A cross-language video detection method, device, equipment and storage medium

Abstract

The application discloses a cross-language video carrying detection method and device, equipment and storage medium, and relates to the field of video analysis, which comprises the following steps: performing time sequence slicing on a video to be detected, determining a first script language and text to be detected corresponding to each to-be-detected slice and text confidence; if the text confidence meets a pre-set confidence condition, generating a target semantic vector corresponding to the to-be-detected slice by using a target embedding model, performing visual feature analysis on the to-be-detected slice to determine a target visual anchor point label, and determining a to-be-detected semantic sequence corresponding to the to-be-detected video based on the target semantic vector; determining a target parent feature and a target parent semantic sequence of a target parent video from a target parent feature library based on the target semantic vector; determining a target transfer cost by using a target constraint, and then determining a homology risk score, so as to generate a cross-language video carrying detection report. The application improves the accuracy of cross-language video carrying detection.

Description

Technical Field

[0001] This invention relates to the field of video analysis, and in particular to a method, apparatus, device, and storage medium for detecting cross-language video transfer. Background Technology

[0002] In the ecosystem of short and medium-length video dissemination, there is a large amount of cross-language reposting and text remixing: replacing the original copyrighted video's voiceover with another language's voiceover, or translating and rewriting the narration and then re-dubbing it before uploading. Existing video source detection solutions mainly rely on audio voiceprint comparison or literal text matching. When the voiceover language is switched or the text is deeply paraphrased, the overlap between voiceprint features and literal text approaches zero, resulting in a high false negative rate for existing detection mechanisms in cross-language rewriting and reposting scenarios. 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 cross-language video transfers, which can improve the accuracy of cross-language video transfer detection. The specific solution is as follows: Firstly, this application discloses a method for detecting cross-language copied videos, including: The video to be tested is divided into time-series slices to determine the first text language corresponding to the video to be tested, as well as the text to be tested and the text confidence level corresponding to each slice to be tested. If the text confidence meets the preset confidence conditions, the target embedding model is used to generate the target semantic vector corresponding to the slice to be detected based on the text to be detected. Visual feature analysis is performed on the slice to be detected to determine the target visual anchor label, and the target semantic vector is used to determine the semantic sequence to be detected; the target semantic vector is a cross-language semantic vector. Based on the target semantic vector, the target parent video corresponding to the video to be detected is determined from the target parent feature library, and the target parent semantic sequence is determined based on the target parent features; Using target constraints, the target transfer cost between each slice to be detected and each target parent slice is determined based on the semantic sequence to be detected and the semantic sequence to be the target source. Based on the target transfer cost, the homology risk score between the video to be detected and the target parent video is determined, and a cross-language transfer detection report is generated based on the target transfer cost and the homology risk score.

[0004] Optionally, before determining the target parent video features corresponding to the target parent video from the target parent feature library based on the target semantic vector, the method further includes: Each source video is acquired, and the source videos are slicing in a time sequence to obtain corresponding source slices. The second text language corresponding to each source video is determined, as well as the source text corresponding to each source slice, and the source text confidence score corresponding to the source text is obtained. Based on the parent text corresponding to the parent text slice, the parent text structured text corresponding to the parent text slice is determined, and the parent text semantic vector corresponding to the parent text slice is generated using the target embedding model based on the parent text and the parent text structured text; the parent text semantic vector is a cross-language semantic vector; Visual feature analysis is performed on the parent tissue slices to determine the parent tissue visual anchor point labels corresponding to the parent tissue slices; The semantic vector of the source video corresponding to each source video slice, the confidence score of the source text, the second copywriting language corresponding to the source text, and the source visual anchor tag corresponding to the source video slice are written into the target source video feature library as source video features of the source video.

[0005] Optionally, the cross-language video copying detection method further includes: Based on the version hash value corresponding to each of the parent video slices, it is determined whether the parent video has undergone content revision. When the version hash value corresponding to the parent video slice indicates that the parent video has undergone content revision, the parent video slice is updated with incremental features.

[0006] Optionally, the cross-language video copying detection method further includes: If neither the first text confidence score nor the second text confidence score corresponding to the slice to be detected meets the preset confidence condition, then the target visual language model is invoked to generate the target description text corresponding to the target keyframe of the slice to be detected and determine the corresponding target confidence score; wherein, the text to be detected includes the speech text to be detected and the screen text to be detected; the first text confidence score is the text confidence score corresponding to the speech text to be detected, and the second text confidence score is the text confidence score corresponding to the screen text to be detected; Accordingly, the step of generating the target semantic vector corresponding to the slice to be detected based on the text to be detected using the target embedding model includes: The target semantic vector corresponding to the slice to be detected is generated based on the target description text using a target embedding model.

[0007] Optionally, the step of determining the target transfer cost between each slice to be detected and each target parent slice based on the target semantic sequence and the target parent semantic sequence using target constraints includes: If the target semantic vector corresponding to the slice to be detected is a semantic vector generated based on the target description text, then the target confidence penalty term between the slice to be detected and each of the target parent slices is determined by using confidence constraints based on the target confidence corresponding to the target description text.

[0008] Optionally, the step of determining the target transfer cost between each slice to be detected and each target parent slice based on the target semantic sequence and the target parent semantic sequence using target constraints includes: If the cosine distance between the target semantic vector corresponding to the slice to be detected and the parent semantic vector corresponding to the target parent slice in the target semantic space is greater than a preset distance threshold, then a semantic distance penalty term between the slice to be detected and the target parent slice is determined based on cross-language semantic distance constraints.

[0009] Optionally, the step of determining the target transfer cost between each slice to be detected and each target parent slice based on the target semantic sequence and the target parent semantic sequence using target constraints includes: If the slice to be detected has the target visual anchor label, and the target parent slice has the parent visual anchor label, then the anchor reward item is determined in the target semantic space based on the cosine distance between the target semantic vector corresponding to the slice to be detected and the parent semantic vector corresponding to the target parent slice in the target semantic space using the first visual anchor label constraint. If the slice to be detected does not have the target visual anchor label and the target parent slice has the parent visual anchor label, or if the slice to be detected has the target visual anchor label and the target parent slice does not have the parent visual anchor label, then the anchor penalty term between the slice to be detected and the target parent slice is determined by using the second visual anchor label constraint. The visual anchor label constraint includes the first visual anchor label constraint and the second visual anchor label constraint.

[0010] Secondly, this application discloses a cross-language video transfer detection device, comprising: The text confidence acquisition module is used to perform time-series slicing of the video to be detected, determine the first text language corresponding to the video to be detected, and the text to be detected and the text confidence corresponding to each slice to be detected; The first sequence acquisition module is used to, if the text confidence meets the preset confidence conditions, generate the target semantic vector corresponding to the slice to be detected based on the text to be detected using the target embedding model, perform visual feature analysis on the slice to be detected to determine the target visual anchor label, and determine the semantic sequence to be detected based on the target semantic vector; the target semantic vector is a cross-language semantic vector; The second sequence acquisition module is used to determine the target parent feature of the target parent video corresponding to the video to be detected from the target parent feature library based on the target semantic vector, and to determine the target parent semantic sequence based on the target parent feature; The detection report generation module is used to determine the target transfer cost between each slice to be detected and each target parent slice based on the target semantic sequence and the target parent semantic sequence using target constraints, and to determine the homology risk score between the video to be detected and the target parent video based on the target transfer cost, so as to generate a cross-language transfer detection report based on the target transfer cost and the homology risk score.

[0011] Thirdly, this application discloses an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned cross-language video transfer detection method.

[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 cross-language video transfer detection method.

[0013] In this application, when detecting cross-language plagiarism in videos, the video to be detected is divided into temporal slices to determine the first text language corresponding to the video to be detected, as well as the text to be detected and the text confidence level corresponding to each slice to be detected. If the text confidence level meets the preset confidence level conditions, a target embedding model is used to generate a target semantic vector corresponding to the slice to be detected based on the text to be detected. Visual feature analysis is performed on the slice to be detected to determine the target visual anchor tag, and a target semantic sequence to be detected is determined based on the target semantic vector. The target semantic vector is a cross-language semantic vector. Based on the target semantic vector, the target parent video corresponding to the video to be detected is determined from the target parent feature library, and a target parent semantic sequence is determined based on the target parent feature. Target constraints are used to determine the target transfer cost between each slice to be detected and each target parent slice based on the target semantic sequence to be detected and the target parent semantic sequence. Based on the target transfer cost, a homology risk score between the video to be detected and the target parent video is determined, and a cross-language plagiarism detection report is generated based on the target transfer cost and the homology risk score. As can be seen, this application projects texts from different languages ​​directly into a cross-lingual semantic space through a pre-trained multilingual embedding model, rather than performing explicit machine translation and then comparing them. Therefore, it avoids the error propagation problem caused by the instability of machine translation quality to a certain extent, and makes cross-lingual expressions with similar semantics (including deep paraphrasing) still comparable in distance calculation. Even if the text of the video to be detected and the text of the original video are in different languages ​​and the literal overlap is close to zero, the semantic distance between the two can still be measured in a unified semantic space, which helps to reduce the false negative rate in cross-lingual rewriting and copying scenarios. Furthermore, this application introduces target constraints, including confidence constraints, cross-lingual semantic distance constraints, and visual anchor label constraints, when determining the target transfer cost. On the one hand, it utilizes visual anchor signals such as strong transitions in video frames to assist in constraining cross-lingual semantic alignment. When the video to be detected and the target parent video have structural consistency at strong transition positions, the alignment path weight at that position is strengthened. When the frame structures are inconsistent, even if the text semantics are similar, the alignment cost at that position increases accordingly, thereby reducing the cross-lingual mis-adsorption problem caused by short texts such as common greetings to a certain extent. On the other hand, it uses confidence constraints to increase the matching path cost of low-reliability segments, avoiding the excessive impact of mismatches of low-confidence texts on the final score. Based on this, it determines the homology risk score based on the target transfer cost and generates a cross-lingual plagiarism detection report, which can output structured temporal mapping reference information for manual review, providing an objective and traceable basis for judging cross-lingual rewriting and plagiarism behavior. 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 1 This is a flowchart of a cross-language video copying detection method disclosed in this application; Figure 2 This is a schematic diagram of a specific cross-language video copying detection system architecture disclosed in this application; Figure 3 This application discloses a specific weak text catchment decision-making process diagram. Figure 4 This is a schematic diagram of a specific visual anchor point reward mechanism and degradation mechanism disclosed in this application; Figure 5 This is a schematic diagram of the structure of a cross-language video transport 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 ecosystem of short and medium-length video dissemination, there is a large amount of cross-language reposting and text remixing: replacing the original copyrighted video's voiceover with another language's voiceover, or translating and rewriting the narration and then re-dubbing it before uploading. Existing video source detection schemes mainly rely on audio voiceprint comparison or literal text matching. When the voiceover language is switched or the text undergoes deep paraphrasing, the overlap between voiceprint features and literal text approaches zero, resulting in a high false negative rate for existing detection mechanisms in cross-language rewriting and reposting scenarios. To solve the above technical problems, this application discloses a cross-language reposting video detection method that can improve the accuracy of cross-language reposting video detection.

[0018] See Figure 1 As shown, this embodiment of the invention discloses a method for detecting cross-language video copying, including: Step S11: Perform time-series slicing on the video to be tested, determine the first text language corresponding to the video to be tested, and the text to be tested and text confidence level corresponding to each slice to be tested.

[0019] In this embodiment, as Figure 2 The diagram illustrates a specific cross-language video plagiarism detection system architecture. The video to be detected can be a short or medium-length video uploaded to a content platform that requires source identity verification. After obtaining the video to be detected, it can be time-series sliced ​​according to a fixed time window to obtain the corresponding detection slices. For example, it can be sliced ​​at a granularity of 5 seconds per time window, and the specific time window length can be flexibly set according to actual business needs. It should be noted that the slicing rules used when time-series slicing the video to be detected are consistent with the time-series slicing rules used when the parent video is entered into the database, so as to facilitate subsequent serialization comparison at the slice granularity.

[0020] After completing the time-series slicing, the language identification module (Language ID) can be invoked to determine the first text language corresponding to the video to be detected, that is, to automatically determine the language category of the audio or text of the video to be detected. Simultaneously, the corresponding text to be detected is extracted for each slice. Specifically, the text to be detected includes the audio text to be detected and the on-screen text to be detected. The audio text to be detected can be obtained by recognizing the audio of the slice using Automatic Speech Recognition (ASR) technology, while the on-screen text to be detected can be obtained by recognizing subtitles, titles, and other text appearing in the slice using Optical Character Recognition (OCR) technology. While extracting the text to be detected, the text confidence score corresponding to the text to be detected is obtained. The text confidence score includes the first text confidence score (asr_conf) corresponding to the audio text to be detected and the second text confidence score (ocr_conf) corresponding to the on-screen text to be detected, which are used to characterize the reliability of the speech recognition results and the on-screen text recognition results, respectively.

[0021] Step S12: If the text confidence meets the preset confidence conditions, then the target embedding model is used to generate the target semantic vector corresponding to the slice to be detected based on the text to be detected, and the visual feature analysis of the slice to be detected is performed to determine the target visual anchor label, and the target semantic vector is used to determine the semantic sequence to be detected; the target semantic vector is a cross-language semantic vector.

[0022] It is understood that in this embodiment, the preset confidence condition can be that the text confidence is not lower than a preset confidence threshold. For example, the confidence threshold can be set to 0.4. If at least one of the first text confidence and the second text confidence meets the preset confidence condition, it indicates that the current slice to be detected contains reliable text information. At this time, the target embedding model can be used to generate the target semantic vector corresponding to the slice to be detected based on the text to be detected.

[0023] In one specific implementation, before generating the target semantic vector, event semantic structure extraction can be performed on the text slice to be detected with valid confidence. That is, the natural language understanding module is called to perform structured event extraction on the text to be detected to obtain the structured text to be detected corresponding to the slice to be detected, which is the structured event unit. The structured event unit contains the following fields: Event Subject, which is the executor of the action, such as character name or character type, for example, protagonist, dragon; Core Action, which is the main action performed by the subject, for example, attacked, escaped, confronted; Action Object, which is the recipient or object of the action, which can be empty, for example, protagonist, village; Scene Tag, which is the scene type of the current segment, for example, COMBAT, DIALOGUE, CHASE, REVEAL, NONE; Conflict Tag, which is the plot conflict attribute, for example, CONFRONTATION, BETRAYAL, RESCUE, NONE. Understandably, since the text of the video to be detected may be a rewritten version in Chinese or other languages, the event extraction module needs to support multilingual input. Standardized output of action type and scene labels helps to preserve the comparability of event semantics in cross-language comparisons. Based on this, structured event units can be concatenated with the original text to be detected and used as input to the target embedding model to generate a high-dimensional vector representation that integrates event semantics. Since the descriptions of the same event differ significantly in literal form in different languages, and event structure fields (such as action type and conflict labels) have high semantic stability in different languages, introducing structured event representation helps to improve the stability of cross-language event semantic alignment.

[0024] The target embedding model is a pre-trained multilingual embedding model. Exemplary models include pre-trained models that support unified semantic space encoding across multiple languages, such as mBART and XLM-R models. The target semantic vector is a cross-lingual semantic vector, i.e., a language-independent high-dimensional semantic vector. It should be noted that this embodiment does not perform explicit machine translation followed by comparison during the semantic vector generation process. Instead, it utilizes the cross-lingual alignment capability of the multilingual embedding model to ensure that texts describing the same event in different languages ​​are close in distance within the cross-lingual semantic space, thereby mitigating, to some extent, the error propagation problem caused by the instability of machine translation quality.

[0025] In this embodiment, visual feature analysis is performed on the slices to be detected to determine the target visual anchor labels corresponding to the slices. Specifically, visual feature abrupt change points of each slice to be detected can be extracted simultaneously: strong transition signals such as large abrupt changes in optical flow between frames and strong scene transitions (shot cuts) are detected. Slices to be detected that have such strong transition signals are marked as having visual anchors, i.e., the visual anchor flag (visual_anchor_flag) is true, otherwise they are marked as not having visual anchors. Visual anchors provide temporal synchronization signals unrelated to dubbing, which can help constrain the temporal drift range in subsequent temporal alignment processes. After obtaining the target semantic vectors corresponding to each slice to be detected, the target semantic vectors are arranged in chronological order to determine the target semantic sequence corresponding to the video to be detected.

[0026] In one specific implementation, such as Figure 3The diagram illustrates the complete decision-making path from text extraction to semantic vector generation in the video to be detected, i.e., the weak text fallback decision-making process. If the first and second text confidence scores corresponding to the segment to be detected do not meet the preset confidence conditions, i.e., the ASR and OCR confidence scores for a certain period are both lower than the preset threshold (example parameter: 0.4), it indicates that there is a lack of reliable speech text and on-screen text for that period. At this time, the weak text fallback strategy is triggered: the target visual language model is called to generate the target description text corresponding to the target keyframe of the segment to be detected and the corresponding target confidence score vlm_conf is determined; where the text to be detected includes the speech text to be detected and the on-screen text to be detected; the first text confidence score is the text confidence score corresponding to the speech text to be detected, and the second text confidence score is the text confidence score corresponding to the on-screen text to be detected; the target description text is the objective scene description text generated by the VLM (Vision-Language Model) for the target keyframe. Accordingly, the target semantic vector corresponding to the segment to be detected is generated based on the target embedding model using the target text to be detected, including: generating the target semantic vector corresponding to the segment to be detected based on the target description text using the target embedding model. In other words, for the VLM text obtained as a fallback, we can first perform structured event extraction to generate structured event unit fields that are consistent with the parent text, and then concatenate them with the target description text before inputting them into the target embedding model to generate cross-language semantic vectors.

[0027] Furthermore, for periods with extremely low VLM confidence, such as periods with no valid image content like a completely black screen or a solid color background, the corresponding slices to be detected can be filtered and marked as invalid segments, preventing them from entering the subsequent alignment sequence, thereby reducing the interference of invalid segments on the homology risk score.

[0028] In this embodiment, vlm_conf (VLM fallback confidence) is a reliability value between 0 and 1 calculated and calibrated by the detection system based on components such as VLM output information, cross-frame description consistency, image-text correspondence, and structural integrity when the confidence scores of both ASR and OCR for a certain time segment of the video under test are lower than a preset threshold, and the system calls the Visual-Language Model (VLM) to generate objective scene description text for the keyframes of that time segment. It has two uses: first, as a gate for filtering invalid segments, segments with vlm_conf below the effective threshold (such as pure black screens, solid color backgrounds, and segments with contradictory cross-frame descriptions) are filtered out and not included in subsequent comparison sequences; second, as a source of penalty terms in the temporal warping algorithm, the matching distance of the VLM fallback segment is increased by multiplying (1-vlm_conf) by the penalty coefficient, so that low-reliability text does not have an excessive impact on the final homologous score. As shown in Table 1 below, for each test period that triggers the fallback, the system extracts K keyframes from that period (example parameter: K=3) to request the objectively described prompt words to call VLM and obtain the corresponding output.

[0029] Table 1 VLM Output Items and Corresponding Acquisition Methods

[0030] The original output, along with the model name / version, prompt word version, and keyframe timestamps, is written to the runtime log, making vlm_conf recalculated and auditable. It's important to clarify that vlm_conf is not a confidence score that the VLM is required to report automatically. Instead, it is calculated by the detection system based on objectively available information in the VLM output (token generation probability, consistency of multi-keyframe descriptions, image-text matching degree, and structured field completeness) using the following formula, which can be reproduced by those skilled in the art: (1) Generation probability components This component characterizes the VLM-generated main description text y. The degree of certainty of the timing itself is denoted as u, the timing segment to be tested that currently triggers VLM fallback is denoted as u, the number of keyframes extracted from segment u and sent to VLM is denoted as K, and the k-th keyframe of segment u (k=1,2,…,K) is denoted as . VLM for keyframes The generated objective scene description text is Take the middle keyframe from the K descriptions, i.e., k = K / 2 Corresponding description ( · This indicates rounding up, such as 3 / 2 =2) is y Main description text y The total number of tokens after word segmentation by the model is n, and the main description text is y. The t-th token is (t=1,2,…,n), the VLM interface returns The probability of generating a logarithm is , ,generate The conditional probability is , , ,in Input VLM keyframes (i.e., k=) when generating the main description text K / 2 (corresponding frame) For token The weight, the value of which is determined by whether t belongs to Decision. When the interface returns the log probability of the token, it is calculated using the content-weighted geometric mean: ; Weights in the above formula Values ​​are segmented based on whether the token carries event semantics: ; In this case, the summation subscript t takes the same value range in both the numerator and denominator, which is t=1,2,…,n, i.e., it iterates through the main description text y. All tokens; This is a set of content-related token indices, specifically token indices that carry information about the event subject, core action, action object, and scene. ;Pick Instead of taking directly This is to avoid logarithmic divergence when the generation probability of individual tokens approaches 0. The underflow protection constant is greater than 0; an example value is given. Content-based keyword weighting and functional word weight All are greater than 0, and Example values , The value of does not systematically decrease with increasing text length n, thus reducing the interference of sentence length and function words on confidence. A higher generation probability indicates greater certainty from the model regarding the description. achievable Therefore The equivalent product of is shown below: ; In this embodiment, the main description text y Perform structured event extraction consistent with that on the parent sample warehouse entry side, and classify the token subscripts that hit the value ranges of the four fields of subject, action, object, and scene_tag into , and the remaining subscripts are classified into its complement; if event extraction is not enabled, it can be degraded to determination according to the stop word list, that is, tokens hitting the stop word list are classified into the complement, and the rest are classified into .

[0031] For example, let n=8, the generation probability of each token is sequentially , wherein the 3rd and 6th tokens are function words, and the remaining six tokens are content words, take , w f = 0.3, then ; ; .

[0032] (2) Cross-frame consistency component . The picture content within the same time series segment is continuous: if the descriptions given by the VLM for each key frame of the segment are consistent with each other in terms of characters, actions and scenes, the description is more likely to reflect the real content of the picture; if the descriptions of each frame contradict each other (a typical manifestation of hallucination), the reliability is low. Take the non-negative mean of pairwise cosine similarities of description vectors of each frame, as shown in the following formula: ; In the formula, the cosine similarity is calculated according to the standard definition, is the semantic vector of the k-th key frame description, , , d is the dimension of the semantic vector, which is determined by the multilingual embedding model used, is the L2 norm. As shown in the following formula, since the semantic vector of the i-th key frame description and the semantic vector of the j-th key frame description have both undergone L2 normalization, this formula can be simplified to a vector inner product: ; It can be understood that the double sum traverses all unordered key frame pairs (i,j), where 1≤i<j≤K, and the total number of frame pairs is K(K-1) / 2; the coefficient 2 / [K(K-1)] is exactly the reciprocal of the total number of frame pairs, so is the arithmetic mean of the similarity of each frame pair; ; take is to truncate negative similarity (opposite description semantics) to 0, so as to prevent negative values from offsetting positive values and causing the result to be falsely high, thus ; The encoding function E is consistent with the multilingual embedding model used in the comparison stage of this embodiment, mapping the text to d-dimensional real vectors and performing L2 normalization, so that... This ensures measurement within the same semantic space; when K=1, the total number of frame pairs is 0, as stipulated. Take a neutral value of 0.5 (meaning that this dimension provides neither positive nor negative evidence); in practice, K ≥ 2 can also be directly required to avoid this situation.

[0033] For example, assuming K=3, the pairwise cosine similarity of the three keyframe descriptions after encoding is: , c If all are positive, then: Total number of frame pairs = K(K-1) / 2 = 3 × 2 / 2 = 3; .

[0034] (3) Corresponding components of images and text This component characterizes the degree of matching between the descriptive text and the keyframe image, used to suppress illusory descriptions unrelated to the image content. Let g be the original image-text matching score given by the VLM or independent image-text encoder for segment u, then: ; Understandable, and Once the annotation and validation set statistics are determined, they are fixed and recorded along with the weight version to ensure that the results can be recalculated; To ensure that individual fractions exceeding the quantile range do not cause the component to go out of bounds, therefore... If the model used directly outputs a matching probability that has been normalized to [0,1], then If the model supports phrase grounding, the success rate of key phrase grounding can be calculated. According to the above formula by coefficient Linear fusion, that is: .

[0035] (4) The structural integrity component c_validity is used to characterize whether the VLM output can be reliably parsed by the system and converted into structured event units. Let the set of validation items be Φ={ 1 , 2 ,…, M If the number of items passes is the proportion of the total number of items verified, then: ; The indicative function I(·) is defined as follows: ; Validation item set Exemplary configuration (M=4): Used to verify that the VLM output can be successfully parsed according to the predefined JSON pattern; Used to verify that the subject field is not empty; Used to verify that the action field is not empty; The validation field `scene_tag` is used to verify that its value falls within the allowed tag set `{COMBAT, DIALOGUE, CHASE, REVEAL, NONE}`. During implementation, validation items can be added or removed according to the actual field conventions, and M will be adjusted accordingly. A black screen, a solid color, or other content-free display usually cause this component to... At the same time, it is relatively low.

[0036] For example, if pass, , The application was rejected because the field was empty. The test fails because the value "UNKNOWN" is not in the allowed set. .

[0037] The original confidence level is obtained by weighted summation of the above four components. : ; The constraints are and Example weighting , , , Since all four components fall within [0,1] and the sum of their weights is 1, therefore... The actual values ​​of the weights are determined by the labeled validation set (for example, the target search weight combination is the one with the lowest Brier score after calibration confidence), and the weight version number is recorded with the results.

[0038] The final confidence level is then obtained through calibration function mapping, as shown in the following formula: ; In the formula This is a monotonically increasing mapping fitted to the labeled validation set. Its purpose is to align the output value with the empirical probability that "the VLM description matches the actual content of the image." Conventional calibration methods such as temperature scaling, Platt scaling, or isotonic regression can be used. Taking Platt scaling as an example, its form is... Parameters a and b are determined by the maximum likelihood estimation on the validation set. If calibration is not enabled, Calibrate is taken as the identity mapping, i.e., vlm_conf= The calibration method, parameters, and version number should be recorded along with the results to ensure that the results can be recalculated.

[0039] It should be noted that the VLM interface only returns text and does not provide the token logarithmic probability. c token Uncomputable. The system was changed to remove this component and renormalize the weights of the remaining three components according to the original proportions, as shown below; ; Substitute the example weights , , ,but The equivalent weights of the three components are approximately 0.25 / 0.60 ≈ 0.4167, 0.25 / 0.60 ≈ 0.4167, and 0.10 / 0.60 ≈ 0.1667, respectively. Their sum is still 1. Therefore... The value range is consistent with the main path, and the thresholds for both paths can be used interchangeably. The system can also generate values ​​by sampling multiple keyframes and low temperatures within the same time period, taking the results from consistent data to enhance performance. The reliability of the data is determined by the confidence level calculated via this path, which is marked with confidence_source="derived_without_logprob" to distinguish it from the calculation path that contains model probability information.

[0040] In one specific implementation, when determining invalid segments, the valid threshold of VLM is set to... (Example parameter: 0.35). At that time, the VLM text enters the event semantic extraction and multilingual embedding, participates in the temporal comparison, and its vlm_conf value is also used as the input of the penalty term of the temporal warping algorithm; When this occurs, the segment is marked as invalid and filtered out, and it is not included in the automatic scoring. The threshold is selected based on a trade-off between the reliability curve on the validation set and the false acceptance / false rejection rate. Table 2 below shows different test cases and the corresponding gating results.

[0041] Table 2. Test cases and corresponding gating results

[0042] For example, the confidence record data structure for a single VLM fallback call is shown below: { "vlm_text": "A man runs through a dark forest, dodging flames behind him", "vlm_conf_detail": { "c_token": 0.9016, "c_consistency": 0.8416, "c_grounding": 0.81, "c_validity": 1.0, "weight_version": "vlm_conf_v1", "calibration_version": "cal_v1", "confidence_source": "token_logprob_and_grounding" }, "vlm_conf": 0.8735 }

[0043] It is understandable that the vlm_conf penalty only applies to segments of text actually provided by VLM; valid segments of ASR / OCR text are not subject to this penalty. A specific example of confidence penalty stacking is as follows: text_conf = max(getattr(test_seg, 'asr_conf', 0.0), getattr(test_seg, 'ocr_conf', 0.0)) vlm_conf = (getattr(test_seg, 'vlm_conf', 0.0) if getattr(test_seg, 'is_vlm_fallback', False) else 1.0) dist += (1.0 - text_conf) config["lang_penalty_factor"] dist += (1.0 - vlm_conf) config["vlm_penalty_factor"] Step S13: Based on the target semantic vector, determine the target parent video corresponding to the video to be detected from the target parent feature library, and determine the target parent semantic sequence based on the target parent features.

[0044] In this embodiment, when determining the target parent video from the target parent feature library, a combination of coarse screening and fine ranking of candidate parent videos can be used. Specifically, firstly, the global semantic feature summary vector of the video to be detected is extracted. For example, it can be obtained by pooling or weighted averaging each target semantic vector in the semantic sequence to be detected. Then, the candidate parent videos of the top K semantic nearest neighbors are recalled in the parent library through hierarchical navigable small world (HNSW) vector indexing to complete the coarse screening of candidate parent videos. Subsequently, cosine similarity is used to finely rank and confirm local segments in the coarse screening results to obtain candidate comparison sequences. The candidate parent videos confirmed by fine ranking are used as the target parent videos corresponding to the video to be detected, and the target parent features of the target parent videos are read. Based on this, the parent semantic vectors in the target parent features corresponding to each target parent slice of the target parent video are arranged in chronological order to determine the target parent semantic sequence corresponding to the target parent video.

[0045] It should be noted that before determining the target source video features corresponding to the target source video from the target source feature library based on the target semantic vector, it is necessary to pre-construct the target source feature library. This involves performing multilingual semantic extraction and incremental indexing of the source video on the source input side. Specifically, this may include: acquiring each source video, performing temporal slicing on the source video to obtain the corresponding source slices, determining the second text language corresponding to the source video, and the source text corresponding to each source slice, and obtaining the source text confidence score corresponding to the source text; and based on the source slices... The corresponding parent text is used to determine the parent text structured text corresponding to the parent text slice, and the parent text semantic vector corresponding to the parent text slice is generated based on the parent text and the parent text structured text using the target embedding model; the parent text semantic vector is a cross-language semantic vector; visual feature analysis is performed on the parent text slice to determine the parent text visual anchor tag corresponding to the parent text slice; the parent text semantic vector, parent text confidence, second copy language corresponding to the parent text, and parent text visual anchor tag corresponding to the parent text slice of each parent text slice of the parent text video are written into the target parent text feature library as parent text features of the parent text video.

[0046] In one specific implementation, the copyrighted master video is sliced ​​sequentially according to a fixed time window. The language identification module (Language ID) is invoked to determine the source language type of each master video slice, i.e., the second text language. The speech recognition (ASR) text and on-screen text recognition (OCR) text of that master video slice are extracted as master text, and their respective confidence scores (asr_conf, ocr_conf) are calculated as the master text confidence score. After obtaining the ASR / OCR text of each master video slice, the natural language understanding module is invoked to perform structured event extraction on the text, outputting structured event units containing fields such as event subject, core action, action object, scene tag, and conflict tag, as the master structured text. The structured event units and the original master text are used together as input to a multilingual embedding model. For example, the master structured text and the original master text are concatenated and input into the target embedding model to generate a high-dimensional vector representation that integrates event semantics, i.e., the master semantic vector. Here, explicit machine translation followed by comparison is not performed. Instead, the cross-language alignment capability of a multilingual pre-trained embedding model (exemplary model category: pre-trained models that support multilingual unified semantic space encoding, such as mBART and XLM-R models) is used to generate language-independent high-dimensional semantic vectors, making texts describing the same event in different languages ​​close in distance in the vector space. Simultaneously, visual feature mutation points of each parent slice are extracted: strong transition signals such as large changes in optical flow between frames and strong scene changes (Shot Cut) are detected and marked as visual anchors (visual_anchor_flag), thus obtaining the parent visual anchor label.

[0047] For example, the data structure for storing the parent stock can be as follows (where the values ​​are all example parameters): { "source_id": "BASE_VID_EN_001", "version_hash": "v1.0_a8f9c", "segments": [ { "time_window": "00:20-00:25", "source_lang": "en", "asr_text": "The dragon suddenly attacked.", "asr_conf": 0.95, "event_unit": { "subject": "dragon", "action": "attacked", "object": "protagonist", "scene_tag": "COMBAT", "conflict_tag": "CONFRONTATION" }, "semantic_embedding": "[0.12, 0.77, ..., -0.34]", "visual_anchor_flag": true } ] } In one specific implementation, incremental version management can also be performed on the target source feature library. That is, the cross-language video copying detection method further includes: determining whether the source video has undergone content revision based on the version hash value corresponding to each source slice; and performing incremental feature updates on the source slice when the version hash value indicates that the source video has undergone content revision. Specifically, a version hash value (version_hash) is calculated based on the slice content. When the source video undergoes content revision, the version hash value of the changed source slice also changes accordingly, which can be used to determine whether the source video has undergone content revision. Incremental feature updates are performed on the changed source slices, i.e., recalculating their source semantic vectors and other source features and updating them into the library. For unchanged source slices, existing semantic vectors are directly reused, thereby avoiding the overhead of full library recalculation when source content is updated.

[0048] Step S14: Utilize target constraints to determine the target transfer cost between each slice to be detected and each target parent slice based on the semantic sequence to be detected and the target parent semantic sequence. Determine the homology risk score between the video to be detected and the target parent video based on the target transfer cost. Generate a cross-language transfer detection report based on the target transfer cost and the homology risk score.

[0049] In this embodiment, a multilingual temporal warping algorithm with visual anchor adjustment can be used to determine the target transition cost. Specifically, the finely ranked semantic sequence to be detected and the target parent semantic sequence are input into a two-dimensional state transition matrix, and the optimal transition cost at each position is calculated sequentially. The following constraints and adjustment mechanisms are introduced during the calculation process: (1) Basic cross-linguistic semantic distance and cross-linguistic semantic distance constraints. Cosine distance is used to measure the distance between the corresponding segments in the target semantic sequence and the target parent sequence in a unified semantic space (i.e., the target semantic space), and a preset distance threshold is set (exemplary parameter: 0.40). That is, if the cosine distance between the target semantic vector corresponding to the target slice and the parent semantic vector corresponding to the target parent slice in the target semantic space is greater than the preset distance threshold, a semantic distance penalty term between the target slice and the target parent slice is determined based on the cross-linguistic semantic distance constraint. That is, when the threshold is exceeded, an additional penalty term is added (exemplary, an additional penalty cost of 1.0) to significantly increase the alignment cost between semantically dissimilar segments.

[0050] (2) Confidence Constraint. The confidence constraint includes two aspects: linguistic confidence penalty and VLM fallback confidence penalty. Linguistic confidence penalty refers to adding a penalty term proportional to the degree of insufficient linguistic confidence of the segment to be detected to the cosine distance. The penalty coefficient is an example parameter (exemplarily, the penalty coefficient is 0.20). VLM fallback confidence penalty means that if the target semantic vector corresponding to the segment to be detected is a semantic vector generated based on the target description text, that is, the segment to be detected is provided with text by VLM fallback, then the confidence constraint is used to determine the target confidence penalty term between the segment to be detected and each target parent segment based on the target confidence corresponding to the target description text. That is, a penalty term proportional to the degree of insufficient VLM confidence is added to the cosine distance, and the penalty coefficient is greater than the penalty coefficient of linguistic confidence penalty (exemplarily, the penalty coefficient is 0.30) to reflect the relatively high uncertainty of VLM description. Through the aforementioned confidence constraints, the matching path cost of low-reliability segments is increased accordingly, thereby avoiding the excessive impact of mismatches of low-confidence texts on the final score.

[0051] (3) Visual anchor label constraints. See [link / reference] Figure 4 As shown, Figure 4This demonstrates the impact of two scenarios—bidirectional consistency and unidirectional absence of visual anchors—on the distance cost. The visual anchor label constraint includes a first visual anchor label constraint and a second visual anchor label constraint. The first visual anchor label constraint corresponds to a bidirectional visual anchor reward mechanism: if the segment to be detected has a target visual anchor label, and the target parent segment has a parent visual anchor label (i.e., both the segment to be detected and the parent segment are labeled with visual anchors (bidirectional consistency), then the anchor reward term is determined based on the cosine distance in the target semantic space between the target semantic vector corresponding to the segment to be detected and the parent semantic vector corresponding to the target parent segment. Specifically, a discount reward is applied to the aligned cosine distance (example parameter: reduction of 0.15) to strengthen the strong transition position. Alignment weight; Second visual anchor label constraint corresponds to a one-way degradation penalty mechanism for visual anchors: If the slice to be detected does not have a target visual anchor label and the target parent slice has a parent visual anchor label, or, the slice to be detected has a target visual anchor label and the target parent slice does not have a parent visual anchor label, that is, one side has an anchor and the other side does not have an anchor, then the anchor penalty item between the slice to be detected and the target parent slice is determined by the second visual anchor label constraint, that is, a degradation penalty is applied (exemplary parameter: increase by 0.10), to avoid forcibly aligning mismatched screen structures due to text alignment.

[0052] (4) Consecutive Jump / Expand Constraints. The number of times the target parent slice is consecutively jumped and the number of times the slice to be detected is consecutively expanded on the same parent node are counted independently. When both exceed their respective upper limits (example parameters: 3 jumps, 4 expansions), the transition path is set to unreachable, thereby preventing extreme misalignment while allowing for temporal flexibility. In addition, for jump or expansion type state transitions, a jump penalty can be added (example parameter: 0.25) to make non-one-to-one corresponding transition paths bear additional costs.

[0053] (5) Maximum temporal misalignment window. Set the Sakoe-Chiba bandwidth (example parameter: 8 steps) to limit the range of temporal misalignments allowed for comparison. Matrix positions outside this window are not included in the cost calculation.

[0054] (6) Local alignment support. Supports prefix / suffix skip, which means that matching can start from any position in the parent sequence and end at any position. This is suitable for scenarios where the video to be detected only rewrites and moves local segments of the parent sequence. In this case, the system outputs the optimal local alignment range and the corresponding score.

[0055] For example, the configuration parameters and core pseudocode of the above multilingual time series warping algorithm are shown below (where the values ​​are all exemplary parameters): config = { "semantic_dist_thresh": 0.40, "anchor_bonus": 0.15, "anchor_miss_penalty": 0.10, "lang_penalty_factor": 0.20, "vlm_penalty_factor": 0.30, "jump_penalty": 0.25, "max_window": 8, "max_consec_skips": 3, "max_consec_expands": 4 } def compute_multilingual_alignment(test_seq, src_seq, start_offset=0): if not test_seq or not src_seq: return 0.0, [], "ERR_EMPTY_SEQUENCE" len_t, len_s = len(test_seq), len(src_seq) dp = [[float('inf')] (len_s + 1) for _ in range(len_t + 1)] skip_count = [[0] (len_s + 1) for _ in range(len_t + 1)] expand_count = [[0] (len_s + 1) for _ in range(len_t + 1)] traceback = [[None] (len_s + 1) for _ in range(len_t + 1)] # Support skipping the start of the reference sequence (local alignment) for j in range(len_s + 1): dp[0][j] = 0.0 for i in range(1, len_t + 1): for j in range(1, len_s + 1): if abs(i - (j - start_offset)) > config["max_window"]: continue test_seg = test_seq[i-1] src_seg = src_seq[j-1] # 1. Basic cross-linguistic semantic distance dist = cosine_distance(test_seg.semantic_embedding, src_seg.semantic_embedding) # 2. Confidence penalty stacking lang_conf = getattr(test_seg, 'asr_conf', 1.0) vlm_conf = getattr(test_seg, 'vlm_conf', 1.0) dist += (1.0 - lang_conf) config["lang_penalty_factor"] dist += (1.0 - vlm_conf) config["vlm_penalty_factor"] if dist > config["semantic_dist_thresh"]: dist += 1.0 else: #3. Visual anchor point reward or downgrade if test_seg.visual_anchor_flag and src_seg.visual_anchor_flag: dist = max(0.0, dist - config["anchor_bonus"]) elif test_seg.visual_anchor_flag != src_seg.visual_anchor_flag: dist += config["anchor_miss_penalty"] # 4. State transition cost and continuity constraint match_cost = dp[i-1][j-1] + dist expand_cost = ( float('inf') if expand_count[i-1][j] >= config["max_consec_expands"] else dp[i-1][j] + dist + config["jump_penalty"] ) skip_cost = ( float('inf') if skip_count[i][j-1] >= config["max_consec_skips"] else dp[i][j-1] + dist + config["jump_penalty"] ) min_cost = min(match_cost, expand_cost, skip_cost) dp[i][j] = min_cost if min_cost == float('inf'): continue elif min_cost == match_cost: traceback[i][j] = (i-1, j-1) skip_count[i][j] = expand_count[i][j] = 0 elif min_cost == expand_cost: traceback[i][j] = (i-1, j) expand_count[i][j] = expand_count[i-1][j] + 1 else: traceback[i][j] = (i, j-1) skip_count[i][j] = skip_count[i][j-1] + 1 # Supports skipping the end of the parent text and finding the optimal local alignment endpoint. best_end_j = max( range(1, len_s + 1), key=lambda j: -dp[len_t][j] ) min_final_cost = dp[len_t][best_end_j] if min_final_cost == float('inf'): return 0.0, [], "ERR_UNREACHABLE_ALIGNMENT" score = max(0.0, 1.0 - min_final_cost / max(len_t, best_end_j)) path = extract_traceback_path(traceback, len_t, best_end_j) return score, path, "SUCCESS" As shown in the pseudocode above, in the dynamic programming solution of the two-dimensional state transition matrix, the transition cost of each matrix position is composed of the basic cross-linguistic semantic distance, confidence penalty, semantic distance penalty, anchor point reward or anchor point penalty, and jump penalty, i.e., the target transition cost. After completing the cost calculation of the entire matrix, the optimal local alignment endpoint is found by combining the parent video tail skipping mechanism to obtain the minimum final cost (min_final_cost). Then, the homology risk score between the video to be detected and the target parent video is determined based on the target transition cost. For example, the homology risk score is calculated as: score = max(0.0, 1.0 - min_final_cost / max(len_t, best_end_j)), where len_t is the length of the semantic sequence to be detected, and best_end_j is the position of the optimal local alignment endpoint, thus normalizing the score to the range of 0 to 1. The higher the score, the higher the homology risk between the video to be detected and the target parent video. In addition, if the input sequence is empty, the exception code ERR_EMPTY_SEQUENCE is returned; if all transfer paths are unreachable, the exception code ERR_UNREACHABLE_ALIGNMENT is returned so that the corresponding exception can be marked in the detection report.

[0056] After calculating the source risk score, when the source risk score meets preset conditions, such as not being lower than a preset score threshold, the system generates a structured review report, i.e., a cross-language content transfer detection report. The report includes: a comprehensive source risk score and score breakdown (semantic matching base score, anchor point reward, penalty deduction); a detailed mapping between the timeline of the video segment to be detected and the timeline of the source video segment (including text comparison between the two); anomaly codes; and manual review suggestions. For example, the evidence package data structure corresponding to the detection report can be as follows (where the values ​​are all exemplary parameters): { "report_id": "REWRITE_2026_099", "candidate_video_id": "TEST_ZH_04", "source_video_id": "BASE_EN_01", "homology_risk_score": 0.88, "score_breakdown": { "semantic_match_ratio": 0.85, "anchor_bonus_applied": 0.05, "penalty_deduction": -0.02 }, "mapped_timeline": [ { "test_time": "00:00-00:05", "source_time": "00:00-00:05", "test_text": "The male protagonist finally mustered up the courage today...", "src_text": "He finally gathered his courage..." }, { "test_time": "00:10-00:15", "source_time": "00:10-00:15", "test_text": "He encountered a terrifying behemoth!" "src_text": "The dragon suddenly attacked." } ], "manual_review_suggestion": "It is recommended to enter the manual review queue: The cross-language temporal feature mapping ratio has reached the threshold, indicating a risk of derivative rewriting and plagiarism. It is recommended to review the corresponding segment content and source correspondence." To facilitate understanding, the technical processing of this embodiment is explained below using several typical scenarios (all numerical values ​​are exemplary parameters). Scenario 1: The original English video is translated into Chinese and re-translated using a paraphrased approach: The Chinese text of the video to be detected and the English text of the parent video are completely different in wording. However, after cross-language embedding, the semantic distance between semantically similar segments in the cross-language semantic space can still reflect the content correspondence. Zero word overlap does not affect semantic matching. Scenario 2: The rewritten video deletes some narration segments from the original video: Jump state transitions skip the corresponding segments in the parent video, maintaining the connectivity of the overall path within the jump limit, so that the deletion does not interrupt the macro-temporal mapping. Scenario 3: The rewritten video adds extra narration (i.e., "adding drama"): Expand state transitions allow multiple segments of the video to correspond to a single node in the parent video, which is tolerated within the expansion limit. Scenario 4: The positions of strong transitions are consistent in both videos: Anchor bonus discount is triggered, reducing the alignment path cost at that position and enhancing the mapping weight. Scenario 5: VLM fallback description of extremely low quality (e.g., pure black screen): judged as invalid segment and filtered, not included in the scoring calculation.

[0057] It should be noted that some processing steps in the above technical solution can also be implemented using the following alternative methods, which can also achieve the purpose of this embodiment. First, the replacement of multilingual embedding models: the current multilingual pre-trained embedding models (such as mBART, XLM-R, etc.) can be replaced by multilingual speech-text joint embedding models based on contrastive learning, which support direct cross-language semantic encoding of audio waveforms without the need for ASR to text conversion, and are suitable for spoken dialects or low-quality audio scenarios. Second, the replacement of visual anchor point extraction methods: the current visual anchor point extraction based on inter-frame optical flow mutation and scene switching detection can be replaced by calling a video scene classification model to perform semantic-level labeling of specific types of high-dynamic scenes (such as fighting, explosions, and strong emotional expressions), providing more content-oriented anchor point signals. Third, the replacement of the weak text fallback mechanism: The VLM objective description fallback method can be replaced by extracting visual semantic feature vectors from video keyframes and directly calculating cross-modal distance with the image features of the image-text parent library (the parent library needs to maintain image feature indexes at the same time), which is suitable for scenarios where the parent library has identifiable images. Fourth, the replacement of the temporal warping algorithm: The current dynamic programming two-dimensional cost matrix scheme can be replaced by a sequence contrast learning scheme based on bidirectional LSTM or Transformer encoders, which directly outputs sequence similarity through model training, reducing the need for manually designed constraint parameters, but requires sufficient labeled samples to support training. Fifth, the replacement of the candidate parent library recall index: The HNSW approximate nearest neighbor index can be replaced by a vector retrieval scheme based on inverted index (IVF), or, when the candidate set is small, directly perform full calculation of accurate cosine similarity, and select an appropriate indexing strategy according to the actual size of the parent library.

[0058] As can be seen, this application uses a multilingual pre-trained embedding model to directly project texts from different languages ​​into a cross-lingual semantic space, rather than performing explicit machine translation first and then comparing them. Therefore, it avoids the error propagation problem caused by the instability of machine translation quality to a certain extent, and makes cross-lingual expressions with similar semantics (including the case of deep paraphrasing) still comparable in distance calculation. Even if the text of the video to be detected and the text of the original video are in different languages ​​and the literal overlap is close to zero, the semantic distance between the two can still be measured in a unified semantic space, which helps to reduce the false negative rate in cross-lingual rewriting and copying scenarios. Because this application introduces a two-way reward and one-way degradation mechanism for visual anchor points in temporal regularization, when the video to be detected and the parent video have structural consistency at strong transition positions, the alignment path weight at that position is strengthened. When the image structure is inconsistent, even if the text semantics are similar, the alignment cost at that position increases accordingly, thereby reducing the cross-language mis-adsorption problem caused by short texts such as common greetings to a certain extent. Furthermore, through the joint restriction of continuous jump / expansion constraints and the maximum temporal misalignment window, extreme temporal drift is prevented within the elastic range that allows for text simplification or expansion in the video to be detected, ensuring that the macro-mapping path maintains reasonable coherence. Based on this, a homology risk score is determined based on the target transfer cost, and a cross-language plagiarism detection report is generated. This report can output structured temporal mapping reference information for manual review, providing an objective and traceable basis for judging cross-language rewriting and plagiarism behavior.

[0059] See Figure 5 As shown, this application discloses a cross-language video transfer detection device, comprising: The text confidence acquisition module 11 is used to perform time-series slicing of the video to be detected, determine the first text language corresponding to the video to be detected, and the text to be detected and the text confidence corresponding to each slice to be detected. The first sequence acquisition module 12 is used to generate a target semantic vector corresponding to the slice to be detected based on the text to be detected using a target embedding model if the text confidence meets the preset confidence conditions, perform visual feature analysis on the slice to be detected to determine the target visual anchor label, and determine the semantic sequence to be detected based on the target semantic vector; the target semantic vector is a cross-language semantic vector. The second sequence acquisition module 13 is used to determine the target parent feature of the target parent video corresponding to the video to be detected from the target parent feature library based on the target semantic vector, and to determine the target parent semantic sequence based on the target parent feature; The detection report generation module 14 is used to determine the target transfer cost between each slice to be detected and each target parent slice based on the target semantic sequence and the target parent semantic sequence using target constraints, and to determine the homology risk score between the video to be detected and the target parent video based on the target transfer cost, so as to generate a cross-language transfer detection report based on the target transfer cost and the homology risk score.

[0060] As can be seen, this application projects texts from different languages ​​directly into a cross-lingual semantic space through a pre-trained multilingual embedding model, rather than performing explicit machine translation and then comparing them. Therefore, it avoids the error propagation problem caused by the instability of machine translation quality to a certain extent, and makes cross-lingual expressions with similar semantics (including deep paraphrasing) still comparable in distance calculation. Even if the text of the video to be detected and the text of the original video are in different languages ​​and the literal overlap is close to zero, the semantic distance between the two can still be measured in a unified semantic space, which helps to reduce the false negative rate in cross-lingual rewriting and copying scenarios. Furthermore, this application introduces target constraints, including confidence constraints, cross-lingual semantic distance constraints, and visual anchor label constraints, when determining the target transfer cost. On the one hand, it utilizes visual anchor signals such as strong transitions in video frames to assist in constraining cross-lingual semantic alignment. When the video to be detected and the target parent video have structural consistency at strong transition positions, the alignment path weight at that position is strengthened. When the frame structures are inconsistent, even if the text semantics are similar, the alignment cost at that position increases accordingly, thereby reducing the cross-lingual mis-adsorption problem caused by short texts such as common greetings to a certain extent. On the other hand, it uses confidence constraints to increase the matching path cost of low-reliability segments, avoiding the excessive impact of mismatches of low-confidence texts on the final score. Based on this, it determines the homology risk score based on the target transfer cost and generates a cross-lingual plagiarism detection report, which can output structured temporal mapping reference information for manual review, providing an objective and traceable basis for judging cross-lingual rewriting and plagiarism behavior.

[0061] In one specific embodiment, the device may further include: The master text slicing module is used to acquire each master text video, perform time-series slicing on the master text video to obtain corresponding master text slices, determine the second copywriting language corresponding to the master text video, and the master text text corresponding to each master text slice, and obtain the confidence level of the master text text corresponding to the master text text. The parent text semantic vector acquisition module is used to determine the parent text structured text corresponding to the parent text slice based on the parent text corresponding to the parent text slice, and to generate the parent text semantic vector corresponding to the parent text slice based on the parent text and the parent text structured text using the target embedding model; the parent text semantic vector is a cross-language semantic vector; The parent label acquisition module is used to perform visual feature analysis on the parent slice to determine the parent visual anchor point label corresponding to the parent slice; The parent feature entry module is used to write the parent semantic vector, parent text confidence, second copy language corresponding to the parent text, and parent visual anchor tag corresponding to the parent slice of the parent video into the target parent feature library as parent features of the parent video.

[0062] In one specific embodiment, the device may further include: The parent slice feature update module is used to determine whether the parent video has undergone content revision based on the version hash value corresponding to each parent slice, and to perform incremental feature update on the parent slice when the version hash value corresponding to the parent slice indicates that the parent video has undergone content revision.

[0063] In one specific embodiment, the device may further include: The target confidence determination module is used to, if neither the first text confidence nor the second text confidence corresponding to the slice to be detected meets the preset confidence condition, call the target visual language model to generate the target description text corresponding to the target keyframe of the slice to be detected and determine the corresponding target confidence; wherein, the text to be detected includes the speech text to be detected and the screen text to be detected; the first text confidence is the text confidence corresponding to the speech text to be detected, and the second text confidence is the text confidence corresponding to the screen text to be detected; Accordingly, the first sequence acquisition module 12 may specifically include: The semantic vector acquisition unit is used to generate the target semantic vector corresponding to the slice to be detected based on the target description text using the target embedding model.

[0064] In one specific implementation, the test report generation module 14 may include: The confidence penalty term determination unit is used to determine the target confidence penalty term between the slice to be detected and each of the target parent slices if the target semantic vector corresponding to the slice to be detected is a semantic vector generated based on the target description text, by using confidence constraints based on the target confidence corresponding to the target description text.

[0065] In one specific implementation, the test report generation module 14 may include: The semantic distance penalty term determination unit is used to determine the semantic distance penalty term between the target semantic vector corresponding to the slice to be detected and the parent semantic vector corresponding to the target parent slice in the target semantic space if the cosine distance in the target semantic space is greater than a preset distance threshold.

[0066] In one specific implementation, the test report generation module 14 may include: Anchor point reward item determination unit is used to determine the anchor point reward item in the target semantic space based on the cosine distance between the target semantic vector corresponding to the slice to be detected and the parent semantic vector corresponding to the parent slice in the target semantic space if the slice to be detected has the target visual anchor point label and the parent visual anchor point label has the parent visual anchor point label. Anchor point penalty term determination unit is used to determine the anchor point penalty term between the slice to be detected and the target parent slice by using the second visual anchor point label constraint if the slice to be detected does not have the target visual anchor point label and the target parent slice has the parent visual anchor point label, or if the slice to be detected has the target visual anchor point label and the target parent slice does not have the parent visual anchor point label. The visual anchor label constraint includes the first visual anchor label constraint and the second visual anchor label constraint.

[0067] 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.

[0068] 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 cross-language video transport detection method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be a computer.

[0069] 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.

[0070] 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 an operating system 221, computer programs 222, etc., and the storage method can be temporary storage or permanent storage.

[0071] 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 cross-language video transport 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.

[0072] 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 cross-language video transfer detection method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0073] 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.

[0074] 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.

[0075] 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.

[0076] 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.

[0077] 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 cross-language copied videos, characterized in that, include: The video to be tested is divided into time-series slices to determine the first text language corresponding to the video to be tested, as well as the text to be tested and the text confidence level corresponding to each slice to be tested. If the text confidence meets the preset confidence conditions, the target embedding model is used to generate the target semantic vector corresponding to the slice to be detected based on the text to be detected. Visual feature analysis is performed on the slice to be detected to determine the target visual anchor label, and the target semantic vector is used to determine the semantic sequence to be detected; the target semantic vector is a cross-language semantic vector. Based on the target semantic vector, the target parent video corresponding to the video to be detected is determined from the target parent feature library, and the target parent semantic sequence is determined based on the target parent features; Using target constraints, the target transfer cost between each slice to be detected and each target parent slice is determined based on the semantic sequence to be detected and the semantic sequence to be the target source. Based on the target transfer cost, the homology risk score between the video to be detected and the target parent video is determined, and a cross-language transfer detection report is generated based on the target transfer cost and the homology risk score.

2. The cross-language video transfer detection method according to claim 1, characterized in that, Before determining the target parent video features corresponding to the target parent video of the video to be detected from the target parent feature library based on the target semantic vector, the method further includes: Each source video is acquired, and the source videos are slicing in a time sequence to obtain corresponding source slices. The second text language corresponding to each source video is determined, as well as the source text corresponding to each source slice, and the source text confidence score corresponding to the source text is obtained. Based on the parent text corresponding to the parent text slice, the parent text structured text corresponding to the parent text slice is determined, and the parent text semantic vector corresponding to the parent text slice is generated using the target embedding model based on the parent text and the parent text structured text; the parent text semantic vector is a cross-language semantic vector; Visual feature analysis is performed on the parent tissue slices to determine the parent tissue visual anchor point labels corresponding to the parent tissue slices; The semantic vector of the source video corresponding to each source video slice, the confidence score of the source text, the second copywriting language corresponding to the source text, and the source visual anchor tag corresponding to the source video slice are written into the target source video feature library as source video features of the source video.

3. The cross-language video transfer detection method according to claim 2, characterized in that, Also includes: Based on the version hash value corresponding to each of the parent video slices, it is determined whether the parent video has undergone content revision. When the version hash value corresponding to the parent video slice indicates that the parent video has undergone content revision, the parent video slice is updated with incremental features.

4. The cross-language video transfer detection method according to claim 1, characterized in that, Also includes: If neither the first text confidence score nor the second text confidence score corresponding to the slice to be detected meets the preset confidence condition, then the target visual language model is invoked to generate the target description text corresponding to the target keyframe of the slice to be detected and determine the corresponding target confidence score; wherein, the text to be detected includes the speech text to be detected and the screen text to be detected; the first text confidence score is the text confidence score corresponding to the speech text to be detected, and the second text confidence score is the text confidence score corresponding to the screen text to be detected; Accordingly, the step of generating the target semantic vector corresponding to the slice to be detected based on the text to be detected using the target embedding model includes: The target semantic vector corresponding to the slice to be detected is generated based on the target description text using a target embedding model.

5. The cross-language video transfer detection method according to claim 4, characterized in that, The step of determining the target transfer cost between each slice to be detected and each target parent slice based on the target semantic sequence and the target parent semantic sequence using target constraints includes: If the target semantic vector corresponding to the slice to be detected is a semantic vector generated based on the target description text, then the target confidence penalty term between the slice to be detected and each of the target parent slices is determined by using confidence constraints based on the target confidence corresponding to the target description text.

6. The cross-language video transfer detection method according to claim 1, characterized in that, The step of determining the target transfer cost between each slice to be detected and each target parent slice based on the target semantic sequence and the target parent semantic sequence using target constraints includes: If the cosine distance between the target semantic vector corresponding to the slice to be detected and the parent semantic vector corresponding to the target parent slice in the target semantic space is greater than a preset distance threshold, then a semantic distance penalty term between the slice to be detected and the target parent slice is determined based on cross-language semantic distance constraints.

7. The cross-language video transfer detection method according to claim 1, characterized in that, The step of determining the target transfer cost between each slice to be detected and each target parent slice based on the target semantic sequence and the target parent semantic sequence using target constraints includes: If the slice to be detected has the target visual anchor label, and the target parent slice has the parent visual anchor label, then the anchor reward item is determined in the target semantic space based on the cosine distance between the target semantic vector corresponding to the slice to be detected and the parent semantic vector corresponding to the target parent slice in the target semantic space using the first visual anchor label constraint. If the slice to be detected does not have the target visual anchor label and the target parent slice has the parent visual anchor label, or if the slice to be detected has the target visual anchor label and the target parent slice does not have the parent visual anchor label, then the anchor penalty term between the slice to be detected and the target parent slice is determined by using the second visual anchor label constraint. The visual anchor label constraint includes the first visual anchor label constraint and the second visual anchor label constraint.

8. A cross-language video transfer detection device, characterized in that, include: The text confidence acquisition module is used to perform time-series slicing of the video to be detected, determine the first text language corresponding to the video to be detected, and the text to be detected and the text confidence corresponding to each slice to be detected; The first sequence acquisition module is used to, if the text confidence meets the preset confidence conditions, generate the target semantic vector corresponding to the slice to be detected based on the text to be detected using the target embedding model, perform visual feature analysis on the slice to be detected to determine the target visual anchor label, and determine the semantic sequence to be detected based on the target semantic vector; the target semantic vector is a cross-language semantic vector; The second sequence acquisition module is used to determine the target parent feature of the target parent video corresponding to the video to be detected from the target parent feature library based on the target semantic vector, and to determine the target parent semantic sequence based on the target parent feature; The detection report generation module is used to determine the target transfer cost between each slice to be detected and each target parent slice based on the target semantic sequence and the target parent semantic sequence using target constraints, and to determine the homology risk score between the video to be detected and the target parent video based on the target transfer cost, so as to generate a cross-language transfer detection report based on the target transfer cost and the homology risk score.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the cross-language video transfer detection method 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 cross-language video transfer detection method as described in any one of claims 1 to 7.

Images