A video detection method, an electronic device, a storage medium and a program product

Abstract

The application provides a video detection method, an electronic device, a storage medium and a program product. The method comprises the following steps: determining a to-be-detected original video and prompt word information corresponding to the to-be-detected original video, determining a prompt word image according to the prompt word information, generating a special detection route for the to-be-detected original video according to the prompt word image, and performing detection processing on the to-be-detected original video based on at least one of a target detection plug-in set, a key detection object set, a target weight distribution strategy, a target gate trigger rule and an eventization review strategy included in the special detection route, to obtain a target detection result of the to-be-detected original video. In this way, a special detection route is generated for each to-be-detected original video according to the corresponding prompt word image, so that the detection processing of the to-be-detected original video is realized based on the special detection route, the video detection is more targeted and more comprehensive, and the detection accuracy can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a video detection method, electronic device, storage medium, and program product. Background Technology

[0002] With the rapid popularization of AI-native video generation technologies such as text-based videos, image-based videos, dialogue-driven videos, digital human videos, and multimodal videos, the demand for automated consistency detection between videos and prompts continues to grow in business scenarios such as advertising material production, model performance evaluation, content pre-screening, pre-launch quality inspection, and batch material acceptance. Current mainstream detection solutions have significant limitations and are difficult to adapt to the diverse and highly constrained AI video quality inspection requirements. Existing common video detection solutions score the original video as a whole based on a multimodal model, directly outputting the similarity score between the original video and the prompts, failing to distinguish between specific issues such as missing main elements, chronological errors, and missing dialogue. In other words, existing video detection methods only score the overall similarity between the original video and the prompts, resulting in a simplistic and inaccurate approach that cannot achieve targeted detection for different types of original videos.

[0003] Therefore, existing video detection methods suffer from the problems of being simple to detect and having low accuracy. Summary of the Invention

[0004] This application provides a video detection method, electronic device, storage medium, and program product that can generate a dedicated detection route for each native video to be detected based on the corresponding prompt word profile, and perform detection processing on the native video to be detected based on the dedicated detection route to obtain the detection result of the native video to be detected. In this way, the detection of native videos is more targeted, more comprehensive, and more accurate.

[0005] To address the aforementioned technical problems, in a first aspect, embodiments of this application provide a video detection method. The method includes: determining a native video to be detected and corresponding prompt word information; determining a prompt word profile based on the prompt word information; generating a target detection route for the native video to be detected based on the prompt word profile, the target detection route including a dedicated detection route, the dedicated detection route including at least one of a video detection plugin set, a set of key video detection objects, a video consistency detection weight allocation strategy, and a video detection failure gating trigger rule; and performing detection processing on the native video to be detected based on the dedicated detection route to obtain the target detection result of the native video to be detected.

[0006] By employing the above technical solution, the native video to be detected and its corresponding prompt word information are determined, and a prompt word profile is determined based on the prompt word information. A dedicated detection route is generated for the native video to be detected based on the prompt word profile. The native video to be detected is then processed based on at least one of the following: a set of target detection plugins, a set of key detection objects, a target weight allocation strategy, a target gating triggering rule, and an event-based review strategy, to obtain the target detection result for the native video to be detected. In this way, a dedicated detection route is generated for each native video to be detected based on its corresponding prompt word profile, enabling detection processing of the native video to be detected based on the dedicated detection route. Furthermore, by determining at least one of the following: a set of target detection plugins, a set of key detection objects, a target weight allocation strategy, a target gating triggering rule, and an event-based review strategy for the native video to be detected, targeted detection of the native video to be detected can be achieved, making the detection more targeted and comprehensive, and effectively improving detection accuracy.

[0007] In one possible implementation of the first aspect above, determining the prompt word profile based on the prompt word information includes: performing feature extraction processing on the prompt word information to obtain core verification elements; and performing pattern recognition processing on the prompt word information based on a target pattern recognition method to obtain a prompt word scene pattern, wherein the target pattern recognition method includes at least one of the following processing methods: dictionary extraction processing method, regular pattern matching method, and heuristic multidimensional feature scoring method; and obtaining the prompt word profile based on the core verification elements and the prompt word scene pattern.

[0008] By adopting the above technical solution, the core verification elements are obtained by combining dictionary extraction, regular expression matching, and heuristic multidimensional scoring composite recognition mechanism with feature extraction, and prompt word profiles are jointly generated. This overcomes the shortcomings of easy misjudgment in single keyword matching, high cost of large model classification, and uncontrollable black box.

[0009] In one possible implementation of the first aspect above, the method further includes: determining the video profile of the native video to be detected and determining the video detection rule configuration information; generating a dedicated detection route for the target detection route of the native video to be detected based on the prompt word profile, including: generating a dedicated detection route for the native video to be detected according to the hierarchical detection pipeline, based on the prompt word profile, the video profile and the video detection rule configuration information.

[0010] By adopting the above technical solution, two types of scene representations, namely prompt word profiles and video profiles, are constructed simultaneously and combined with business rule configurations to jointly generate exclusive detection routes. This breaks the rigid logic of traditional detection systems that reuse the same set of fixed plugins, slots, weights, and thresholds for all videos. Relying on the semantic constraints of prompt words and the physical morphological features of videos, the detection strategy is adjusted by layering and superimposing on two dimensions. It automatically distinguishes the tightness of verification for different content and carrier materials, schedules special plugins as needed, and dynamically switches sampling strategies and event review switches. This avoids the missed detections caused by insufficient verification of common and complex scenarios by the general baseline, and also prevents excessive false positives caused by applying strict rules to simple materials. At the same time, the route generation logic relies on predefined heuristic rules to achieve lightweight calculations, without repeatedly calling large models, ensuring the processing efficiency of batch quality inspection scenarios, providing customized execution scheduling basis for the layered detection pipeline, and realizing the adaptive detection capability of "one material, one route".

[0011] In one possible implementation of the first aspect described above, the prompt word profile includes prompt word scene patterns, the video profile includes video scene patterns, and the layered detection pipeline includes a routing layer. Then, according to the layered detection pipeline, based on the prompt word profile, video profile, and rule configuration information, a dedicated detection route for the native video to be detected is generated. This includes: according to the routing layer, if a target scene combination pattern can be generated based on the prompt word scene pattern and the video scene pattern, then a dedicated detection route for the native video to be detected is generated based on the target scene combination pattern; if a target scene combination pattern cannot be generated based on the prompt word scene pattern and the video scene pattern, then a first detection route is generated based on the prompt word scene pattern, a second detection route is generated based on the video scene pattern, and the second detection route is updated based on the first detection route according to the rule configuration information to generate a dedicated detection route for the native video to be detected.

[0012] Using the above technical solution, the routing layer distinguishes between two routing generation logics: combined pattern matching and single-pattern step-by-step fusion. It prioritizes matching the exclusive strategies for composite scenarios where prompt words and video scenarios are superimposed. When there is no predefined combination, it generates two sets of basic routes based on two types of profiles, with the prompt word-side route as the main one and business rules as the constraint to complete parameter overwriting and fusion. This can reuse the fine-grained loose / tight verification strategies preset for special material combinations, effectively reducing the probability of false interception of composite materials such as voice-over comments and screen recording tutorials. It can also ensure that the core requirements of prompts take priority and that the characteristics of video carriers are adapted synchronously when there are no exclusive combination rules. It takes into account the completeness of the strategy and the flexibility and compatibility, and avoids the imbalance of verification scale caused by generating routes solely based on content or video dimensions. This provides a unique and complete scheduling strategy for the layered pipeline to be adapted to the current material.

[0013] In one possible implementation of the first aspect above, the method further includes: generating a video detection strategy for the native video to be detected based on the prompt word profile; performing detection processing on the native video to be detected based on the target detection route to obtain the target detection result of the native video to be detected, including: performing detection processing on the native video to be detected based on the video detection strategy according to the dedicated detection route in a hierarchical detection pipeline to obtain the target detection result of the native video to be detected.

[0014] The above technical solution first generates a video detection strategy independently based on the prompt word profile, and then the layered detection pipeline performs video detection based on the target detection route scheduling and in conjunction with the detection strategy, thereby achieving decoupling and linkage between the routing scheduling logic and the specific verification strategy.

[0015] In one possible implementation of the first aspect above, when the dedicated detection route includes a set of video detection plugins, the hierarchical detection pipeline includes a dedicated plugin layer. Following the hierarchical detection pipeline, based on the dedicated detection route and a video detection strategy, the native video to be detected is processed to obtain the target detection result of the native video. This includes: according to the dedicated plugin layer, based on the target detection plugin set included in the video detection plugin set, processing the native video to be detected according to the video detection strategy to obtain the plugin detection result of the native video. The target detection plugin set includes at least one detection plugin among multi-camera narrative consistency plugin, identity consistency plugin, text detection plugin, audio-visual consistency plugin, speech recognition verification plugin, and complex event understanding plugin. Based on the target adapter corresponding to the target detection plugin in the target detection plugin set, the plugin detection result is format-unified according to the plugin signal standard protocol to obtain the normalized plugin detection result of the native video to be detected, which serves as the target detection result of the native video to be detected.

[0016] By adopting the above technical solution, various specialized testing plug-ins are scheduled on demand through routing for targeted testing. The adapter is used in conjunction with a unified plug-in signal standard protocol to standardize the original results of heterogeneous plug-ins, solving the problem of inconsistent output formats and dimensions of different plug-ins. This decouples the plug-in layer from the downstream testing and adjudication logic, facilitating plug-in iterative expansion. The standardized and normalized plug-in testing results can stably participate in subsequent slot verification, hard gating judgment, and score fusion, avoiding the distortion of results caused by direct fusion of heterogeneous testing data, and improving the stability, consistency, and traceability of the testing results.

[0017] In one possible implementation of the first aspect above, when the dedicated detection route includes a set of key detection objects, the hierarchical detection pipeline includes a key detection object layer. According to the hierarchical detection pipeline, based on the dedicated detection route and the video detection strategy, the native video to be detected is processed to obtain the target detection result of the native video to be detected. This includes: according to the key detection object layer, based on the video detection strategy and the set of key detection objects for the video, detecting whether the native video to be detected meets the key detection object requirements, obtaining the key detection object detection result of the native video to be detected, and using it as the target detection result of the native video to be detected.

[0018] By adopting the above technical solution, a key detection object layer is added to the layered detection pipeline. A dedicated detection route accurately identifies the set of key detection objects corresponding to the current scenario, and the compliance of key detection objects is specifically verified according to the video detection strategy. The core constraints of the prompt words are focused on to carry out accurate verification, avoiding invalid verification and excessive constraints on non-key elements. By having a dedicated layer at an independent level to complete the compliance judgment of hard elements such as core subjects, key actions, and core quantities, the effective implementation of strong business constraints is prioritized. This avoids the problem of core key defects being covered by secondary dimension scores from the source of detection, effectively improving the detection rate of key violations and defects. At the same time, it simplifies invalid detection logic, reduces the probability of false judgment, and further ensures the accuracy and business adaptability of the final detection results.

[0019] In one possible implementation of the first aspect above, where the dedicated detection route includes a target weight allocation strategy, the hierarchical detection pipeline includes a consistency detection layer. Following the hierarchical detection pipeline, based on the dedicated detection route and a video detection strategy, the native video to be detected is processed to obtain a target detection result for the native video. This includes: generating a video frame extraction strategy for the native video to be detected based on the video detection strategy, according to the consistency detection layer; performing frame extraction processing on the native video to be detected based on the video frame extraction strategy to obtain a sampled frame sequence; performing consistency verification processing on the sampled frame sequence based on the video detection strategy to obtain an initial consistency detection result for the native video to be detected, wherein the consistency verification processing includes at least one of semantic consistency detection processing, entity consistency detection processing, spatial consistency detection processing, temporal consistency detection processing, style consistency detection processing, and logical consistency detection processing; and adjusting the initial consistency detection result of the native video to be detected based on the key detection object detection result and the video detection strategy to obtain a target consistency detection result for the native video to be detected, which is then used as the target detection result for the native video to be detected.

[0020] By adopting the above technical solution, the consistency detection layer relies on the dynamic weight allocation strategy of the target detection route and the video detection strategy to adaptively generate a frame extraction strategy adapted to the current video scene and complete fine sampling. Through multi-dimensional consistency verification of semantics, entity, space, time, style, and logic, a comprehensive content compliance judgment is achieved. At the same time, the initial consistency score is corrected and adjusted by combining the hard verification results of the key detection objects in the previous stage. On the one hand, the scene-based adaptive frame extraction strategy improves the sampling coverage of key change frames and problem periods, avoiding missed detections and redundant calculations caused by fixed frame extraction. On the other hand, the dynamic dimension weight is used to adapt to the quality inspection focus of different scenarios, taking into account the rigid constraints of global content consistency and core key elements, correcting the deviation problem of simple dimension scoring. It effectively solves the problems of poor scene adaptability, overwhelming of key defects, and scoring distortion caused by traditional fixed weight and uniform sampling strategies. The final consistency detection result has global integrity, local accuracy, and scene adaptability, which significantly improves the accuracy and robustness of video consistency detection.

[0021] In one possible implementation of the first aspect above, a video frame extraction strategy for the native video to be detected is generated according to the consistency detection layer, including: generating a video frame extraction strategy for the native video to be detected based on the prompt word profile and the video detection strategy according to the consistency detection layer.

[0022] By adopting the above technical solution, and combining prompt word profiling with video detection strategy to jointly generate scene-adaptive video frame extraction strategy, the limitations of traditional fixed uniform frame extraction are overcome. Based on the core verification focus represented by prompt words and the current scene detection rules, the video frame extraction strategy for the original video to be detected can be adaptively generated.

[0023] In one possible implementation of the first aspect above, when the dedicated detection route includes target gating triggering rules, the layered detection pipeline includes a hard gating layer. According to the layered detection pipeline, the native video to be detected is processed based on the dedicated detection route and the video detection strategy to obtain the target detection result of the native video to be detected. This includes: determining the hard gating detection result of the native video to be detected based on at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result, according to the hard gating layer. If it is determined that the native video to be detected triggers hard gating based on the hard gating detection result and the video detection strategy, then the target detection result of the native video to be detected is determined to be a detection failure or the detection information of the native video to be detected is limited to below a preset upper limit to obtain the target detection result of the native video to be detected.

[0024] By adopting the above technical solution, the target gating trigger rules of the dedicated detection route are relied upon at the hard gating layer. The normalized plug-in detection results, key detection object detection results and target consistency detection results are integrated for multi-source joint judgment to achieve rigid interception and verification of fatal defects in videos. Compared with the traditional single score threshold judgment method, this mechanism can accurately trigger hard gating by comprehensively considering various fatal problems such as abnormal plug-in, missing core elements, and serious non-compliance of dimensional consistency. Once hard gating is triggered, the detection failure can be directly judged or the score of the detection result can be capped. This effectively prevents the problem of missed judgment caused by serious errors in the subject, missing key processes, and structural violations being covered by the overall average score. The rigid gating mechanism safeguards the bottom line of video content compliance. At the same time, it forms a layered adjudication system with soft and hard coordination with the front-end flexible consistency scoring, which significantly improves the recognition accuracy of high-risk violation samples and the reliability and rigor of the detection system.

[0025] In one possible implementation of the first aspect above, the dedicated detection route also includes a soft fusion strategy. If it is determined that the native video to be detected has not triggered hard gating, the layered detection pipeline includes a soft fusion layer. According to the layered detection pipeline, the native video to be detected is processed based on the dedicated detection route and the video detection strategy to obtain the target detection result of the native video to be detected. This includes: according to the soft fusion layer, the comprehensive detection result of the native video to be detected is determined based on at least one of the normalization plugin detection result, the key detection object detection result, and the target consistency detection result, based on the video detection strategy, and is used as the target detection result of the native video to be detected.

[0026] By adopting the above technical solution, without triggering hard gating, a soft fusion strategy corresponding to the dedicated detection route is loaded through the soft fusion layer. This allows for the integration of multi-source information, including the normalized plug-in detection results, the key detection object detection results, and the target consistency detection results. The final comprehensive detection result is generated by comprehensively considering specific detail deviations, the compliance of core elements, and the overall six-dimensional consistency level. Compared with the traditional single-dimensional weighted scoring mechanism, this fusion method can take into account both the compliance of rigid key indicators and the differentiated deduction of minor flaws. It can achieve refined quantitative evaluation of non-fatal, local, and minor deviations, avoiding the problem of black-and-white and coarse-grained detection results. Together with the aforementioned rigid gating interception mechanism, it forms a collaborative adjudication system that is both hard-bottom-line and soft-refined. This system not only safeguards the bottom line of content quality inspection but also significantly improves the differentiation between good and bad normal samples, making the final detection result more delicate, objective, and in line with human review standards.

[0027] In one possible implementation of the first aspect above, where the dedicated detection route also includes a review strategy, the hierarchical detection pipeline includes a review layer. According to the hierarchical detection pipeline, the native video to be detected is processed based on the dedicated detection route to obtain the target detection result of the native video to be detected. This includes: when it is determined that the native video to be detected meets the target conditions, the review layer generates a review result of the native video to be detected based on the prompt word information and the native video to be detected, which is used as the target detection result of the native video to be detected. The target conditions include at least one of the following: the comprehensive detection result of the native video to be detected is within the grayscale range, and the native video to be detected is a complex scene video.

[0028] By adopting the above technical solution, a review layer is added to the layered detection pipeline, and an event-based review mechanism is adaptively activated based on the review strategy of the dedicated detection route. This mechanism performs refined event-based review on high-difficulty video samples whose comprehensive detection results are in the grayscale range or belong to complex scenes. Review results are generated by combining the integrity constraints of prompt words and the temporal content of the video. Compared to the traditional method of relying solely on global weighted scoring, this mechanism can perform segmented, event-level fine-grained verification of fuzzy boundary samples and samples from complex scenes with strong temporal sequences. This compensates for the accuracy deficiencies of global scoring in judging detailed temporal sequences and staged compliance, effectively solving the problems of inaccurate judgment of grayscale samples and difficulty in identifying defects in complex scenes. It constructs a three-level progressive adjudication system of hard gate control as a fallback, soft fusion scoring, and event review for fine-tuning, significantly improving the detection accuracy of boundary samples and complex materials. Simultaneously, it generates structured event review evidence, providing reliable data support for manual review, result tracing, and model iteration.

[0029] In one possible implementation of the first aspect described above, the normalized plugin detection result includes a detection result processing intent. The method further includes: if the detection result processing intent indicates that the normalized plugin detection result is a policy signal, then the normalized plugin detection result is returned to the routing layer so that the routing layer updates the dedicated detection route and video detection policy according to the normalized plugin detection result; if the detection result processing intent indicates that the normalized plugin detection result is a scoring signal, then the normalized plugin detection result is sent to the soft fusion layer so that the soft fusion layer determines the comprehensive detection result of the native video to be detected according to at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result; if the detection result processing intent indicates that the normalized plugin detection result is a gating signal, then the normalized plugin detection result is sent to the hard gating layer so that the hard gating layer determines the hard gating detection result of the native video to be detected according to at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result; if the detection result processing intent indicates that the normalized plugin detection result is an evidence signal, then the normalized plugin detection result is used as the target detection result of the native video to be detected.

[0030] By employing the above technical solution, the processing intent of the detection results is embedded in the normalized plug-in detection results, enabling precise four-way routing and scheduling of standardized plug-in signals. Based on the different semantic intents of strategy signals, scoring signals, gating signals, and evidence signals, the signals are respectively transmitted back to the routing layer, sent to the soft fusion layer, sent to the hard gating layer, and retained as verification evidence. This achieves intelligent classification and precise hierarchical flow of the normalized results of heterogeneous plug-ins. Strategy signals can dynamically iterate on dedicated detection routes and video detection strategies, achieving adaptive optimization of the detection process. Scoring signals can participate in multi-source comprehensive scoring to improve the granularity of the results. Gating signals can participate in the rigid interception and judgment of fatal defects. Evidence signals can be used to accumulate structured traceability materials. Compared to the traditional, crude mode of single plug-in result output and unified processing, this mechanism enables different attribute detection signals to perform their respective functions and be reused in layers. It completely solves the problems of mixed processing of multiple plug-in results, ambiguous signal effects, and the inability to adaptively update strategies in real time. It constructs a refined detection reasoning system with layered plug-in output and bidirectional linkage, significantly improving the intelligent scheduling capability, detection accuracy, and result traceability of the entire pipeline.

[0031] In one possible implementation of the first aspect above, the method further includes: obtaining the manual review results of each native video to be detected obtained by the user based on the prompt word information corresponding to each of the multiple native videos to be detected; performing bucketing processing on the multiple native videos to be detected and their corresponding manual review results according to the prompt word profiles corresponding to each native video to be detected, to obtain content buckets of different types; determining the detection route adjustment parameters corresponding to each content bucket according to the manual review results of the native videos to be detected included in each content bucket; updating the dedicated detection route of the native videos to be detected of the same type as the content bucket according to the detection route adjustment parameters corresponding to the content bucket, so as to perform detection processing on the native videos to be detected based on the adjusted dedicated detection route.

[0032] By adopting the above technical solution, the manual review results of multiple batches of videos are collected, and the samples are refined into independent content buckets based on the prompt word profile to form different scene types. Based on the real human feedback of each content bucket, the exclusive routing adjustment parameters are obtained by statistical fitting, and the exclusive detection routes for the same type of scene are iteratively updated. This realizes the upgrade of the detection strategy from fixed preset rules to data-driven and continuously adaptive iteration. It can dynamically correct the plug-in scheduling logic, key slot constraints, dimension weights and gate thresholds in different scenarios according to the real review standards, continuously reduce the deviation between machine judgment and manual review, and solve the defects of traditional detection strategies that are rigid and cannot be adaptively optimized with the iteration of business materials and changes in review scale.

[0033] Secondly, embodiments of this application provide an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the video detection method provided by any of the implementations of the first aspect above.

[0034] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the video detection method provided by any of the implementations of the first aspect described above.

[0035] Fourthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the video detection method provided by any of the implementations of the first aspect described above.

[0036] It is understood that the beneficial effects of the second to fourth aspects mentioned above can also be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0037] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0038] Figure 1 This is a schematic flowchart of a video detection method provided in an embodiment of this application;

[0039] Figure 2 Another flowchart illustrating the video detection method provided in this application embodiment;

[0040] Figure 3 A schematic flowchart illustrating the plug-in signal standardization processing procedure provided in this application embodiment;

[0041] Figure 4 This is a schematic diagram of a process for detecting and processing video of a mechanical deformation scene, provided in an embodiment of this application.

[0042] Figure 5 A schematic diagram of the structure of a video detection system provided in this application embodiment. Detailed Implementation

[0043] With the development of AI native videos and the current demand for them, the detection of AI native videos is no longer satisfied with simple image quality scoring. Instead, it is necessary to verify the matching degree of various constraints and prompts in the native video, such as the subject, scene, action, quantity, style, timing, logic, dialogue, audio-visual relationship, text and image, and identity continuity.

[0044] For example, an AI-native video might be clear in terms of picture quality, but it has the following problems: the prompt requires "shopping mall voiceover," but the actual video scene is an office or study; the prompt requires "two cats talking," but only one cat appears or the dialogue is misaligned; the prompt requires the mechanical subject to gradually transform into a robot, but it is actually just a light effect that blocks the view and then cuts directly to the final form; the prompt requires the appearance of a specific text, but the actual Optical Character Recognition (OCR) results in missing text or many typos; the video seems reasonable overall, but the low scores are concentrated in a few key frames or key time windows.

[0045] In existing technologies, one method for AI-native video consistency detection involves direct scoring using a single multimodal model. This method inputs prompts and video frames into similarity models such as visual language models, Contrastive Language-Image Pre-Training (CLIP), or other large multimodal models. The model then directly outputs a similarity score, a descriptive judgment, or a pass / fail conclusion. While this method is relatively simple to implement and can cover some open semantic judgments, complex prompts are compressed into a single score, making it difficult to analyze the reasons for failure. The model uses the same judgment logic for videos of different scene types, lacking scene adaptation and resulting in low detection accuracy. Furthermore, as AI-native videos become increasingly diverse, this method is insufficiently supportive of specific issues such as dialogue, OCR, audio-visual synchronization, and identity consistency. Key detection errors are easily masked by the overall similarity score, and the detection results cannot be verified.

[0046] Furthermore, existing AI video detection methods also include video detection based on fixed rule templates. This method pre-defines prompts into several fixed detection objects (also known as fixed slots), such as subject, location, action, and quantity. Then, it performs item-by-item judgment through object detection, image classification, OCR, automatic speech recognition (ASR) technology, or rule matching. Finally, it weights the detection results to obtain a total score. This method has a certain degree of interpretability in simple detection tasks, but when faced with complex video scenes, all prompts use the same set of slots and thresholds, which cannot achieve targeted detection of prompts and videos in different scenarios. For example, some videos do not require strong verification of location or quantity, but fixed detection rules may cause false positives. Some videos must have strong verification of dialogue or video change order, but fixed detection rules are prone to omissions. Moreover, fixed weights are difficult to adapt to different video types and detection standards.

[0047] Of course, many video inspection companies still rely on manual review of low-resolution frames one by one. This method is flexible, but it is inefficient, difficult to unify standards, difficult to accumulate experience, and difficult to support large-scale automatic quality inspection and long-term threshold calibration.

[0048] In summary, existing video detection methods suffer from problems such as simple detection methods, low accuracy, poor flexibility, and even low efficiency.

[0049] Based on this, this application proposes a video detection method. It determines a prompt word profile based on the prompt word information corresponding to the native video to be detected, and generates a dedicated detection route for the native video based on the prompt word profile. Then, it performs native video detection processing based on at least one of the following: a set of video detection plugins, a set of key video detection objects, a video consistency detection weight allocation strategy, and a video detection failure gating trigger rule, to obtain the target detection result of the native video to be detected. In this way, a dedicated detection path can be generated for different native videos and their prompt word information, enabling targeted detection of different native videos and making video detection more comprehensive and accurate.

[0050] Next, with reference to the accompanying drawings, the video detection method provided in this application will be described in detail.

[0051] See Figure 1 The video detection method provided in this application includes the following steps.

[0052] S100, determine the original video to be detected and the prompt word information corresponding to the original video to be detected.

[0053] For example, the AI ​​native video file to be detected (i.e., the native video to be detected) and the prompt words corresponding to the native video to be detected are obtained.

[0054] In one implementation of this application, the prompt information can be a text prompt. Of course, the prompt can also be a voice prompt, a structured prompt, such as table information, etc.

[0055] Furthermore, in this application, if it is determined that the prompt word information is related to the detection of the subject in the video (e.g., real person, animal, virtual person, virtual animal, etc.), a reference image is also determined. This reference image is an image required for identity consistency detection processing of the video to be detected.

[0056] Furthermore, in the implementation of this application, video detection rule configuration information can also be determined, and subsequent target detection route generation can be performed based on the video detection rule configuration information. The video detection rule configuration information includes gating thresholds, weight coefficients, plugin timeout limits, sampling parameters, and other information.

[0057] Furthermore, in the implementation method of this application, detection-related information of historical native videos can also be obtained, and subsequent target detection routes can be generated based on the detection-related information of historical native videos. The detection-related information of historical native videos includes manually reviewed tags, remarks, detection results, and other information.

[0058] S200, determine the prompt word profile based on the prompt word information.

[0059] In one implementation of this application, determining the prompt profile based on the prompt information includes: performing feature extraction processing on the prompt information to obtain core verification elements.

[0060] For example, the prompt words are preprocessed and structured parsed to obtain the core verification elements.

[0061] Specifically, the process involves cleaning up formatting noise irrelevant to detection from the prompts, extracting core verification elements such as subject (Who), scene (Where), action (What), quantity (Count), visual style, dialogue, text, and stage descriptions, and generating prompt versions corresponding to different video detection scenarios based on these core verification elements and the video detection scenario. These versions include visual detection prompts, semantic analysis prompts, and entity detection prompts. The video detection scenario is determined based on the core verification elements.

[0062] For example, if the prompt words are determined to be used for visual / entity detection in the original video to be tested, then the visual / entity detection prompt words retain only concrete descriptions of core visual subjects, scenes, and actions, while removing non-visual content such as abstract viewpoints and dialogue text. This reduces interference from irrelevant information and improves the accuracy of entity recognition and slot verification. If the prompt words are determined to be used for semantic analysis detection in the original video to be tested, then the semantic analysis prompt words retain the complete context and core requirements, and are used for global semantic consistency verification to ensure the accuracy of the overall semantic matching degree judgment.

[0063] Thus, in this application, after cleaning, standardizing, and structurally extracting the prompt word information, a prompt word version is obtained, providing basic data for subsequent scene pattern recognition and target detection route generation.

[0064] Furthermore, the prompt word information is processed by pattern recognition based on the target pattern recognition method to obtain the prompt word scene pattern.

[0065] In this application, instead of relying on a large language model for scene pattern recognition, a pattern recognition algorithm is constructed and called to identify the prompt word scene pattern (also known as prompt word scene pattern) based on the prompt word information according to the target pattern recognition method.

[0066] The prompt word scene modes include any one of the following: standard visual description scene, dialogue script scene, voiceover or monologue scene, mechanical transformation or structural reorganization scene, commentary or abstract expression scene, tutorial demonstration scene, and multi-shot narrative scene.

[0067] Furthermore, the target pattern recognition method includes at least one of the following processing methods: dictionary extraction processing method, regular pattern matching method, and heuristic multidimensional feature scoring method.

[0068] In one implementation of this application, the pattern recognition algorithm is a complex algorithm based on a low-latency "domain dictionary extraction + regular pattern matching + heuristic multidimensional feature scoring".

[0069] For example, if the pattern recognition algorithm is invoked based on the target pattern recognition method, and it is determined that the prompt word information does not match the preset pattern feature information, then the prompt word scene pattern of the prompt word information is determined to be the standard visual description scene.

[0070] The pattern recognition algorithm is based on the regular pattern matching method. If the prompt word information is determined to match the preset language format (e.g., "XXX said:"), and it is determined that the speaker entity needs to be dynamically extracted, and the number of different speaker entities to be extracted is greater than or equal to 2, then the prompt word scenario mode of the prompt word information is determined to be a multi-speaker dialogue script scenario.

[0071] The pattern recognition algorithm is based on dictionary extraction. If the prompt word information is determined to match a preset strong feature word (such as "dialogue", "narration", "voiceover") or if the prompt word information includes specific camera position / scene description information, then the prompt word scene mode is determined to be a voiceover or narration scene.

[0072] The pattern recognition algorithm is based on dictionary extraction to extract multi-dimensional keyword combinations from the prompt word information. If it is determined that the multi-dimensional keyword combination hits the mechanical subject keyword (such as the subject form words such as "mechanical cat" and "mecha" and the deformation action words and part words such as "deformation" and "disassemble") and the number of keywords (such as subject form words and deformation action words) hits reaches the preset dynamic threshold, then the prompt word scene mode of the prompt word information is determined to be a mechanical deformation or structural reorganization scene.

[0073] The pattern recognition algorithm uses a heuristic multidimensional feature scoring method to extract logical connectives (such as "because," "sarcasm," "I think," etc.) from the prompt words. It then integrates the logical connectives and subjective intent words to obtain a feature score. When the score exceeds a set threshold (e.g., ...), the algorithm calculates the feature score. 6 points, or in low real-time footage If the score is 4, then the prompt word scenario pattern is determined to be an opinion commentary or abstract expression scenario.

[0074] The pattern recognition algorithm is based on the regular pattern matching method. If it is determined that the prompt word information matches the scene keywords, and the regular pattern matching method can extract the temporal sentence structure (such as "first step", "first...then" etc.) from the prompt word information to determine that there is an action chain or step clause, then the prompt word scene pattern of the prompt word information is determined to be the tutorial demonstration scene.

[0075] The pattern recognition algorithm is based on dictionary extraction. If the prompt word information is determined to match the keywords of a multi-shot scene (such as "transition", "shot one", etc.), or if the prompt word information is determined to include prompt words with a total number of shots > 1 through structured parsing calculation, then the prompt word scene pattern of the prompt word information is determined to be a multi-shot narrative scene.

[0076] Furthermore, a prompt word profile is obtained based on the core verification elements and prompt word scenario patterns.

[0077] For example, the aforementioned core verification elements and prompt word scenario patterns are encapsulated to obtain a structured prompt word profile.

[0078] Furthermore, in one implementation of this application, a scenario-specific schema extractor is used to perform schema extraction processing to obtain the structured schema corresponding to the prompt word information.

[0079] For example, natural language prompts are transformed into multi-dimensional prompt profiles by combining heuristic rule matching (including regular pattern matching and multi-dimensional feature dictionary extraction), structured entity parsing, and scene-specific schema extractors.

[0080] Specifically, heuristic rule matching utilizes a large built-in feature vocabulary and regular expression patterns based on the aforementioned pattern recognition algorithm to accurately extract actions, props, compound nouns (trigger gating), and camera language. Structured entity parsing involves determining the aforementioned core verification elements. Scene-specific schema construction involves using a dedicated extractor to decompose the prompt information into a structured schema of prompts, such as initial form, stage sequence, and target features, if the prompt scene pattern is determined to be a special scene like mechanical transformation or tutorial demonstration.

[0081] Furthermore, the original prompt words are subjected to denoising and semantic purification processes. Three types of effective semantic information are extracted from the original prompt information in a structured manner: concise prompt text, scene context information, and core verification elements. Redundant and invalid interference information is eliminated to generate standardized prompt word semantic data with high signal-to-noise ratio. This provides accurate and pure semantic input for downstream tasks such as scene recognition, route generation, special plugin detection, and consistency verification. It effectively avoids strategy deviations and detection misjudgments caused by interference from invalid semantic information and significantly improves the semantic adaptation accuracy of the entire detection pipeline.

[0082] The final prompt word profile includes at least one of the following: prompt word scene mode, visual prompt words, entity prompt words, analysis prompt words, applicable slots, and prompt word structured schema (e.g., standardized variant schema), which is used for the accurate generation of subsequent target detection routes.

[0083] Thus, in subsequent target detection route generation, the detection objects (also known as applicable gates) can be dynamically calculated based on the subject extracted from the prompt word profile and the prompt word scene pattern.

[0084] In one implementation of this application, the prompt word scene mode is a mechanical deformation or structural reorganization scene, and the prompt word profile of the prompt word information includes at least one of the following: profile type identifier, prompt word scene mode (also known as content mode), summary prompt words for semantic detection, visual prompt words for entity detection, list of applicable gating items (storing key slot gating that needs to be enabled in the current scene), list of deformation stages, list of key verification parts, and list of prohibited failure items.

[0085] Furthermore, in the implementation of this application, feature extraction processing is performed on the original video to be detected to obtain a video profile of the original video to be detected. The video profile includes video scene patterns.

[0086] For example, basic metadata such as frame rate, total number of frames, duration, and resolution of the original video to be detected are read, and an initial sampling strategy (also known as an initial sampling plan) is generated. Among them, frame rate, total number of frames, and duration are used to accurately locate frame-level timestamps, ensuring the traceability of subsequent low-resolution frames and key time windows; resolution is used to adapt to the input specifications of different detection models and avoid feature distortion caused by scaling; the sampling strategy determines the frame extraction density and distribution strategy for subsequent detection, balancing detection accuracy and running efficiency.

[0087] And determine the video scene mode of the native video to be detected based on the video title, video description, and video content.

[0088] Video scene modes include standard single-lens video scene, voice-over scene, multi-lens mixed editing scene, and screen recording / UI interface scene.

[0089] Furthermore, a video profile of the native video to be detected is generated based on the basic metadata of the native video to be detected, the initial sampling strategy, and the video scene mode.

[0090] S300 generates a target detection route for the native video to be detected based on the prompt word profile. The target detection route includes a dedicated detection route, which includes at least one of the following: a video detection plugin set, a video key detection object set, a video consistency detection weight allocation strategy, and a video detection failure gating trigger rule.

[0091] In one implementation of this application, a video detection policy (joint_policy) for the native video to be detected is generated based on the prompt word profile.

[0092] Traditional solutions often treat all prompts as ordinary image-text matching tasks. In reality, the detection priorities for different prompts vary significantly. For example, spoken or narrated videos focus more on characters, the main visual element, dialogue, and the audio-visual relationship, and shouldn't be forced to match every location slot. Multi-speaker dialogue videos need to focus on speakers, dialogue order, cross-line misplacement, and voice alignment. Mechanical transformation videos need to focus on phased structural changes, key component reconstruction, sequence, and prohibited failures. Multi-camera narrative videos need to focus on cross-camera subjects, events, and temporal continuity. Commentary / opinion videos may focus more on whether the main visual element supports the viewpoint, rather than matching every noun verbatim. Therefore, without first identifying the scene mode to which the prompt belongs, subsequent detection routing, slot gating, and weight configuration are prone to mismatch.

[0093] In this application, if the prompt word profile includes a multi-speaker dialogue script scenario, the video detection strategy prioritizes speaker matching, dialogue order, dialogue completeness, and audio-visual alignment, while downplaying unnecessary scene and slot quantity checks to suit the detection needs of narrative and multi-character videos. If the prompt word profile includes a voiceover or narration scenario, the video detection strategy extracts the main visual cues from the prompt word profile, downplaying or even skipping unnecessary location gates to avoid false positives such as "failure due to complete matching of voiceover content but mismatched background details," focusing on character identity and dialogue completeness checks. If the prompt word profile includes a mechanical deformation or structural reorganization scenario, the video detection strategy automatically generates detection targets containing deformation stages, key parts, and prohibited failure items, directly triggering an event-based review process, using "process completeness and structural authenticity" as core verification points to intercept A. The common problem in native videos is "light effect occlusion replacing distortion." If the prompt word profile includes prompt word scene patterns of opinion commentary or abstract expression, the video detection strategy is to reduce or clear hard slot gating of strong entities and strong locations, shifting the verification focus from "precise entity matching" to "whether the semantics of the image supports the expression of opinion," significantly reducing the false judgment rate of abstract content. If the prompt word profile includes prompt word scene patterns of tutorial demonstration, the video detection strategy is to increase the weight of temporal consistency and logical consistency dimensions, focusing on verifying the order of steps, the completeness of key actions, and the supporting relationship between the image and the steps, adapting to the core quality inspection requirements of teaching videos. If the prompt word profile includes prompt word scene patterns of multi-shot narrative, the video detection strategy is to trigger a special plugin for multi-shot narrative consistency, focusing on verifying the continuity of the subject across shots, the rationality of the event sequence, and logical coherence, splitting the single-shot global detection into a two-level mode of "shot verification + global continuity verification."

[0094] Thus, before entering the high-cost large-scale model inference, scene pattern recognition is completed through lightweight and low-cost feature engineering, significantly reducing the overhead of invalid inference and adapting to the high-throughput requirements of large-scale batch quality inspection. Moreover, it breaks away from the drawback of the traditional solution of "one set of rules covering all scenarios" by attaching differentiated slots, weights, and gating configurations based on the recognition results, thereby solving the dual problems of "overly strict rules leading to false positives in simple scenarios and overly lenient rules leading to false negatives in complex scenarios".

[0095] Furthermore, in this application, the dedicated detection route also includes at least one of the following: a set of detection objects that need to be relaxed or skipped, a review strategy for whether to enter the verification review, a video sampling strategy (e.g., tail enhancement sampling, variable peak sampling, etc.), a soft fusion strategy, and a runtime calibration strategy.

[0096] Furthermore, a dedicated detection route for the native video to be detected is generated based on the prompt word profile, including: according to the layered detection pipeline, a dedicated detection route for the native video to be detected is generated based on the prompt word profile, video profile, and video detection rule configuration information.

[0097] The prompt word profile includes the prompt word scene mode (prompt_mode), and the video profile includes the video scene mode (video_mode).

[0098] For example, the layered detection pipeline includes a routing layer that uses a built-in rule mapping engine to map the identified prompt word scene patterns and video scene patterns to specific video detection strategy parameters based on pre-configured correspondences (such as heuristic pre-configured rules). This is done to override or adjust the default detection route baseline configuration to obtain a dedicated detection route.

[0099] The pre-configured mapping relationships include mapping methods based on prompt word scene patterns, mapping methods based on video scene patterns, and multi-mode overlay mapping methods.

[0100] If a target scene combination mode cannot be generated based on the prompt word scene mode and the video scene mode, then the first detection route is generated based on the prompt word scene mode mapping method.

[0101] For example, different weight allocations and slot verification focuses are pre-configured according to different prompt word scenario patterns to obtain the first detection route.

[0102] For example, when the prompt word scene mode is a tutorial demonstration scene, a higher pre-allocated logical and temporal dimension weight is used to obtain a video consistency detection weight allocation strategy, and action detection (i.e., action slots) is set as a key detection object with strong verification (i.e., key slots), resulting in a set of key video detection objects. When the prompt word scene mode is a multi-shot narrative scene, a higher pre-allocated temporal dimension weight is used to obtain a video consistency detection weight allocation strategy to ensure cross-shot continuity, and a soft fusion strategy is enforced to require the lower limit of slot coverage.

[0103] Furthermore, the mapping method based on video scene patterns generates a second detection route according to the video scene patterns.

[0104] For example, the strictness of visual verification for the native video to be detected is dynamically adjusted according to the video scene mode to obtain a second detection route. For instance, when the video scene mode is screen recording or a UI interface, the hard visual gate checks on people, locations, and quantities are removed to obtain video detection failure gate triggering rules, and semantic weights are increased to obtain a video consistency detection weight allocation strategy. When the video scene mode is abstract animation, the gating on the location and quantity of real physical entities is relaxed, and the video detection failure gate triggering rules are adjusted.

[0105] Furthermore, based on the rule configuration information, a dedicated detection route is generated for the native video to be detected according to the first detection route and the second detection route.

[0106] For example, the second detection route is preferred, and the first detection route is merged to obtain a dedicated detection route without changing the second detection route.

[0107] If a target scene combination pattern can be generated based on the prompt word scene pattern and the video scene pattern, then a dedicated detection route for the original video to be detected is generated based on the target scene combination pattern.

[0108] For example, when the prompt word scene pattern and the video scene pattern can generate a target scene combination pattern, a dimensionality reduction pre-configuration strategy is triggered based on the multi-pattern overlay mapping method. Based on this strategy, a dedicated detection route is obtained, further reducing the probability of false blocking. A typical combination is: prompt word scene pattern = opinion / comment + video scene pattern = spoken commentary on camera (as an example of a target scene combination pattern). In this case, only the single hard gate of the person's identity is retained, resulting in a video detection failure gate triggering rule. The semantic consistency passing threshold is lowered to obtain a video consistency detection weight allocation strategy.

[0109] It should be noted that in this application, the rule engine covers parameters layer by layer according to the priority of multi-mode overlay mapping > video scene mode mapping > prompt word scene mode mapping > global default baseline. High-priority rules directly override low-priority general configurations, and finally integrate and output complete structured dedicated detection routes.

[0110] For example, taking a prompt word scene mode as a mechanical deformation or structural reorganization scene, the dedicated detection route generated in this application includes: prompt word scene mode, such as a mechanical deformation or structural reorganization scene; a set of key detection objects, such as needing to detect key objects such as "Who", "What", and "Count"; a set of hard-gated detection objects, such as ["Who", "What"], which represents a hard-gated forced key slot. If either the subject or the deformation action is not satisfied, the L2 hard-gated failure is directly triggered, resulting in a veto (as an example of a video detection failure gating triggering rule); a six-dimensional consistency score weight overwrite configuration, such as a semantic consistency detection weight of 0.28, an entity consistency detection weight of 0.18, a temporal consistency detection weight of 0.22, and a logical consistency weight of 0.14 (as an example of a video consistency detection weight strategy); and core dimension hard-gated threshold coverage, including a semantic dimension passing circuit breaker and a temporal dimension circuit breaker (as an example of a video detection failure gating triggering rule).

[0111] In this way, by using traffic splitting based on prompt word scene patterns and video scene patterns, the globally unified fixed detection strategy is completely abandoned, eliminating the scene adaptation imbalance caused by the general baseline. The layered heuristic rules are configurable and iterable. Adding new video formats and prompt scenes only requires expanding the mapping library, without refactoring the detection pipeline code; routing loads specialized plugins on demand and allows for custom frame sampling, without requiring the full plugin to run, thus improving the throughput speed of batch quality inspection; different scenes allow for the independent definition of key objects and circuit breaker thresholds, enabling flexible detection.

[0112] The S400 performs detection processing on the original video to be detected based on a dedicated detection route, and obtains the target detection result of the original video to be detected.

[0113] See Figure 2 After inputting the prompt words and the original video to be detected, the system performs prompt word preprocessing and structured parsing, then prompt word scene recognition, generates adaptive detection routes based on the routing layer (Router), and executes the adaptive routes for video detection processing.

[0114] In the implementation of this application, the target detection route generated by the routing layer based on the prompt word scene mode, video scene mode, and detection rule configuration information can also be a standard detection route or an event-based review route (also known as an event-based review process).

[0115] In this application, the Router layer is the traffic scheduling hub of the entire pipeline. Its inputs are the prompt word profiles, video profiles and routing rule configurations generated by the preceding stages. Its core function is to dynamically match the optimal detection process according to the scenario characteristics of the task, so as to achieve differentiated scheduling of "basic scenarios follow the baseline, special scenarios follow the dedicated process, and complex scenarios are split into events".

[0116] If the prompt word profile structure is simple and does not hit any specific scene strong features (e.g., the prompt word scene mode is a standard scene), and the video profile is a single shot with no special form (e.g., the video scene mode is a labeled single shot video scene), then the target detection route is determined to be the standard detection route. The default baseline configuration and basic detection capabilities of the standard detection route are adopted, and global six-dimensional consistency scoring is performed. This is the fallback process for general scenarios, namely the standard flow.

[0117] If the prompt word profile / video profile hits a specific dimension with strong features (such as dialogue, on-screen text, abstract artistic conception); or if the video scene mode is a voice-over, screen recording, or mixed editing scenario, then the target detection route is determined as a dedicated detection route, the route status is marked as special, and corresponding special plugins are attached (such as text detection plugin S3, audio-visual consistency plugin S4 / speech recognition verification plugin S5, complex event understanding plugin S-VLM) to supplement the subdivided dimension verification that general detection cannot cover, i.e., the special plugin flow.

[0118] That is, in this application, the implementation method of this application, based on the target detection route, performs detection processing on the native video to be detected to obtain the target detection result of the native video to be detected, including: according to the layered detection pipeline, based on the dedicated detection route and the video detection strategy, performing detection processing on the native video to be detected to obtain the target detection result of the native video to be detected.

[0119] In the implementation of this application, when the dedicated detection route includes a set of video detection plugins, the layered detection pipeline includes a Special Plugins layer; when the dedicated detection route includes a set of key detection objects, the layered detection pipeline includes a Key Detection Objects layer (Layer2 Slot); when the dedicated detection route includes a target weight allocation strategy, the layered detection pipeline includes a consistency detection layer (Layer2 Extract); when the dedicated detection route includes target gating triggering rules, the layered detection pipeline includes a Hard Gate layer; when it is determined that the original video to be detected has not triggered the Hard Gate layer's hard gate, the layered detection pipeline includes a soft fusion layer (ScoreFusion); and when the dedicated detection route also includes a review and approval strategy, the layered detection pipeline includes a review and approval layer (also known as an event-driven review layer, Layer3).

[0120] In this application, based on the target detection plugin set included in the video detection plugin set, the native video to be detected is processed according to the video detection strategy, according to the specialized plugin layer, to obtain the plugin detection result of the native video to be detected.

[0121] The target detection plugin set includes at least one of the following: multi-camera narrative consistency plugin, identity consistency plugin, text detection plugin, audio-visual consistency plugin, speech recognition verification plugin, and complex event understanding plugin.

[0122] In this application, the specialized plugin layer is a customized supplementary module for general detection capabilities. It is triggered on demand by the Router layer (not full-scale operation) and specifically addresses the specific pain points unique to AI native videos that are difficult for general visual models to accurately verify (such as identity drift across shots, misaligned dialogue, and abstract semantic judgment). Based on the target-direction plugin, video detection is performed, and the plugin detection results and runtime metadata are output. Deeply customized detection processing is achieved for the characteristics of AI native videos (such as multimodal and complex temporal sequences).

[0123] Furthermore, the raw output of the target detection plugin is not directly included in the final comprehensive score. Instead, it is all converted in the subsequent plugin signal standardization process before being connected to the main process according to the rules. This avoids the weight distortion problem caused by the direct fusion of scores from heterogeneous plugins from the root.

[0124] If the target detection plugin set includes a multi-shot narrative consistency plugin S1, then based on the multi-shot narrative consistency plugin S1, the native video to be detected is processed according to the video detection strategy to obtain the plugin detection result of the native video to be detected. This includes: extracting the corresponding structured storyboard script (including the expected duration of each shot and scene description) from the prompt word profile included in the video detection strategy through a built-in large language model or regular expression pattern; calling the physical scene detection algorithm (Video Scene Aligner) to identify the actual shot segmentation points of the native video to be detected; using a dynamic programming algorithm to perform sequence alignment verification between the "expected storyboard sequence" and the "actual video segmentation result"; outputting the corresponding alignment status of the native video to be detected (e.g., perfect alignment, missing expected transitions, existence of redundant transitions); and returning the split storyboard segments and time-series mapping plan to obtain the multi-shot narrative consistency plugin detection result, which is then used for subsequent consistency scoring on a shot-by-shot basis. In this way, the problems of missing transitions, reversed order, and missing storyboards in multi-shot videos that cannot be identified by the general global scoring of existing video detection methods are solved.

[0125] If the target detection plugin set includes the identity consistency plugin S2, then based on the identity consistency plugin S2, the original video to be detected is processed according to the video detection strategy to obtain the plugin detection result of the original video to be detected. This includes: for detection tasks that provide a reference map, extracting keyframes of the video, comparing and calculating them with the reference map using at least one of the following methods: Histogram Correlation, SSIM Approx, and optional cross-modal features (such as CLIP Cosine Similarity), outputting a frame-by-frame similarity feature matrix, and weighted fusion to obtain a comprehensive global ID fidelity score, thus obtaining the identity consistency plugin detection result. In this way, the common problems of character face distortion, identity drift, and mid-process actor replacement in AI-generated videos are solved, which are difficult for general models to accurately capture.

[0126] Among them, histogram correlation is used to compare the color and contour distribution of the image; SSIM Approx is used to compare facial / subject structural features; and optional CLIP Cosine Similarity is used to compare semantic features across modalities.

[0127] If the target detection plugin set includes OCR text detection plugin S3, audio-visual consistency plugin S4, ASR, and dialogue verification plugin S5, then based on OCR text detection plugin S3, audio-visual consistency plugin S4, ASR, and dialogue verification plugin S5, the original video to be detected is processed according to the video detection strategy to obtain the plugin detection results of the original video to be detected. This includes: calling specific multimodal model or algorithm modules (such as Whisper to extract speech, OCR engine to recognize text on screen, and waveform peak alignment analysis), performing regular expression matching, fuzzy string matching (FuzzyMatch), and alignment calculation on the extracted results and the hard text requirements in the prompt words, and outputting the recognition hit rate score and the coordinates / timestamp of the located abnormal area to obtain the corresponding detection results.

[0128] Specifically, based on the S3 OCR text detection plugin, the native video to be detected is processed according to the video detection strategy to obtain the plugin detection results of the native video to be detected. This includes: calling the OCR engine to recognize the text in the video screen, performing alignment verification with the specified text in the prompt words through regular expression matching and fuzzy matching, outputting the text recognition hit rate score, the regional coordinates and timestamps of abnormal text, so as to obtain the OCR text plugin detection results. This solves the problem that general visual models cannot accurately verify the correctness and completeness of text in videos with specified subtitles and poster text.

[0129] Based on the audio-visual consistency plugin S4, the native video to be detected is processed according to the video detection strategy to obtain the plugin detection result of the native video to be detected. This includes: verifying the time synchronization of speech and image actions and lip movements in the video through algorithms such as waveform peak alignment and lip feature matching, and outputting audio-visual offset, synchronization score, and timestamp of abnormal offset to obtain the audio-visual consistency plugin detection result, thus solving the common audio-visual asynchrony and lip misalignment problems in spoken and digital human videos.

[0130] Based on the S5 ASR / dialogue verification plugin, the native video to be detected is processed according to the video detection strategy to obtain the plugin detection results of the native video to be detected. This includes: calling specific multimodal models or algorithm modules (such as Whisper to extract speech, OCR engine to recognize text on screen, waveform peak alignment analysis) for video feature extraction only; aligning the extracted results with the specified dialogue in the prompt words sentence by sentence based on regular expression matching, fuzzy string matching, and alignment calculation; verifying completeness and correct order; and outputting the dialogue hit rate, details of missing / incorrect sentences, and timestamps of abnormal dialogue to obtain the ASR / dialogue plugin detection results. This solves the problems of missing, incorrect, and reversed dialogue in spoken and dialogue videos, and addresses the issue that general semantic models cannot perform fine-grained verification.

[0131] If the object detection plugin set includes the Abstract Action / Complex Event Understanding Plugin S-VLM, when conventional entity detection fails or the prompts contain highly abstract expressions, the plugin, based on the Abstract Action / Complex Event Understanding Plugin S-VLM, performs detection processing on the original video to be detected according to the video detection strategy, obtaining the plugin detection results for the original video to be detected. This includes: extracting the video temporal frame sequence, assembling a dedicated decision prompt to call the multimodal large model (VLM), performing end-to-end semantic evaluation, outputting not only a confidence score in the 0~1 range, but also, by parsing the VLM response, outputting natural language correction suggestions about hand movements, subject shape, or action coherence, and revealing the original inference reasoning for review, in order to obtain the complex event plugin detection results. This solves the problem that rule-based and general detection models cannot accurately determine semantic matching degree for videos with abstract artistic conception, complex actions, and unconventional scenes.

[0132] Furthermore, in the implementation of this application, the layered inspection pipeline also includes a basic quality inspection layer, Layer1.

[0133] Layer 1 is the first entry threshold in the entire detection pipeline. It is a basic video quality check unrelated to prompts. Its core function is to intercept severely damaged videos, prevent invalid samples from entering the high-cost semantic detection link, and reduce overall computational overhead.

[0134] Layer 1 detects fundamental image quality issues in the original video being inspected, including: black frames, white frames, abnormal solid colors across the entire screen; severe blurring and image quality distortion; long periods of stillness with no image change; flickering and abnormal brightness fluctuations; and at least one of the following: optional image quality evaluation without reference (general image quality scoring). By intercepting substandard samples upfront, this significantly reduces resource waste in subsequent semantic detection and plugin calls, improving the overall throughput efficiency of batch quality inspection scenarios.

[0135] Furthermore, based on the key detection object layer and the video detection strategy, according to the set of key detection objects in the video, the detection of whether the original video to be detected meets the key detection object requirements is obtained, and the key detection object detection results of the original video to be detected are used as the target detection results of the original video to be detected.

[0136] Layer2 Slot (Key Detection Object Layer) is a single-point verification layer for hard constraints on prompt words. It is used to detect whether the key detection objects (i.e., key slots) involved in the prompt word profile in the original video to be detected meet the requirements. It belongs to the precise verification of the "point" dimension. The detection results of key detection objects directly serve as the core triggering basis for subsequent hard gating, complementing the global scoring of Layer2Extract (Consistency Detection Layer) in the "surface" dimension.

[0137] In this application, the key detection objects include at least one of the following: Who: Subject / Role slot, used to detect whether the core subject, role, or item specified by the prompt appears in the video; Where: Scene / Location slot, used to detect whether the video screen matches the scene, location, or environment specified by the prompt; What: Action / Event slot, used to detect whether the action, event, or behavior specified by the prompt occurs in the video; Count: Quantity slot, used to detect whether the number of specified subjects in the video meets the requirements of the prompt; Composition: Composite object slot, used to detect whether the combined or composite objects specified by the prompt appear in the video.

[0138] The scope of key slot verification is not fixed, but dynamically determined by the prompt word profile and adaptive video detection configuration information. For example, for spoken videos, only the Who (subject) slot is retained for verification, while non-core slots such as location and quantity are weakened to avoid false positives; for mechanical transformation videos, the Who (subject), What (transformation action), and Count (quantity) slots are retained; for opinion and commentary videos, most entity-related hard slots are cleared, and only core semantic-related verification is retained.

[0139] In this way, the problem of "one-size-fits-all fixed slot rules" in existing technologies is solved, and the scope of hard verification is fully matched with the core requirements of the scenario. This can effectively ensure that key requirements are not overlooked, and can also greatly avoid misjudgments caused by non-core slots.

[0140] Furthermore, according to the consistency detection layer, a video frame extraction strategy is generated for the original video to be detected based on the video detection strategy. The original video to be detected is then processed by extracting frames based on the video frame extraction strategy to obtain a sampled frame sequence.

[0141] Layer 2 Extract (Consistency Detection Layer) is the core output layer of the global consistency score. It is a comprehensive evaluation based on the "surface" dimension. It is driven by scenario strategy, with multiple independent validators running in parallel and key slots being dynamically calibrated. It outputs standardized multi-dimensional consistency scores (as examples of target consistency detection results), which are the core data foundation for subsequent comprehensive integration. It also includes full diagnostic evidence to support manual review and problem debriefing.

[0142] In this application, a video frame extraction strategy for the native video to be detected is generated according to the consistency detection layer, including: generating a video frame extraction strategy for the native video to be detected based on the prompt word profile and the video detection strategy according to the consistency detection layer.

[0143] For example, based on the joint_policy (also known as the video detection policy, which is the full set of policy parameters generated by adaptive routing, including weights, thresholds, video sampling policies, etc.) distributed by the routing layer and the prompt word profile, a video frame extraction plan (also known as the video frame extraction policy) is dynamically generated according to the scene policy. For example, variable peak sampling is enabled in mechanical deformation scenes, and uniform sampling is enabled in voice-over scenes to ensure that key frames are not lost.

[0144] Furthermore, based on the sampled frame sequence, a consistency check is performed on the sampled frame sequence according to the video detection strategy to obtain the initial consistency detection result of the original video to be detected. The consistency check includes at least one of semantic consistency detection, entity consistency detection, spatial consistency detection, temporal consistency detection, style consistency detection, and logical consistency detection.

[0145] For example, for the sampled frame sequence and the parsed prompt word information (such as prompt word profile), six dedicated verifiers are called in parallel to calculate the basic scores of the six dimensions (as an example of the initial consistency detection results), without interfering with each other.

[0146] Among them, semantic consistency detection is used to evaluate the degree of fit between the overall video image and the core semantics of the prompt words in the prompt word profile, which is a macro-level judgment of the global scene matching degree.

[0147] Entity consistency detection is used to check whether the categories and attributes of core entities in a video are consistent with the requirements of the prompt information.

[0148] Spatial consistency detection is used to verify whether the spatial orientation relationships (such as up and down, left and right, inside and outside) between multiple entities conform to the prompt word description (e.g., prompt word description A is above B).

[0149] The time consistency detection process is used to evaluate the persistence and smoothness of core entities in time sequence, and whether there are cases of disappearing midway or abrupt jumps.

[0150] Style consistency detection is used to verify whether the style, tone, and texture of the entire original video being tested are consistent and whether they meet the style requirements specified by the prompt words.

[0151] Logical consistency detection is used to determine whether the development of actions and physical laws in the original video to be detected conform to common sense and the logical settings of the prompt words.

[0152] Furthermore, based on the detection results of key detection objects and the video detection strategy, the initial consistency detection results of the original video to be detected are adjusted to obtain the target consistency detection results of the original video to be detected, which are then used as the target detection results of the original video to be detected.

[0153] For example, combining the key detection results of the key detection objects (such as Who, Where, What, Count, Composition, etc.) output by Layer2 Slot, the basic score of the corresponding dimension (such as the entity dimension) is dynamically penalized / compensated according to the joint strategy: for example, if the key detection result is that the main slot is not hit, the entity consistency detection score is deducted according to the preset adjustment rule; if all core slots are hit, the entity consistency detection score can be slightly compensated, and the calibrated standardized score is finally output.

[0154] In the implementation of this application, the scores for each dimension simultaneously retain detailed diagnostic evidence, sampling information, and weakness summaries as inputs for hard gating, such as checking whether the core dimensions have fallen below the fatal threshold dynamically set by the joint strategy; failure to meet the threshold results in a direct circuit breaker and a FAIL. It also serves as the basis for comprehensive fusion, such as calculating the global comprehensive score by weighting according to scenario weights and triggering soft penalties when significant weaknesses exist. Furthermore, it serves as evidence for gray-zone verification, such as supporting rapid manual location of problem dimensions and problem frames, and as the basis for generating detection reports.

[0155] Furthermore, the output formats, dimensions, and semantics of specialized plugins such as OCR, ASR, audio-visual alignment, and identity consistency are completely different. Some output text edit distance, some output time offset in milliseconds, some output cosine similarity probability, and some output event hit status. Directly mixing the raw scores into the overall score will cause five major problems: inconsistent score dimensions, making direct horizontal comparison impossible; different confidence limits for plugins, leading to weight drift when directly weighted; the same fact being scored repeatedly by the main model and plugins, interfering with the objectivity of the overall score; high-risk special failures being diluted by the average score of other high-scoring items, resulting in missed judgments of "overall score qualified but core requirements invalid"; and scores when plugins are abnormal / downgraded being incomparable, affecting the stability of the final result.

[0156] Therefore, in this application, a three-layer mechanism of unified protocol-adaptation conversion-distribution according to intent is used to convert the heterogeneous plug-in output into a standardized signal before subsequent processing. This not only preserves the plug-in's specialized detection capabilities but also does not compromise the stability and interpretability of the overall score. At the same time, it achieves complete decoupling between the plug-in and the main detection process.

[0157] In the implementation of this application, based on the target adapter corresponding to the target detection plugin in the target detection plugin set, the plugin detection result is processed in a unified format according to the plugin signal standard protocol to obtain the normalized plugin detection result of the original video to be detected, which is then used as the target detection result of the original video to be detected.

[0158] For example, a unified signal structure is first constructed, and a set of general plug-in signal standard protocols (PluginSignal Protocol) is predefined. All specialized plug-in outputs must be encapsulated into this unified format. This achieves format alignment of heterogeneous outputs at the data structure level, which is the basis for standardization.

[0159] Specifically, such as Figure 3 As shown, through the defined plugin signal standard protocol, the plugin detection results of all target detection plugins (such as OCR text results, ASR / dialogue results, audio-visual alignment results, identity consistency results, and multi-shot / event results) are unified into the same format of normalized plugin detection results. The normalized plugin detection results include at least: processing intent, used to identify the scope of the normalized plugin detection results; normalized score and confidence interval, which should be uniformly set to a standardized score in the range of [0,1], where 1 is completely compliant and 0 is completely non-compliant; gating blocking flag, with optional values ​​of PASS (pass), WARN (warning), and BLOCK (block); evidence chain data, which structurally stores detailed information such as low-score frame timestamps, bounding box coordinates, and abnormal text sets for manual review and problem localization.

[0160] Furthermore, for different types of specialized plugins, a dedicated feature adapter is deployed after the plugin runs independently and before the plugin detection results enter the main process. The adapter layer is an intermediate conversion layer between the specialized plugin and the main process. It configures a dedicated feature adapter for each type of specialized plugin and is responsible for converting the plugin-specific and dimensionally inconsistent raw outputs into a standardized signal format that conforms to the plugin signal standard protocol, thereby achieving normalized output of detection results for multiple heterogeneous plugins.

[0161] The core of the adapter is a non-linear mapping function plus a business threshold rule. It designs exclusive conversion rules for the original output type of different detection plugins. That is, based on the non-linear mapping function and the business threshold rule, it converts the plugin detection results output by the corresponding detection plugin (such as the text editing distance of OCR, the offset milliseconds of audio-visual alignment, the cosine similarity probability of face feature comparison, etc.) into the aforementioned normalized plugin detection results.

[0162] Furthermore, in the implementation of this application, the normalized plugin detection results of each target detection plugin are dynamically distributed based on the backbone fusion engine according to the processing intent included in the normalized plugin detection results.

[0163] In the implementation of this application, if the detection result processing intention indicates that the normalization plugin detection result is a policy signal, then the normalization plugin detection result is returned to the routing layer so that the routing layer updates the exclusive detection route and video detection policy according to the normalization plugin detection result.

[0164] For example, if the detection result processing intent indicates that the normalization plugin detection result is a policy signal, then the normalization plugin detection result is returned to the routing layer to adjust video detection policies such as thresholds and weights, and to adjust dedicated detection routes such as key detection object sets.

[0165] If the detection result processing intent indicates that the normalization plugin detection result is a scoring signal, then the normalization plugin detection result is sent to the soft fusion layer so that the soft fusion layer determines the comprehensive detection result of the original video to be detected based on at least one of the normalization plugin detection result, the key detection object detection result, and the target consistency detection result.

[0166] For example, if the detection result processing intent indicates that the normalization plugin detection result is a normalized scoring signal, then it is mapped into the specified soft fusion layer according to the scene routing strategy (e.g., under semantically comparable conditions) for weighted calculation of consistency dimensions (such as entity, time).

[0167] If the detection result processing intent indicates that the normalization plugin detection result is a gating signal, then the normalization plugin detection result is sent to the hard gating layer so that the hard gating layer determines the hard gating detection result of the native video to be detected based on at least one of the normalization plugin detection result, the key detection object detection result, and the target consistency detection result.

[0168] For example, if the detection result processing intent indicates that the normalization plugin detection result is a gate signal (Verdictgate), or the gate blocking flag is blocked (BLOCK), the hard gate layer is directly triggered to perform hard gate and output a detection failure result.

[0169] If the detection result processing intent indicates that the normalization plugin detection result is an evidentiary signal, then the normalization plugin detection result will be used as the target detection result of the original video to be detected.

[0170] For example, if the processing intent of the detection result indicates that the detection result of the normalization plugin is an evidence-only artifact, then the detection result of the normalization plugin will be used as an evidence artifact, without participating in any scoring or adjudication logic, and will only be attached to the evidence chain of the result for manual review and report generation.

[0171] For example, for the ASR / dialogue verification plugin (S5), if the plugin detection result is a serious missing dialogue, the normalized plugin detection result can be used as a conclusion gating signal. That is, when the ASR / dialogue verification plugin detects that the core dialogue missing rate required by the prompt words in the original video to be detected exceeds the preset blocking threshold (such as 30%), the adapter marks it as a conclusion gating type signal, the gating flag is set to BLOCK, and the hard gating global failure is directly triggered.

[0172] For the OCR text detection plugin (S3), if the plugin detection result is a slight text deviation, the normalized plugin detection result can be used as an evidence-only signal. That is, when the OCR text detection plugin detects that there are individual typos in the specified text in the original video to be detected, but the core semantics are complete, the adapter marks it as an evidence-only type, only displays the anomaly details on the results page, does not deduct the comprehensive score or use it as a soft penalty item in the soft fusion layer.

[0173] For complex event understanding plugins like S-VLM (e.g., mechanical deformation event review plugins), if the plugin detection result indicates that the deformation process is missing, the normalized plugin detection result can be used as a conclusion gating signal. That is, when the event review detects that the video is only occluded by light effects and has no actual structural changes, or jumps directly to the final form, the adapter marks it as a gating signal and directly triggers a hard failure or score capping condition.

[0174] For the multi-shot narrative consistency plugin (S1), if the plugin detection result is the shot recognition result, the normalized plugin detection result can be used as a strategy signal. That is, after the actual number and structure of the video shots are identified, the corresponding plugin detection + 63 result is converted into a strategy signal and sent back to the routing layer to dynamically adjust the scoring weight of the subsequent time-series dimension, analyze the prompt words, split the six-dimensional score into shot dimensions, and adjust the frame sampling strategy.

[0175] For the identity consistency plugin (S2), if the plugin detection result is identity similarity (e.g., face similarity score), when the comparability condition is met, the normalized plugin detection result is converted into a normalized score signal of the entity dimension. After being normalized to the [0,1] interval by the dedicated adapter, it is mapped to the entity consistency dimension and participates in the scoring calibration of this dimension according to the preset weight.

[0176] In this way, by using a unified protocol and adaptation conversion, the differences in units and confidence levels between different plugins are eliminated, avoiding weight drift and score distortion caused by direct weighting, and significantly improving the stability of the total score. This solves the problem that heterogeneous plugin detection results are "visible but not stable," ensuring that plugin detection results can influence the final conclusion without compromising the stability and interpretability of the total score.

[0177] Furthermore, according to the hard gating layer, the hard gating detection result of the original video to be detected is determined based on at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result.

[0178] Existing technologies generally use a single comprehensive weighted score as the sole basis for adjudication, which is prone to situations where the core requirements are completely invalid (e.g., the subject is completely wrong, the key scene is completely unmet, the dialogue is severely missing, the mechanical transformation process is completely nonexistent, or the OCR-specified text is completely absent). However, high scores in other dimensions can inflate the total score, leading to the problem of "false compliance" that is missed in judgment. This application adopts a collaborative mechanism that completely separates hard gating (rigid interception) and soft fusion (flexible scoring): hard gating maintains the bottom line of "must fail", while soft fusion enables fine-grained differentiation of "degree of excellence". The two are executed in layers without interfering with each other, ensuring zero missed judgments for high-risk issues while retaining the ability to finely evaluate minor deviations.

[0179] Among them, the triggers for hard gating include, but are not limited to: core element type: determining serious errors in the main body based on the detection results of key detection objects and the consistency detection results of the target (e.g., the entity consistency detection result does not meet the preset threshold), key slots are completely unsatisfactory, and specified text is completely missing (e.g., the semantic consistency detection processing result does not meet the preset threshold or is missing); special scenario type: determining serious missing / mismatched lines based on the detection results of the normalization plugin, mechanical deformation directly cutting to the final form, and only light effects without structural changes; business custom type: other high-risk failure items configured by business rules.

[0180] The specific processing mechanism of hard gating is as follows:

[0181] Thresholds and Slot Circuit Breaker: Circuit breaker thresholds are set based on two dimensions: core dimension scores and key slot coverage. Hard gating is triggered when either condition is met. For example, if the consistency detection score of core consistency dimensions such as semantics and entities is lower than the hard gating threshold, the circuit breaker will be triggered directly. If the key slots associated with the scenario are not hit, or the slot coverage rate (i.e., the ratio of hit key slots to total key slots) is lower than the preset lower limit, the circuit breaker will be triggered.

[0182] Pipeline short-circuit skipping fallback arbitration: After a hard gating failure is detected in the Layer 2 detection stage, the subsequent pipeline is directly interrupted, a FAIL conclusion is output, and the fallback arbitration of the large model in Layer 3 is skipped. The reason for the conclusion is clearly marked as "L2 hard gating failure", so as to avoid Layer 3 covering up or overturning the fatal structural failure.

[0183] Dynamic policy injection: The rules for hard gating are not globally fixed, but are dynamically injected by a joint policy generated by the pre-adaptive routing based on the prompt word scenario to achieve differentiated interception. For example, the list of key hard gating slots and the hard thresholds of core dimensions are dynamically covered for different scenarios to achieve differentiated and precise interception. The key slots in the list of key hard gating slots are completely different in different scenarios (e.g., only the "who" main slot is retained in the voice broadcast scenario, while "who," "what," and "count" are retained in the mechanical transformation scenario); the hard thresholds of core dimensions are dynamically adjusted according to the dimensional circuit breaker thresholds in different scenarios (e.g., the hard threshold of the time dimension is increased in the mechanical transformation scenario).

[0184] Diagnostic information is visualized: the reasons for failure, triggers, and diagnostic details of hard gating are all persistently stored and highlighted as "L2 gating failed" in the results queue and details page, preventing "false prosperity" caused by high average scores and directly guiding corrective actions. This allows users to directly pinpoint the core failure points, avoiding being misled by the overall score, and provides clear direction for material correction and rule optimization, solving the problem of "knowing failure but not knowing why it failed."

[0185] If the hard-gating detection result and the video detection strategy determine that the native video to be detected triggers hard gating, then the target detection result of the native video to be detected is determined to be a detection failure or the detection information of the native video to be detected is limited to below a preset upper limit, so as to obtain the target detection result of the native video to be detected.

[0186] For example, if hard gating is triggered (that is, the hard gating detection result meets the aforementioned hard gating trigger threshold, conditions, etc.), the final FAIL conclusion is directly output, and the subsequent detection process is terminated, or failure is not directly determined, but the comprehensive score is limited to below the preset upper limit (such as no more than 40 points) to ensure that the sample cannot reach the pass line, while retaining complete detection data for analysis.

[0187] If the hard gating detection result and the video detection strategy determine that the native video to be detected has not triggered the hard gating, then according to the soft fusion layer, based on the video detection strategy, the comprehensive detection result of the native video to be detected is determined according to at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result, and is used as the target detection result of the native video to be detected.

[0188] That is, if the hard gating is not triggered, the soft fusion layer will then perform a comprehensive fusion based on the six-dimensional consistency detection results, the key detection object detection results, and the normalized plug-in detection results.

[0189] In this application, soft fusion is a fine-grained scoring mechanism that weights and integrates all detection data without triggering hard gating, and outputs a comprehensive consistency score (as an example of the comprehensive detection result) to distinguish the quality of qualified samples and define the gray area range of ambiguous samples.

[0190] For example, the soft fusion layer strategy includes at least one of the following processing strategies:

[0191] Multi-dimensional weighted summation: Based on the six-dimensional consistency detection score (as an example of the target consistency detection result), a weighted comprehensive score is calculated according to the video consistency detection weight allocation strategy configured for the scenario routing. The weights are dynamically adjusted according to the prompt word profile and video profile. For example, the temporal and logical dimensions are weighted more in tutorial scenarios, while the semantic and entity dimensions are weighted more in broadcast scenarios.

[0192] Low-dimensional soft penalty: For normalized plug-in detection results with significantly low scores in a single dimension but not reaching the hard gate threshold, additional points are deducted to reinforce the weak link effect and avoid the problem of "high scores in multiple dimensions masking serious biases in a single dimension".

[0193] Slot coverage penalty: If the hit rate of a slot is insufficient, the overall score will be deducted proportionally.

[0194] It should be noted that the aforementioned hard-gating slot blocking is a rigid blocking when there is a severe shortage of critical slots, which is a rigid failure. Here, it is a flexible deduction when the proportion of ordinary slots is insufficient.

[0195] Maximum penalty cap: A maximum limit is set for the cumulative deduction of all soft penalty items to avoid excessive deductions that could distort the total score and ensure the stability and comparability of the overall score.

[0196] Confidence Calculation: Combining parameters such as sampling density, plugin confidence, and model output stability, the overall confidence of the detection result is calculated. The lower the confidence, the wider the gray area judgment range, avoiding the direct judgment of low-confidence samples as pass / fail.

[0197] Three-level interval division: Based on the comprehensive score and confidence level, the samples are divided into three conclusion intervals: Automatic pass interval, where the score is higher than the pass threshold and the confidence level meets the standard, PASS is directly output to directly output the target detection result (Final); Automatic failure interval, where the score is lower than the failure threshold and the hard gate is not triggered, FAIL is directly output; Gray zone interval, where the score is between the pass threshold and the failure threshold, or the confidence level is insufficient, and the process enters the review process.

[0198] Thus, soft fusion processing based on the soft fusion layer can prevent serious errors from being masked by average scores, while retaining the ability to score minor deviations with fine granularity. Furthermore, in this application, based on the collaborative processing of hard gating and soft fusion, the two form a hierarchical, progressive, and complementary adjudication system. The core logic is that hard gating sets the bottom line, while soft fusion assesses the quality. In terms of priority, hard gating has the highest priority; it is checked first, and soft fusion is only executed if it is not triggered. Functionally, hard gating is responsible for "filtering out absolutely unqualified samples" to solve the problem of missed judgments, while soft fusion is responsible for "distinguishing the degree of goodness or badness of samples" to solve the problem of refined evaluation. In terms of rules, hard gating rules are strongly dependent on the core requirements of the scenario, while soft fusion rules adapt to the overall scoring logic. Both are dynamically configured through adaptive routing to maintain consistent scenario adaptability.

[0199] Furthermore, in this application, when it is determined that the native video to be detected meets the target conditions, the review and evaluation layer generates the review and evaluation result of the native video to be detected based on the prompt word information and the native video to be detected, which serves as the target detection result of the native video to be detected. The review and evaluation layer then outputs the final target detection result (Final).

[0200] The target conditions include at least one of the following: the overall detection result of the original video to be detected is within the grayscale range, and the original video to be detected is a complex scene video.

[0201] In AI video quality inspection, many samples are not clearly passed or failed, but rather fall into a gray area. Existing solutions often only output the total score and a brief reason, failing to answer which frame or time period caused the problem, which cue word element was not matched, whether it was a subject error, action error, timing error, or style / logic issue, whether the failure of specialized plugins affected the conclusion or was only for reference, and which frames should be prioritized for manual review. Moreover, the lack of evidentiary work leads to problems such as high manual review costs, difficulty in reviewing misjudgments, and inability to accumulate rule improvements.

[0202] For example, the review and assessment layer (also known as the event-based review layer) is the core link in human-machine collaboration. For gray area samples that AI cannot accurately determine with 100% precision, as well as special scenario samples with complex logic, the event-based detailed review and structured evidence output significantly reduce the workload of manual review, while providing a data foundation for the subsequent manual feedback calibration closed loop.

[0203] In this application, the Layer3 event-based review process is initiated if any of the following conditions are met: the comprehensive score output by the soft fusion layer (as an example of the comprehensive detection result) is in the gray zone, and the automatic judgment cannot give a clear pass / fail conclusion; the video scene of the original video to be detected belongs to a complex special scene video with strong temporal sequence and multiple stages, such as mechanical deformation, multi-camera narrative, or complex tutorial, and is forced to enter the event-based review to ensure the accuracy of the verification.

[0204] Furthermore, based on the prompt word information and the original video to be tested, the review layer generates a review result (also known as an event-based review result) for the original video to be tested. This includes: splitting the complete original video to be tested into multiple independent event segments, completing the test segment by segment, and outputting event-level pass / fail conclusions and evidence, which is more accurate than global scoring and more adaptable to the verification needs of complex time-series scenarios. For example, for mechanical deformation scenarios, it is split into event segments such as initial form, disassembly, reconstruction, assembly, and final form according to the deformation stage; for multi-camera scenarios, it is split into independent shot event segments according to the actual storyboard; for tutorial demonstration scenarios, it is split into multiple operation event segments according to the steps.

[0205] Furthermore, the review results include at least one of the following: all sampled frames, low-scoring frames, frames with varying peak values, key time windows, scores per frame, best-matching query, weakness summary, event review results, and reasons for failure of specific plugins.

[0206] The sampled frames are screenshots of all video frames extracted in this test, used to completely reconstruct the input samples and support full manual review and verification. Low-scoring frames are key frames with the lowest consistency scores, used for direct manual location of the most problematic scenes without the need for frame-by-frame browsing, significantly improving review efficiency. Peak change frames are frames with the largest changes in video content and structure, used to adapt to deformation, transition, and action scenes, quickly locating key nodes of structural changes and transitions. Key time windows are continuous time periods with concentrated problems and abnormal score fluctuations, used to locate the intervals where time-series problems occur, adapting to the review of dialogue, audio-visual elements, and deformation processes. Each frame score is the single-frame consistency score corresponding to all sampled frames, used to present the overall score fluctuation trend and accurately locate the time nodes when problems occur. The best matching query is the prompt word fragment with the highest semantic matching degree to the video scene, used to reverse locate content in the prompt words that was not matched, quickly locating deviation points. The weakness summary is an automatically generated summary of problem dimensions and types, used for manual quick understanding of the core problem direction without the need for manual extraction from multi-dimensional results. The event review results are separate review conclusions for each stage / shot after the event is broken down into its components. These are used to quickly pinpoint the specific event segment where a problem occurred in complex scenarios, supporting detailed post-event analysis. The specialized plugin failure reasons are detailed breakdowns of anomalies in specialized plugins such as OCR, ASR, and identity consistency. These are used for precise location of specific issues, such as missing lines of dialogue, typos, or the specific location of identity drift.

[0207] Event-based review results can be persistently stored as image files (such as screenshots of various keyframes and sampling frames), indexed JSON files (such as structured storage of scores, timestamps, anomaly details, event results, and other data), and result detail fields (such as structured summary information integrated into the test results). These data are used for data input in the result page visualization, automatic generation of quality inspection reports, manual review assistance, and manual feedback calibration closed loop.

[0208] In the implementation of this application, the event-driven review chain of the event-driven review layer is a pre-defined review strategy.

[0209] Furthermore, in the implementation of this application, if the native video to be detected is determined to be a complex scene video based directly on the prompt word profile and video profile, then the target detection route is determined to be the review route. In this case, the prompt word profile and the native video to be detected are directly input into the event-based review layer, so that the event-based review layer reviews the native video to be detected based on the review strategy and obtains the review result, which is used as the target detection result of the native video to be detected, so that the user can perform detection review of the native video to be detected based on the review result.

[0210] For example, when the prompt word profile exhibits structured features of "strong temporal sequence" or "multi-stage" (e.g., it is identified as a mechanical transformation scene and contains clear transformation stages, or contains multiple independent scene transitions or strong step chains), the system will activate the event planner to break down the overall verification originally for the whole film into local evidence collection and review for each "event segment", thereby ensuring the temporal completion and consistency verification of long-process or complex evolution scenarios (such as step-by-step tutorials or transformation reassembly).

[0211] like Figure 4 As shown, taking a mechanical deformation scenario as an example, when the prompt word prompt scene pattern is identified as a mechanical deformation scenario, a dedicated detection route and standardized deformation schema are generated for the mechanical deformation scenario. This dedicated detection route includes a set of video detection plugins, a set of key video detection objects, a video consistency detection weight allocation strategy, video detection failure gating trigger rules, and a review strategy. The required layered detection pipeline includes a special plugin layer, a key detection object layer, a consistency detection layer, and a review layer.

[0212] Specifically, after the aforementioned prompt word pattern recognition, the prompt word scene pattern is determined to be the mechanical deformation scene pattern. The routing layer abandons the general detection baseline, loads the mechanical deformation-specific detection route, and the pipeline flows to the abstract action / complex event understanding plugin S-VLM. The core output of the plugin is the mechanical deformation structured description (i.e., the standardized deformation schema), which serves as the sole standard for full-scene verification.

[0213] The standardized variant schema includes the following fields:

[0214] Scene type identifier, value is mechanical_transformation;

[0215] The initial morphological structure includes an initial main body category, which is used to constrain the appearance of this type of object to appear at the beginning of the original video to be detected; and an appearance array, which is used to store appearance description keywords, which can be used for main body consistency verification.

[0216] The deformable target structure includes the target category, used to verify the type of the main body at the end of the video; and the style feature array, used to score the style consistency dimension.

[0217] The transformation stage list contains a segment of mandatory time-series events. Each stage has a unified definition for its fields, including: stage number S1 / S2 / S3 (used for time window matching and event scoring and tracing); stage name (used for review reports and evidence presentation, e.g., S1 stage initial form event, S2 stage unlocking and disassembly event, S3 stage structural reconstruction event); Boolean identifiers (e.g., S1 stage true, S2 stage true, S3 stage true, where true indicates that the stage cannot be omitted and its absence directly triggers hard gating); the set of components that must appear or change in this stage; the dynamic actions required to be performed in this stage; and, of course, detailed assembly events and final form events.

[0218] The set of key detection objects throughout the entire process (i.e., the core detection objects throughout the entire process); necessary structural changes within the entire video sequence, without which points will be deducted / failure will be triggered.

[0219] The blacklist of prohibited failures (hard-gate trigger keywords) includes, for example, no intermediate transformation process, directly switching to the final form; only light and shadow effects, with no physical disassembly and reassembly of the mecha; the initial main image drifts throughout the process, and the front and back do not belong to the same mechanical object.

[0220] Furthermore, based on the Special Plugins layer, the Key Detection Object layer, the Consistency Detection layer, and the Review and Approval layer, an event plan (including initial form event, unlocking and disassembly event, structural reconstruction event, detailed assembly event, and final form event) is constructed according to the standardized deformation schema to perform detection processing on the native video to be detected, and the target detection results of the native video to be detected in the mechanical deformation scene category are obtained.

[0221] First, the abstract action / complex event understanding plugin S-VLM dynamically generates key object inspection instructions for each deformation stage based on the deformation stage list and the set of key detection objects in the whole process. These instructions are then sent to the visual big model of the key detection object layer to verify the status of the key detection objects segment by segment. This is to evaluate the initial form establishment, the sequential occurrence of deformation stages, and the achievement of the final form. If a key action slot is missing or any item in the prohibited failure blacklist is hit, L2 hard gate failure will be triggered directly, and the review layer will be skipped. The reason for failure (such as missing deformation process / light effect replacement deformation, etc.) will be highlighted.

[0222] Furthermore, during the phased verification of key objects, a global six-dimensional supplementary verification is performed synchronously and in parallel based on the consistency detection layer (Layer2 Extract) to compensate for the blind spots of the phased local verification. For example, the consistency of the main body is verified by tracking the initial mechanical main body features throughout the process and identifying the drift of the main body identity; the consistency of time and logic (that is, the consistency of action and physical law verification) is verified to check whether the trajectory of the mecha deck unlocking, disassembling and reassembling conforms to the physical space logic and time sequence setting, and to verify whether the phase sequence is correct; the consistency of space / style is verified to ensure that the background is not distorted or abnormally distorted due to the reconstruction of the main body in the range of drastic deformation of the image, and the style and material of the mecha are consistent throughout the process. In this way, a six-dimensional consistency detection score is obtained (as an example of the initial consistency detection result).

[0223] Furthermore, if hard gating is not triggered, a review is conducted based on the review and evaluation layer (Layer 3) in combination with low-scoring frames, change peak frames, and component evidence. The completion status of the aforementioned stage events S1 / S2 / S3 output by the special plugin (such as whether the deformation process occurs in stages and whether structural changes occur in key parts) is scored independently to obtain stage scores. This special score is then fused with the six-dimensional consistency detection score output by the consistency detection layer (Layer 2 Extract) based on a video consistency detection weight allocation strategy to obtain the aforementioned hard failure reasons and review evidence.

[0224] Furthermore, even if hard gating is not triggered, if the system determines that an error such as "critical parts are not reconstructed" occurs, which seriously violates the standardized deformation schema definition, in addition to triggering hard gating, a special score cap can also be implemented to ensure that mechanical deformation scenarios with high consistency requirements are rigorously evaluated.

[0225] Furthermore, existing technologies that uniformly employ global fixed thresholds and fixed dimensional weights have significant drawbacks: the quality inspection standards for mechanical transformation, voice-over, dialogue, and commentary videos vary greatly. A single set of parameters is too strict for one type of scenario and too lenient for another, and it cannot be automatically adjusted to keep up with changes in business materials and reviewers' standards. It requires manual configuration adjustments, which are costly and result in outdated standards. In other words, the scoring thresholds and dimensional weights for different content modes should not be completely consistent. For example, mechanical transformation videos should place greater emphasis on process completeness and temporal logic, voice-over videos should prioritize the main subject, dialogue, and audio-visual synchronization, and abstract animations should not rely excessively on entity detection. If fixed thresholds and weights are used for all scenarios in the long term, it will be too strict in some scenarios and too lenient in others.

[0226] In this application, a novel approach is adopted: content profile bucketing. Based on manual feedback, a bucketing calibration closed loop is used to manually review all raw videos to be detected, forming a complete closed loop of automatic detection, manual review feedback, bucketing statistical fitting, updating the global detection strategy, and adaptive detection in the next round.

[0227] For example, the system obtains the manual review results of each native video to be detected, based on the prompt word information corresponding to each native video to be detected; it then divides the native videos to be detected and their corresponding manual review results into different types of content buckets based on the prompt word profiles of each native video to be detected; it determines the detection route adjustment parameters corresponding to each content bucket based on the manual review results of the native videos to be detected included in each content bucket; and it updates the dedicated detection routes for native videos to be detected of the same type as the content buckets based on the detection route adjustment parameters corresponding to the content buckets, so as to perform detection processing on the native videos to be detected based on the adjusted dedicated detection routes.

[0228] Specifically, the manual review results of multiple original videos to be tested are collected, and fine-grained binning and isolation processing is performed based on the prompt word profile and video profile corresponding to each original video to be tested to obtain different types of content bins.

[0229] Content buckets include, for example, mechanical deformation videos, general deformation videos without special camera shapes, multi-character dialogue + multi-camera narrative videos, single-person on-camera voice-over advertisements, and videos featuring on-camera commentary.

[0230] Different buckets store samples independently and calculate calibration parameters independently, without interfering with each other. For example, the semantic threshold of the oral broadcast bucket will not affect the temporal weight of the mechanical deformation bucket, completely solving the problem of false positives and false negatives caused by the "one-size-fits-all" approach to global parameters.

[0231] For each content bucket, all manually labeled samples within the bucket are periodically aggregated, and persistent calibration workpieces are generated through statistical analysis and reward model fitting.

[0232] In one implementation of this application, the persistent calibration workpiece includes at least one of the following seven types of information:

[0233] Semantic pass threshold: Extract all manually labeled pass samples and failure samples (i.e., positive and negative samples) in the bucket, calculate the 25th quantile Q1 of the pass samples and the 75th quantile Q3 of the failure samples respectively, and the mean of the two is the dynamic threshold. If the number of effective samples in the bucket is insufficient, no statistical fitting is performed, and the default threshold is lowered. The final threshold is forcibly locked in the range [0.45, 0.9] to avoid extreme samples causing the threshold to deviate from the reasonable business range. In this way, when replacing the global default semantic pass line in video detection, the mechanical deformation bucket will automatically raise the semantic + temporal threshold, and the comment bucket will automatically lower the semantic hard threshold.

[0234] Dimensional Weights: By collecting historical scores and final human judgment labels for the original videos to be detected corresponding to each scene mode, a preference dataset is constructed. Each sample stores its original six-dimensional score plus the final human pass / fail label. Using a reward model calibrator, the human judgment is used as the real reward label to inversely optimize the optimal weight bias of the six dimensions (semantic / entity / space / time / style / logic). Different buckets automatically adjust the dimensional emphasis. The tutorial bucket increases the weight of time and logic, while the broadcast bucket increases the weight of entity and semantics. No manual configuration of routing weights is required.

[0235] The upper and lower bounds of the CLIP similarity (i.e., multimodal feature similarity) to the score mapping: For content buckets that use CLIP multimodal features as the scoring backend, the 75th percentile of the failed samples is calculated as the lower bound of the mapping, and the 25th percentile of the samples is used as the upper bound of the mapping. The upper and lower bound parameters of the mapping are dynamically updated to calibrate the linear mapping from feature vector similarity to semantic scores of 0 to 1.

[0236] Sample size metrics: The total number of valid samples that hit and settled in different buckets, the number of manually labeled samples that passed, and the number of manually labeled samples that failed are used as the basis for determining the confidence level of the strategy calibration. For buckets with too few samples, the newly fitted thresholds and weights will not take effect, and historically stable parameters will continue to be used to prevent a small number of extreme labels from contaminating the detection strategy.

[0237] Positive and negative sample mean index: The mathematical expectation of three core indicators on positive and negative samples is calculated by the sample mean and the failure sample mean. These are the average score of semantic consistency, the average score of prompt word element hit, and the average CLIP feature similarity. The index monitors the drift of the sample distribution within the bucket. When the average score of the new material deviates significantly from the historical average, recalibration is triggered to adapt to the style change of the new material.

[0238] Fit quality index: The quality assessment index of the model fit is generated simultaneously during the process of fitting the dimensional weights of the reward model, which helps to judge the reliability of the weight adjustment.

[0239] Calibration workpiece fingerprint: After serializing the calibration strategy generated by all buckets, calculate the SHA256 hash value and extract the first 16 bits as the global calibration fingerprint to ensure that the strategy loaded at runtime has strict consistency and version traceability.

[0240] Furthermore, upon initiating the task detection, the Router layer identifies the content bucket to which the current native video to be detected belongs; it reads the complete set of pre-calculated calibration artifacts (thresholds, weights, CLIP mapping upper and lower bounds, etc.) for that bucket; it automatically overwrites the global default baseline parameters, updates the detection route, six-dimensional scoring weights, hard-gated semantic thresholds, and CLIP score mapping rules; the entire pipeline (slot verification, six-dimensional scoring, hard gating, and comprehensive fusion) is executed based on the adaptive parameters of this bucket; after the task is completed, the annotation results are manually reviewed and written back to the corresponding content bucket, accumulating sample data, waiting for the next round of timed batch refitting of calibration parameters.

[0241] For example, an auditor might mark a mechanical deformation sample as failed on the results details page, noting "only light effect occlusion, no component reconstruction." This manual feedback, along with the content mode, video mode, dimensional score, semantic statistics, and plugin results for the detection task, is then written into the calibration data. Once a certain number of samples from the same content bucket are collected, the system updates the bucket's semantic threshold, dimensional weights, and similarity mapping. Subsequent similar tasks will read this calibration strategy, making mechanical deformation tasks prioritize process integrity and temporal logic, thereby reducing misjudgments.

[0242] This application provides a video detection method capable of performing the following core capabilities: prompt word profiling (identifying standard, dialogue, voiceover, comment, mechanical deformation, etc.), adaptive routing (adjusting effective prompt words, applicable slots, thresholds, and weights), specialized plugins (with existing specialized capabilities or interfaces such as S1 / S2 / S3 / S4 / S5 / S-VLM), layered pipeline (with existing routing layer, specialized plugin layer, key detection object layer, consistency detection layer, hard gating layer, soft fusion layer, review and approval layer, and final output layer), hard gating layer (integrating slot coverage, dimension thresholds, penalty coefficients, and score fusion strategies to achieve rigid constraints), mechanical deformation specialization (relying on mechanical deformation profiling, deformation stage sequence, core component constraints, event-based review, and change peak sampling to complete time-series scene fine-grained verification), review evidence (capable of persistently storing sampled frames, low-score abnormal frames, retrieval indexes, and defect / weakness summaries), and feedback calibration (based on prompt word profiling bucketing strategy, automatically generating judgment thresholds, dimension weights, and feature similarity mapping parameters corresponding to each scene), thus possessing the foundation for technological implementation. Compared with existing technologies, this solution can reduce scene misjudgment, improve detection targeting, avoid distortion caused by plug-in fusion, strengthen the interception of key errors, reduce review costs, achieve continuous calibration, and has high scalability.

[0243] The video detection method proposed in this application is an AI-native video consistency detection method based on the fusion of prompt word scene recognition and plugin signal standardization. This method is designed for AI-native video quality inspection scenarios. After inputting prompt word information and the native video to be detected, it first identifies the scene mode to which the prompt words belong, then generates an adaptive detection route, subsequently executes a hierarchical hybrid detection pipeline, and converts the results of specialized plugins into a unified standardized signal for strategy adjustment, dimensional scoring, conclusion gating, or evidence presentation. Finally, it outputs an interpretable, verifiable, and sustainably calibrable consistency detection result.

[0244] Furthermore, in the implementation method of this application, the preferred video detection method is a complete end-to-end detection scheme, which covers the entire process of scene recognition, adaptive routing, hierarchical detection, signal standardization, gating fusion, verification, and calibration, maximizing the coverage of technical barriers.

[0245] Specifically, before detection, the prompt words are structurally analyzed and their scene patterns are identified to determine whether they belong to different types such as standard scenes, dialogue scenes, spoken scenes, mechanical deformation scenes, and commentary / opinion scenes. An adaptive detection route is generated based on the prompt word scene patterns to determine the special plugins to be triggered, the key slot set, gating rules, weight coverage strategies, and whether to enter the event-based review. A layered hybrid detection pipeline is constructed, which processes basic quality detection, key slot detection, six-dimensional consistency scoring, hard gating, soft fusion, gray area review, and final output in layers. The output of special plugins is standardized into strategy signals, normalized scoring signals, conclusion gating signals, or evidence display signals, rather than directly mixing the original scores. For gray areas or complex scenes, evidence artifacts such as low-scoring frames, changing peak frames, key time windows, event review results, and weakness summaries are output. Content buckets are constructed based on manual review feedback, and semantic thresholds, dimensional weights, and similarity mappings are continuously calibrated.

[0246] Thus, compared to existing technologies, this approach reduces misjudgments caused by sharing the same rules across different scene modes through prompt word scene recognition; improves the targeting of detection in complex video scenes and the consistency detection accuracy in complex AI native video scenes through adaptive detection routing; avoids distortion caused by direct fusion of original scores from heterogeneous plugins through plugin signal standardization, improving the stability and interpretability of results from different specialized plugins entering the main process; prevents serious errors from being masked by high-scoring dimensions through separation of hard gating and soft fusion, improving the interception capability of key failure items; improves the efficiency of manual review and the ability to review misjudgments through gray zone evidence artifacts (i.e., review results), reducing the cost of manual review of gray zone samples; enables the system thresholds and weights to be continuously optimized with human feedback through content bucket calibration, thereby continuously approaching real business standards; and enhances the scalability of new scenes, plugins, and rules through a layered pipeline and plug-in structure, providing an extensible framework for subsequent additions of scenes, plugins, and business rules.

[0247] Furthermore, in one implementation of this application, the preferred video detection method is as follows: acquiring the original video to be detected, corresponding prompt word information, and video detection rule configuration information; cleaning, structurally parsing, and pattern recognition of the prompt word information; generating detection routes based on prompt word scene patterns; determining a set of specialized plugins, a set of key slots, gating rules, and weight coverage strategies; executing basic quality detection, slot detection, six-dimensional consistency scoring, and specialized plugin detection in a layered hybrid detection pipeline; converting the output of specialized plugins into strategy signals, normalized scoring signals, conclusion gating signals, or evidence-only display signals; performing hard gating based on key slots, multi-dimensional scores, and standardized signals; performing comprehensive fusion when hard gating is not triggered to obtain a comprehensive score, confidence level, and automatic conclusion; generating event-based review results and evidence artifacts for gray areas or complex specialized scenes; collecting manual review feedback and updating operating parameters according to content buckets; and outputting the final detection conclusion, dimension details, failure reasons, review evidence, and suggestions.

[0248] Furthermore, the standardization process for plugin detection results is as follows: acquire multiple specialized plugin outputs; determine the input modality, output semantics, failure level, and confidence interval of the plugin output; classify the plugin results into policy signals, normalized scoring signals, conclusion gating signals, or evidence-only signals; input them into the routing, scoring, gating, or evidence links according to the signal type; and output stable and interpretable comprehensive detection results.

[0249] Furthermore, in cold start or insufficient sample scenarios, manual feedback calibration can be disabled initially, retaining only prompt word scene recognition; adaptive routing; hierarchical detection; plug-in signal standardization; hard gating and soft fusion; and gray area evidence output.

[0250] Furthermore, when the original video to be detected is a complex deformation video, a special solution for mechanical deformation can be adopted. The steps include: identifying mechanical deformation cue words; generating deformation stages, key parts, target shapes, and prohibited failure items; establishing an event plan; evaluating the initial shape, process coverage, sequence, component coupling, final shape, and continuity of identity style; triggering hard gating for cases such as directly cutting the final shape or only lighting effects without structural changes; and outputting low-resolution frames, change peak frames, and stage evidence.

[0251] For example, when detecting a scene mode of a mechanical transformation advertising video, the prompt message might be: "A pure white mechanical cat stands on the ground, with blue energy patterns flowing, the machine deck gradually unlocks, flips outward, slides, the limbs and spine are reconstructed, and finally it transforms into a futuristic humanoid robot."

[0252] First, the prompt information is identified as a mechanical deformation or structural reorganization scenario, generating an initial main form, target deformation form, multiple deformation stages, key detection objects, and prohibited failure items. The routing layer adaptively generates a dedicated detection route based on this mechanical deformation profile, proactively increasing the weight of the time and logic dimensions, and setting the main body and actions as key slots. Peak sampling of the rate of change is enabled to protect frames with obvious structural changes. The entire process of initial form, disassembly, reconstruction, assembly, and final form is evaluated and verified segmented according to the layered detection pipeline. During the detection process, if the video only shows a flashing light effect followed by the robot appearing directly, it is determined that the video only has light and shadow flashing effects and no real mechanical structural disassembly and reorganization changes, or all intermediate deformation stages are skipped and the final robot form is directly displayed, resulting in a video failure conclusion. The specific illegal deformation stage is accurately located, and key evidence such as low-scoring abnormal frames and structural change peak frames are retained simultaneously. Corresponding standardized review and rectification suggestions are generated to complete the intelligent detection and judgment of the entire mechanical deformation scenario.

[0253] In this application, based on the video detection method provided, when the focus of the video to be detected is on scenarios such as spoken advertisements, dialogue, and digital human narration, a dialogue / spoken audio-visual consistency detection method can be adopted. The steps include: identifying the dialogue script or spoken scene; extracting lines, speakers, visual main lines, and key scene requirements; triggering the ASR / line verification plugin and the audio-visual consistency plugin; detecting line completeness, mismatches, missing lines, cross-line misappropriation, and audio-visual synchronization; converting serious problems into conclusion gates, and using minor deviations as soft penalties or evidence; weakening or skipping unnecessary location slots. In this way, quality inspection of spoken advertisements and dialogue-based materials is achieved.

[0254] For example, in short video dialogue detection, a prompt might be that two cats are required to engage in a multi-turn dialogue, presented in a specified order. The system would then identify the prompt as a multi-speaker dialogue script scenario; extract the speakers, dialogue entries, and visual thread; trigger the ASR / dialogue verification plugin and the audio-visual consistency plugin; the dialogue verification plugin would perform word or segment-level alignment, identifying missing lines, mismatches, and cross-line misappropriation; serious missing lines would be used as a conclusion gating signal, while minor deviations would be used as evidence or a soft penalty; and the system would output a comprehensive conclusion, dialogue detection results, and reasons for failure.

[0255] For example, in the detection of spoken advertising videos, the prompt might be: "A woman is required to sit in front of the camera to introduce the product's selling points and recite the specified script in its entirety." The prompt information would then be identified as a spoken or narrated scene. Visual detection prompts would be summarized to avoid mistaking entire marketing copy errors for visual entity requirements. The system would retain the verification of the main character while downplaying irrelevant location slots. An ASR / dialogue verification plugin would check the script's completeness. An audio-visual consistency plugin would detect any significant misalignment between the audio and video. The system would then output a comprehensive score, script completeness, audio-visual relationship, and verification evidence.

[0256] like Figure 5 As shown, this application also provides a video detection system, including a prompt word structured parsing and scene recognition module, a routing module, a plug-in signal standardization adapter layer, a hard gating + event-based review and adjudication module, a soft fusion module, and a manual feedback bin calibration module.

[0257] The prompt word structure parsing and scene recognition module is used to determine the original video to be detected and the prompt word information corresponding to the original video to be detected; and to determine the prompt word profile based on the prompt word information.

[0258] The routing module is used to generate target detection routes for the native video to be detected based on the prompt word profile. The target detection routes include standard routes, dedicated detection routes, and review routes. The dedicated detection routes include at least one of the following: a set of video detection plugins, a set of key video detection objects, a video consistency detection weight allocation strategy, and a video detection failure gating trigger rule.

[0259] That is, the layered detection pipeline in this application includes Router-Special Plugins-Layer1-Layer2 Slot-Layer2 Extract, which is deployed in the routing module to generate a dedicated detection route for the native video to be detected according to the routing layer based on the prompt word profile, video profile, and video detection rule configuration information; according to the special plugin layer, based on the target detection plugins included in the video detection plugin set, the native video to be detected is processed according to the video detection strategy to obtain the plugin detection result of the native video to be detected. The target detection plugin set includes at least one detection plugin among multi-camera narrative consistency plugin, identity consistency plugin, text detection plugin, audio-visual consistency plugin, speech recognition verification plugin, and complex event understanding plugin; according to the key detection object layer, based on the video detection strategy and the video key detection object set, the native video to be detected is detected to see if it meets the key detection object requirements to obtain the key detection object detection result of the native video to be detected, which is used as the native video to be detected. The process involves: obtaining the target detection result of the raw video; generating a video frame extraction strategy for the raw video to be detected based on the video detection strategy according to the consistency detection layer; performing frame extraction processing on the raw video to be detected based on the video frame extraction strategy to obtain a sampled frame sequence; performing consistency verification processing on the sampled frame sequence based on the video detection strategy to obtain the initial consistency detection result of the raw video to be detected, wherein the consistency verification processing includes at least one of semantic consistency detection processing, entity consistency detection processing, spatial consistency detection processing, temporal consistency detection processing, style consistency detection processing, and logical consistency detection processing; and adjusting the initial consistency detection result of the raw video to be detected based on the key detection object detection result and the video detection strategy to obtain the target consistency detection result of the raw video to be detected, which is then used as the target detection result of the raw video to be detected.

[0260] The plugin signal standardization adapter layer is used to perform format unification processing on the plugin detection results based on the target adapter corresponding to the target detection plugin and according to the plugin signal standard protocol, so as to obtain the normalized plugin detection result of the original video to be detected, which is used as the target detection result of the original video to be detected. The normalized plugin detection result includes a detection result processing intent. If the detection result processing intent indicates that the normalized plugin detection result is a policy signal, the normalized plugin detection result is returned to the routing layer so that the routing layer can update the dedicated detection route and video detection policy based on the normalized plugin detection result. If the detection result processing intent indicates that the normalized plugin detection result is a scoring signal, the normalized plugin detection result is sent to the soft fusion layer so that the soft fusion layer can determine the comprehensive detection result of the native video to be detected based on at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result. If the detection result processing intent indicates that the normalized plugin detection result is a gating signal, the normalized plugin detection result is sent to the hard gating layer so that the hard gating layer can determine the hard gating detection result of the native video to be detected based on at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result. If the detection result processing intent indicates that the normalized plugin detection result is an evidence signal, the normalized plugin detection result is used as the target detection result of the native video to be detected.

[0261] The hard gate + event-based review and adjudication module (i.e., Hard Gate and Layer3 are deployed in the hard gate + event-based review and adjudication module) is used to generate a review and adjudication result of the native video to be detected based on the prompt word information and the native video to be detected when it is determined that the native video to be detected meets the target conditions. This result serves as the target detection result of the native video to be detected. The target conditions include at least one of the following: the comprehensive detection result of the native video to be detected is within the grayscale range, and the native video to be detected is a complex scene video.

[0262] The soft fusion module (i.e., Score Fusion is deployed in the soft fusion module) is used to determine the comprehensive detection result of the native video to be detected if it is determined that the native video to be detected has not triggered hard gating. According to the soft fusion layer, based on the video detection strategy, the comprehensive detection result of the native video to be detected is determined based on at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result, and is used as the target detection result of the native video to be detected.

[0263] The manual feedback binning calibration module is used to obtain the manual review results of each raw video to be tested, obtained by the user based on the prompt word information corresponding to each of the multiple raw videos to be tested. Based on the prompt word profiles corresponding to each raw video to be tested, the multiple raw videos to be tested and their corresponding manual review results are binned to obtain different types of content bins. Based on the manual review results of the raw videos to be tested included in each content bin, the detection route adjustment parameters corresponding to each content bin are determined. The dedicated detection routes for raw videos to be tested of the same type as those in the content bins are updated according to the detection route adjustment parameters, so that the raw videos to be tested can be processed based on the adjusted dedicated detection routes.

[0264] Furthermore, the layered inspection pipeline includes a final layer that outputs the final inspection information. This includes, for example, the final pass / fail / uncertain conclusion, overall score, six-dimensional details, key slot results, specialized plug-in results, failure reasons, and verification evidence artifacts.

[0265] The video detection method provided in this application is not a simple business rule or manual review process, but a full-link video detection process that performs calculations on video, image frames, audio, OCR / ASR output, structured data of prompt words, and detection workpieces. It does not just call a large model for scoring, but has a clear detection mechanism including data flow, signal conversion, gating fusion, and evidence generation.

[0266] The video detection method provided in this application can be applied to servers, workstations, local computers, cloud-based quality inspection platforms, or other electronic devices with video processing capabilities.

[0267] That is, the embodiments of this application also provide an electronic device, including a processor and a memory communicatively connected to the processor, wherein the memory stores computer execution instructions, and the processor executes the computer execution instructions stored in the memory to implement the video detection method provided in the implementation of this application.

[0268] This application also provides a chip for executing instructions, which is used to execute the video detection method described in the above embodiments.

[0269] This application also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on the processor of an electronic device, the processor of the electronic device performs the technical solution of the video detection method described in the above embodiments.

[0270] In some possible implementations, various aspects of the methods provided in this application can also be implemented as a program product, which includes program code. When the program product is run on the processor of an electronic device, the program code is used to cause the processor of the electronic device to perform the steps in the methods of the various exemplary implementations of this application described above. For example, the electronic device can perform the video detection method described in the embodiments of this application.

[0271] The program product may take the form of any combination of one or more readable media. A readable medium may be a readable data medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CDROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0272] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium, and when the at least one processor executes the computer program, it can implement the technical solution of the video detection method in the above embodiments.

[0273] It should be noted that, in addition to the specific embodiments described above, those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Although the description of this application is presented in conjunction with preferred embodiments, this does not mean that the features of this application are limited to this implementation. On the contrary, the purpose of describing the application in conjunction with the implementation is to cover other options or modifications that may be derived from this application. To provide a thorough understanding of this application, many specific details are included in the above description, and this application may also be implemented without using these details. Furthermore, to avoid confusion or obscuring the focus of this application, some specific details will be omitted in the description. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0274] It should be noted that in this specification, similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0275] It should be noted that the terms "first," "second," and "third" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0276] It should be noted that some structural or methodological features may be shown in the accompanying drawings in a specific arrangement and / or order. However, it should be understood that such a specific arrangement and / or order may not be necessary. Rather, in some embodiments, these features may be arranged in a manner and / or order different from that shown in the illustrative drawings. Furthermore, including structural or methodological features in a particular figure does not imply that such features are required in all embodiments, and in some embodiments, these features may be omitted or may be combined with other features.

[0277] Although this application has been illustrated and described with reference to certain preferred embodiments, those skilled in the art should understand that the above description is a further detailed explanation of the application in conjunction with specific implementations, and should not be construed as limiting the specific implementation of the application to these descriptions. Those skilled in the art can make various changes in form and detail, including some simple deductions or substitutions, without departing from the spirit and scope of this application.

Claims

1. A video detection method, characterized in that, The method includes: Determine the original video to be detected and the prompt word information corresponding to the original video to be detected; Determining a prompt word profile based on the prompt word information includes: performing feature extraction processing on the prompt word information to obtain core verification elements, and generating a prompt word version corresponding to the video detection scenario based on the core verification elements and the video detection scenario; performing pattern recognition processing on the prompt word information based on target pattern recognition to obtain a prompt word scene pattern; and obtaining the prompt word profile based on the core verification elements and the prompt word scene pattern. The prompt word profile includes at least one of the following: the prompt word scene pattern, the prompt word version, the applicable detection object, and the prompt word structure schema. The prompt word version includes at least one of the following: visual prompt words, entity prompt words, and analytical prompt words. Determine the video profile of the native video to be detected, wherein the video profile includes video scene modes; Based on the prompt word profile, a target detection route is generated for the native video to be detected. The target detection route includes a dedicated detection route, which includes at least one of a video detection plugin set, a video key detection object set, a video consistency detection weight allocation strategy, and a video detection failure gating trigger rule. The dedicated detection route included in the target detection route generated based on the prompt word profile includes: generating a video detection strategy for the native video to be detected based on the prompt word profile; and mapping the prompt word scene patterns included in the prompt word profile and the video scene patterns included in the video profile to the parameters of the video detection strategy according to the routing layer included in the layered detection pipeline through a built-in rule mapping engine based on a pre-configured correspondence relationship to obtain the dedicated detection route. The target detection route is used to perform detection processing on the original video to be detected, and the target detection result of the original video to be detected is obtained.

2. The video detection method according to claim 1, characterized in that, The target pattern recognition method includes at least one of the following processing methods: dictionary extraction processing method, regular pattern matching method, and heuristic multidimensional feature scoring method.

3. The video detection method according to claim 2, characterized in that, The method further includes: The video detection rule configuration information is determined, including the gating threshold, weight coefficient, plugin timeout limit information, and sampling parameters. Mapping the prompt word scene patterns included in the prompt word profile and the video scene patterns included in the video profile to the parameters of the video detection strategy to obtain the dedicated detection route includes: If a target scene combination pattern can be generated based on the prompt word scene pattern and the video scene pattern, then the target scene combination pattern is mapped to the parameters of the video detection strategy to generate the dedicated detection route for the native video to be detected; if a target scene combination pattern cannot be generated based on the prompt word scene pattern and the video scene pattern, then the prompt word scene pattern is mapped to the parameters of the video detection strategy to generate a first detection route, and the video scene pattern is mapped to the parameters of the video detection strategy to generate a second detection route. Based on the video detection rule configuration information, the dedicated detection route for the native video to be detected is generated according to the first detection route and the second detection route.

4. The video detection method according to claim 3, characterized in that, The method further includes: Based on the target detection route, the target detection process is performed on the original video to be detected to obtain the target detection result of the original video to be detected, including: According to the layered detection pipeline, the native video to be detected is processed based on the dedicated detection route and the video detection strategy to obtain the target detection result of the native video to be detected.

5. The video detection method according to claim 4, characterized in that, When the dedicated detection route includes a set of video detection plugins, the layered detection pipeline includes a dedicated plugin layer. Following the layered detection pipeline, based on the dedicated detection route and the video detection strategy, the native video to be detected is processed to obtain the target detection result of the native video to be detected, including: According to the special plugin layer, based on the target detection plugin set included in the video detection plugin set, the original video to be detected is detected and processed according to the video detection strategy to obtain the plugin detection result of the original video to be detected. The target detection plugin set includes at least one detection plugin among multi-camera narrative consistency plugin, identity consistency plugin, text detection plugin, audio-visual consistency plugin, speech recognition verification plugin, and complex event understanding plugin. Based on the target adapters corresponding to the target detection plugins in the target detection plugin set, the plugin detection results are processed according to the plugin signal standard protocol to obtain the normalized plugin detection results of the original video to be detected, which are used as the target detection results of the original video to be detected.

6. The video detection method according to claim 5, characterized in that, When the dedicated detection route includes a set of key detection objects, the hierarchical detection pipeline includes a key detection object layer. Following the hierarchical detection pipeline, based on the dedicated detection route and the video detection strategy, the native video to be detected is processed to obtain the target detection result of the native video to be detected, including: According to the key detection object layer, based on the video detection strategy and the set of key detection objects for the video, it is determined whether the original video to be detected meets the key detection object requirements, and the key detection object detection result of the original video to be detected is obtained as the target detection result of the original video to be detected.

7. The video detection method according to claim 6, characterized in that, When the dedicated detection route includes a target weight allocation strategy, the hierarchical detection pipeline includes a consistency detection layer. Following the hierarchical detection pipeline, based on the dedicated detection route and the video detection strategy, the native video to be detected is processed to obtain the target detection result of the native video to be detected, including: According to the consistency detection layer, a video frame extraction strategy for the original video to be detected is generated based on the video detection strategy. Based on the video frame extraction strategy, the original video to be detected is subjected to frame extraction processing to obtain a sampled frame sequence; Based on the sampled frame sequence, a consistency check is performed on the sampled frame sequence according to the video detection strategy to obtain the initial consistency detection result of the original video to be detected. The consistency check includes at least one of semantic consistency detection, entity consistency detection, spatial consistency detection, temporal consistency detection, style consistency detection, and logical consistency detection. Based on the detection results of the key detection objects and the video detection strategy, the initial consistency detection results of the original video to be detected are adjusted to obtain the target consistency detection results of the original video to be detected, which are used as the target detection results of the original video to be detected.

8. The video detection method according to claim 7, characterized in that, When the dedicated detection route includes target gating triggering rules, the layered detection pipeline includes a hard gating layer. Following the layered detection pipeline, based on the dedicated detection route and the video detection strategy, the native video to be detected is processed to obtain the target detection result of the native video to be detected, including: Based on the hard gating layer, the hard gating detection result of the original video to be detected is determined based on at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result; wherein... If the hard-gating detection result and the video detection strategy determine that the native video to be detected triggers hard gating, then the target detection result of the native video to be detected is determined to be a detection failure or the detection information of the native video to be detected is limited to below a preset upper limit, so as to obtain the target detection result of the native video to be detected.

9. The video detection method according to claim 8, characterized in that, The dedicated detection route also includes a soft fusion strategy. If it is determined that the native video to be detected has not triggered the hard gate, the layered detection pipeline includes a soft fusion layer. According to the layered detection pipeline, based on the dedicated detection route and the video detection strategy, the native video to be detected is processed to obtain the target detection result of the native video to be detected, including: According to the soft fusion layer, based on the video detection strategy, the comprehensive detection result of the original video to be detected is determined according to at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result, and is used as the target detection result of the original video to be detected.

10. The video detection method according to claim 9, characterized in that, The dedicated detection route also includes a review and approval strategy. The layered detection pipeline includes a review and approval layer. Following the layered detection pipeline, the native video to be detected is processed based on the dedicated detection route to obtain the target detection result of the native video to be detected, including: When it is determined that the native video to be detected meets the target conditions, the review and evaluation layer generates a review and evaluation result of the native video to be detected based on the prompt word information and the native video to be detected, which serves as the target detection result of the native video to be detected. The target conditions include at least one of the following: the comprehensive detection result of the native video to be detected is within the grayscale range, and the native video to be detected is a complex scene video.

11. The video detection method according to claim 10, characterized in that, The normalized plugin detection result includes the detection result processing intent, and the method further includes: If the detection result processing intent indicates that the normalization plugin detection result is a policy signal, then the normalization plugin detection result is returned to the routing layer so that the routing layer updates the dedicated detection route and the video detection policy according to the normalization plugin detection result; If the detection result processing intent indicates that the normalized plugin detection result is a scoring signal, then the normalized plugin detection result is sent to the soft fusion layer so that the soft fusion layer determines the comprehensive detection result of the original video to be detected based on at least one of the normalized plugin detection result, the key detection object detection result, and the target consistency detection result; If the detection result processing intent indicates that the normalization plugin detection result is a gating signal, then the normalization plugin detection result is sent to the hard gating layer, so that the hard gating layer determines the hard gating detection result of the original video to be detected based on at least one of the normalization plugin detection result, the key detection object detection result, and the target consistency detection result; If the detection result processing intent indicates that the normalization plugin detection result is an evidence signal, then the normalization plugin detection result is used as the target detection result of the original video to be detected.

12. The video detection method according to claim 11, characterized in that, The method further includes: Obtain the manual review results of each of the original videos to be detected, obtained by the user based on the prompt word information corresponding to each of the original videos to be detected; Based on the prompt word profiles corresponding to each of the original videos to be detected, the multiple original videos to be detected and their corresponding manual review results are binned to obtain different types of content bins. Based on the manual review results of the original videos to be detected included in each content bucket, determine the detection route adjustment parameters corresponding to each content bucket; The dedicated detection route for the original video to be detected, which is of the same type as the content bucket, is updated according to the detection route adjustment parameters corresponding to the content bucket, so as to perform detection processing on the original video to be detected based on the adjusted dedicated detection route.

13. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes the computer execution instructions stored in the memory to perform the video detection method as described in any one of claims 1-12.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to perform the video detection method as described in any one of claims 1-12.

15. A computer program product, characterized in that, It includes a computer program, which, when executed by a processor, performs the video detection method as described in any one of claims 1-12.

Images