feature vector, the acoustic dimension feature vector, and the text dimension feature vector are weighted and fused based on the first preset weighting strategy, so as to determine the target saliency weight corresponding to the parent slice based on the parent feature vector, and the parent feature vector corresponding to each parent video is written into the target parent feature library.

3. The method for detecting the same source in mixed-edit videos according to claim 1, characterized in that, The method of determining the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector includes: Using the Locality Sensitive Hash algorithm, based on the Euclidean distance between the target feature vector and each parent feature vector in the target parent feature library, the candidate parent feature vector corresponding to the slice to be detected is determined, and the target parent feature vector is determined based on the target similarity between the candidate parent feature vector and the target feature vector. The parent slice corresponding to the target parent feature vector is then determined as the target parent slice.

4. The method for detecting the same source in mixed-edit videos according to claim 1, characterized in that, After determining the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, the method further includes: If the same slice to be detected corresponds to at least two target parent slices, then the target parent slice with the highest target similarity is determined as the unique parent slice corresponding to the slice to be detected. Accordingly, constructing the target hit relationship between the slice to be detected and the target parent slice includes: Construct the target hit relationship between the slice to be detected and the unique parent slice.

5. The method for detecting the same source in mixed-edit videos according to claim 1, characterized in that, The determination of the continuous reference duration, relative concentration ratio, and core plot coverage of each target parent video in the video to be detected based on the target attribution matrix includes: Based on the target attribution matrix, the target parent video is determined to have a set of hit segments to be detected. Then, based on the time axis of the target video, the time intervals of each target segment in the set of hit segments to be detected are deduplicated and merged to determine the first hit interval corresponding to each target parent video. The first hit interval is a continuous interval. Based on the interval duration corresponding to each of the first hit intervals, the longest consecutive hit interval and the total hit duration corresponding to the target parent video are determined, and the continuous reference duration corresponding to the target parent video is determined based on the interval duration corresponding to the longest consecutive hit interval. Based on the total hit duration and the total video duration of the video to be detected, the relative concentration ratio corresponding to the target parent video is determined. Based on the target attribution matrix, the target parent video is determined to have a target parent slice hit set. Then, based on the time axis of the target parent video, the time intervals of each target parent slice in the target parent slice hit set are deduplicated and merged to obtain the corresponding second hit interval. The target coverage score of the target parent video is determined based on the target salience weight corresponding to each of the second hit intervals and the target parent slice, and the core plot coverage rate of the target parent video is determined based on the ratio of the target coverage score to the full coverage score of the target parent video; wherein, the full coverage score is the score determined based on the duration of each parent slice corresponding to the target parent video and the corresponding target salience weight.

6. The method for detecting the same source in mixed-edit videos according to claim 1, characterized in that, The determination of the target coverage intensity coefficient corresponding to the target source video based on continuous citation duration, relative concentration ratio, and core plot coverage includes: The continuous reference duration is normalized based on a preset baseline duration to obtain a corresponding normalized reference duration. The normalized reference duration, the relative concentration ratio, and the core plot coverage are weighted and summed based on a second preset weighting strategy to determine the target coverage intensity coefficient corresponding to the target parent video.

7. The method for detecting the same source in mixed-edit videos according to any one of claims 1 to 6, characterized in that, Also includes: If the target parent feature vector corresponding to the slice to be detected does not exist in the target parent feature library, or the target similarity between the target parent feature vector and the target feature vector does not meet the preset similarity condition, then the slice to be detected is determined as an unattributed slice, and the proportion of proprietary content corresponding to the video to be detected is determined based on the unattributed slice.

8. A device for detecting the same source in mixed-edit videos, characterized in that, include: The detection feature extraction module is used to perform temporal slicing of the video to be detected and extract the target feature vector corresponding to each slice. The hit relationship construction module is used to determine the target parent slice based on the target similarity between each parent feature vector in the target parent feature library and the target feature vector, and to construct the target hit relationship between the slice to be detected and the target parent slice; The attribution matrix determination module is used to group the target hit relationships based on the parent identifier corresponding to the target parent slice, and construct a target attribution matrix between the video to be detected and each target parent video based on the target grouping results; different target attribution matrices are independent of each other; the target attribution matrix is ​​used to record the time interval of the video to be detected attributed to the corresponding target parent video, the time interval of the target parent slice, and the target saliency weight corresponding to the target parent slice; The target saliency weight represents the degree to which the target parent slice carries the core plot content of the corresponding parent video; The detection result acquisition module is used to determine the continuous reference duration, relative concentration ratio, and core plot coverage of each target parent video in the video to be detected based on the target attribution matrix, and to determine the target coverage intensity coefficient corresponding to the target parent video based on the continuous reference duration, relative concentration ratio, and core plot coverage, and to generate a homogeneous detection result evidence package based on the target coverage intensity coefficient; wherein, the relative concentration ratio represents the proportion of the video content of the target parent video in the video to be detected, and the core plot coverage represents the degree to which the video to be detected covers the target core plot of the target parent video.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method for detecting the same source of mixed-edit videos as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the method for detecting the same source of mixed-edit videos as described in any one of claims 1 to 7.

Images