Quality gating methods, electronic devices, and storage media for marketing videos

By identifying the type of shot and obtaining differentiated quality gating configurations, detailed anomaly signatures are generated, solving the problem of not being able to distinguish anomalies in existing technologies. This enables refined quality control of marketing videos and improves the accuracy and efficiency of evaluation.

CN123123609APending Publication Date: 2026-11-17SHENZHEN DONGXU TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611304646.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing video quality control methods cannot distinguish between different types of anomalies, leading to false releases or increased detection costs. They also cannot perform precise repair actions for different shot types, affecting the video quality of marketing videos.

Method used

By identifying the shot type of the shot segment to be detected in the marketing video, the corresponding quality gating configuration is obtained from the preset configuration library, and differentiated quality detection is performed to generate anomaly signatures containing the anomaly type and degree, adapting to anomaly evaluation scenarios of different shot types.

Benefits of technology

It improved the accuracy and efficiency of quality assessment, reduced unnecessary testing costs, provided precise and targeted remediation guidance, realized the transformation from extensive management to refined management, and improved the quality control level of marketing videos.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123609A_ABST
    Figure CN123123609A_ABST
Patent Text Reader

Abstract

This application relates to the field of video quality control technology, and discloses a quality gating method, electronic device, and storage medium for marketing videos. The method includes: determining the shot segment to be detected in the marketing video and the shot type of the shot segment; obtaining a quality gating configuration corresponding to the shot type from a preset configuration library; performing quality detection on the shot segment according to the quality gating configuration to obtain a detection result; and generating an anomaly signature for the shot segment based on the detection result, the anomaly signature containing anomaly type information and anomaly degree information of the anomalies present in the shot segment. The embodiments of this application can distinguish different types of anomalies and adapt to anomaly evaluation scenarios for different shot types.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of video quality control technology, and in particular to a quality gating method, electronic device and storage medium for marketing videos. Background Technology

[0002] Existing video quality control methods use a single quality assessment dimension to calculate an overall quality score and employ a uniform quality threshold standard. This score cannot distinguish between different types of anomalies, and the uniform quality threshold standard cannot adapt to anomalies of different shot types, leading to false releases or increased detection costs. This results in an inability to perform precise remediation actions for different types of anomalies, affecting the video quality of marketing videos. Summary of the Invention

[0003] The purpose of this application is to provide a quality gating method, electronic device, and storage medium for marketing videos, which can distinguish different types of anomalies and adapt to anomaly evaluation scenarios for different types of shots.

[0004] This application provides a quality gating method for marketing videos, including: Identify the shot segments to be detected in the marketing video and the shot types of those segments; Obtain the quality gating configuration corresponding to the lens type from the preset configuration library; the quality gating configuration includes several quality dimensions used when evaluating the quality of the lens segment, several detection items corresponding to the quality dimensions, and several thresholds corresponding to the detection items. In the quality gating configurations corresponding to different lens types, at least one of the detection items and the thresholds is different. Based on the quality gating configuration, the quality of the shot segment is inspected to obtain the inspection result; Based on the detection results, an anomaly signature is generated for the shot segment; the anomaly signature contains anomaly type information and anomaly degree information of the anomalies present in the shot segment.

[0005] In some embodiments, performing quality inspection on the shot segment according to the quality gating configuration includes: Determine the detection computation cost of each detection item in the quality gating configuration; According to the order of detection computation cost from low to high, the quality detection corresponding to the detection item is performed on the lens segment to obtain the corresponding single detection result. When the single detection result of the detection item indicates that the lens segment does not meet the gate control, the execution of the detection item located after the detection item and dependent on the single detection result of the detection item is terminated. The detection results are generated based on the individual detection results.

[0006] In some embodiments, the threshold includes a hard threshold and a soft threshold. When the individual detection result of the detection item indicates that the shot segment does not conform to the hard threshold, the execution of the detection item located after the detection item and dependent on the individual detection result of the detection item is terminated. Otherwise, the individual detection result is generated according to the degree of deviation between the shot segment and the soft threshold.

[0007] In some embodiments, the shot types include product close-up shots, character voiceover shots, and transition shots. In the quality gating configuration corresponding to the product close-up shot, the thresholds corresponding to product geometry detection, brand logo detection, and text detection are set as the hard thresholds. In the quality gating configuration corresponding to the character voiceover shot, identity detection and lip-sync detection are set as the detection items. In the quality gating configuration corresponding to the transition shot, the threshold for semantic relevance detection is lowered and the threshold for motion continuity detection is raised.

[0008] In some embodiments, generating an abnormal signature for the shot segment based on the detection result includes: Each abnormal item and its abnormal duration are identified from the detection results. The abnormal durations of similar abnormal items that are adjacent or overlapping in time are merged to obtain the abnormal merged duration. Based on the confidence level of the detectors, the candidate anomaly type information output by each detector is weighted and voted on, and the anomaly type information corresponding to the anomaly merging period is determined based on the voting results. The degree of abnormality is determined based on the deviation between the detection score of the detection item corresponding to the abnormal merging period and the threshold, the proportion of the abnormal merging period to the total duration of the shot segment, and the proportion of the screen area occupied by the abnormal item in the shot segment. The anomaly signature is generated based on the anomaly type information and the anomaly severity information.

[0009] In some embodiments, the anomaly signature further includes responsible node information, which is used to indicate the responsible node in the video generation chain that caused the anomaly. The quality gating method for marketing videos further includes: Based on the information of the responsible nodes and the duration of the anomalies, combined with the dependency relationship between the shot clips and the auxiliary media data, the minimum set of points that need to be updated is determined as the rework closure. In the isolated storage area, according to the rework strategy corresponding to the anomaly type information, the nodes in the rework closure are locally regenerated to obtain candidate video segments; The candidate video segments are re-verified for quality, and boundary consistency verification is performed on the splicing boundary between the candidate video segments and the retained segments in the shot segments. When both the quality re-verification and the boundary consistency verification pass, the node corresponding to the rework closure is replaced with the node corresponding to the candidate video segment using atomic operations.

[0010] In some embodiments, prior to the local regeneration of nodes in the rework closure, the method further includes: Based on the anomaly type information, a set of candidate actions is obtained from a preset anomaly-action mapping table; Based on the current model performance information, remaining retry budget, and historical failure signatures, the candidate actions in the candidate action set are sorted to obtain the sorting result; Based on the sorting results, the rework strategy is determined; When the historical failure signature and the current abnormal signature contain the same abnormal type information, and the number of consecutive repair failures reaches a preset threshold, the model configuration and parameter combination corresponding to the historical failure signature are excluded from the candidate action set.

[0011] In some embodiments, replacing the node corresponding to the rework closure with the node corresponding to the candidate video segment using atomic operations includes: Lock the rework interval corresponding to the rework closure, and verify whether the current state of the rework interval is consistent with the previous state when the rework closure was generated. If they are consistent, switch the original node reference in the rework interval to the reference of the candidate video segment and its associated media data.

[0012] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described quality gating method for marketing videos.

[0013] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described quality gating method for marketing videos.

[0014] The beneficial effects of this application are as follows: By identifying the shot type of the footage segment to be inspected in the marketing video and obtaining the corresponding quality gating configuration from a preset configuration library, adaptive quality thresholds are achieved. This avoids applying a uniform and unreasonable standard to all shot types, significantly improving the accuracy, relevance, and efficiency of quality assessment and reducing unnecessary inspection costs. Furthermore, based on the quality inspection results, anomaly signatures containing anomaly type and severity information are generated. This overcomes the limitations of single-dimensional anomalies in quality assessment, clearly distinguishing different types of anomalies and providing precise guidance for subsequent targeted repair or optimization, avoiding blind repair or omission of key anomalies. Therefore, by introducing shot type recognition and adaptive quality gating configuration based on shot type, and generating detailed anomaly signatures, the quality control of marketing videos shifts from extensive to intensive management, improving the accuracy, relevance, and efficiency of video quality assessment, and providing a solid technical foundation for the automated production and optimization of marketing videos. Attached Figure Description

[0015] Figure 1 This is a flowchart of the quality gating method for marketing videos provided in the embodiments of this application.

[0016] Figure 2 This is a flowchart of a method for quality inspection of shot segments provided in an embodiment of this application.

[0017] Figure 3 This is a flowchart of a method for generating abnormal signatures for shot clips provided in an embodiment of this application.

[0018] Figure 4 This is a flowchart of a method provided in an embodiment of this application for generating an abnormal signature of a shot clip.

[0019] Figure 5 This is a flowchart of a method provided in this application embodiment for the process of locally regenerating nodes in a repair closure.

[0020] Figure 6 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and drawings are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the application. Furthermore, the information, data, and signals involved in the embodiments of this application are all authorized by the relevant parties or have been fully authorized by the parties, and the collection, use, and processing of related data comply with relevant laws, regulations, and standards.

[0024] In existing generative video platforms, video quality control processes typically calculate only a single overall quality score. This score fails to decouple specific anomaly types such as product deformation, subject drift, logo errors, garbled text, abrupt motion changes, beginning / end breaks, and audio-visual misalignment into independent evaluation metrics. The singularity of quality assessment dimensions makes it impossible to identify specific anomaly categories, thus failing to trigger targeted remedial actions. Furthermore, while quality thresholds are set using a uniform standard, different shot types exhibit fundamentally different sensitivities to quality metrics. For example, close-up shots of products are sensitive to geometric deformation but have a high tolerance for abrupt motion changes; lip-syncing in spoken-word shots requires strict precision; and fast transition shots are more sensitive to motion continuity. Consequently, the same threshold may be too lenient in product close-ups, allowing actual anomalies to pass, while being too strict in transition shots, misjudging them and causing unnecessary high-cost detection or missed anomalies. For instance, in the generation of e-commerce marketing videos, when processing close-up shots of mobile phone products, high-precision detection of product geometry and brand logos is required. However, current technology only outputs a total quality score. If this score is higher than a global threshold, the video is allowed to proceed, but it may overlook garbled text or slight logo distortion on the screen. Conversely, in fast transition shots, sudden motion changes are misjudged as abnormal, causing the full-frame quality inspection process to be incorrectly initiated, increasing processing latency. Specifically, when generating narration shots, minor misalignments between lip movements and speech may be oversensitively judged as audio-visual misalignment anomalies, triggering redundant repair mechanisms, while the actual tolerance for lip-sync should be higher than for other shot types.

[0025] If these issues are not addressed, video quality anomalies cannot be accurately located and categorized, resulting in a lack of targeted remediation mechanisms. This could lead to frequent and unnecessary full-scale checks, consuming additional computing resources. Furthermore, critical anomalies such as product deformation or garbled text may be missed, resulting in low-quality videos that negatively impact user experience and brand reputation. In the long run, this will reduce the reliability and commercial value of generative video platforms, hindering the effective deployment of the technology in the marketing field.

[0026] Based on this, embodiments of this application provide a quality gating method, electronic device, and storage medium for marketing videos. By configuring differentiated quality gating parameters for different shot types, it achieves refined classification of anomalies and generation of anomaly signatures, distinguishes different types of anomalies, adapts to anomaly evaluation scenarios for different shot types, and improves the accuracy of marketing video quality assessment.

[0027] See Figure 1 In one embodiment, a quality gating method for marketing videos is provided, including but not limited to steps S101 to S104.

[0028] Step S101: Determine the shot segments to be detected in the marketing video and the shot types of the shot segments.

[0029] A shot clip refers to a continuous sequence of images with specific content or presentation extracted from a complete video. It is the basic unit for quality inspection and evaluation.

[0030] Shot type is a classification system that categorizes shot clips according to their content, shooting techniques, or function. For example, they can be divided into close-up shots of products, voice-over shots, transition shots, etc. Different shot types have different requirements and focuses regarding video quality.

[0031] Identifying the footage segment to be detected can be achieved in several ways. For example, algorithms based on video content analysis can be used to automatically identify scene transition points or key event points in the video, thereby automatically segmenting the video into multiple independent footage segments. To identify the type of the footage segment, machine learning models can be used to analyze the visual features of the footage segment (such as composition, motion trajectory, subject recognition, etc.) to automatically classify the type of the footage segment.

[0032] Step S102: Obtain the quality gating configuration corresponding to the lens type from the preset configuration library.

[0033] A quality gating configuration is a set of quality assessment rules defined for a specific lens type. A quality gating configuration includes several quality dimensions used to assess the quality of a shot, several detection items corresponding to those quality dimensions, and several thresholds corresponding to those detection items. At least one of the detection items and thresholds differs in the quality gating configurations corresponding to different lens types.

[0034] The quality dimension is an indicator used to classify video quality macroscopically.

[0035] In some embodiments, the quality dimensions include file integrity, subject consistency, brand logo compliance, embedded text readability, motion smoothness, temporal continuity, audio-visual synchronization, and visual aesthetics. The file integrity dimension addresses issues such as video file corruption, encoding errors, or frame loss. The subject consistency dimension assesses the stability of key objects (e.g., products, people) within the video frame, checking for drift or distortion. The brand logo compliance dimension focuses on verifying the correctness, clarity, and placement of the brand logo, and identifying any potential infringement risks. The embedded text readability dimension evaluates the clarity and legibility of text within the video, checking for garbled text or obstructions. The motion smoothness dimension focuses on the smoothness of movement within the video frame, checking for stuttering or jitter. The temporal continuity dimension assesses the continuity of time logic between video segments or within the video, checking for frame skips or repetitions. The audio-visual synchronization dimension checks whether the audio and video are synchronized, checking for delays or misalignments. The visual aesthetics dimension evaluates the overall aesthetic appeal of the video from an artistic and visually attractive perspective, considering factors such as color, composition, and lighting.

[0036] Inspection items are more specific and quantifiable indicators within the quality dimension. For example, under the brand identity compliance dimension, inspection items may include logo clarity inspection or logo position inspection.

[0037] A threshold is a standard value used to determine whether a test item is qualified. When the test result exceeds or falls below the threshold, the test item is considered abnormal.

[0038] The pre-defined configuration library can be a structured database storing pre-defined sets of quality assessment rules for different shot types (such as product close-ups, voice-overs, and transitions). For example, when a product close-up is identified, a quality gating configuration specifically for such close-ups is retrieved from the library. This configuration might set a high threshold for the logo clarity detection item under the brand logo compliance dimension to ensure the product logo is clearly visible. Similarly, when a voice-over is identified, its corresponding quality gating configuration is obtained. This configuration might set a strict threshold for the lip-sync detection item under the audio-visual synchronization dimension. The quality gating configurations can be stored in XML, JSON, or other structured data formats, explicitly defining each quality dimension, its included detection items, and the threshold value for each detection item.

[0039] Step S103: Perform quality inspection on the shot clips according to the quality gating configuration to obtain the inspection results.

[0040] The execution of quality inspection can involve a series of independent inspection modules or algorithms. For example, for the file integrity dimension, a file verification module can be invoked to check the integrity of the video file; for the brand logo compliance dimension, image recognition algorithms can be used to detect the presence, position, and clarity of the logo; for the audio-visual synchronization dimension, audio and video analysis technologies can be used to compare the timestamps or feature matching degrees of the sound and image. Each inspection module evaluates the shot segment according to its corresponding inspection items and thresholds, and outputs corresponding individual inspection results. These individual inspection results can be a Boolean value (pass / fail), a score, or a detailed report. Finally, all individual inspection results are aggregated to form the inspection result for that shot segment.

[0041] Step S104: Based on the detection results, generate an anomaly signature for the shot segment.

[0042] Anomaly signatures are structured information describing quality issues present in a shot. Anomaly signatures contain information about the type of anomaly (e.g., unclear logo, audio-visual desynchronization) and the severity of the anomaly (e.g., minor, moderate, severe).

[0043] Generating anomaly signatures for video clips involves analyzing the aforementioned detection results to identify all detection items that failed to meet the threshold. For each anomaly, its corresponding anomaly type information is extracted, such as an unclear logo, out-of-sync lip movements, or shaky footage. Simultaneously, based on the deviation of the detection score from the threshold, the severity of the anomaly is determined, such as minor, moderate, or severe. This anomaly type and severity information are integrated into a structured data object, forming the anomaly signature for the video clip. This anomaly signature can serve as input for subsequent video editing, repair, or review processes, providing a clear basis for targeted solutions to video quality issues.

[0044] In summary, the quality gating method for marketing videos provided in this application identifies the shot type of the shot segment to be detected in the marketing video and obtains the corresponding quality gating configuration from a preset configuration library. This achieves adaptive quality thresholding, avoiding the application of a uniform and unreasonable standard to all shot types, significantly improving the accuracy, relevance, and efficiency of quality assessment, and reducing unnecessary detection costs. Furthermore, based on the quality detection results, an anomaly signature containing anomaly type and degree information is generated. This overcomes the limitations of a single dimension in quality assessment, clearly distinguishing different types of anomalies and providing precise guidance for subsequent targeted repair or optimization, avoiding blind repair or omission of key anomalies. Therefore, by introducing shot type recognition and adaptive quality gating configuration based on shot type, and generating detailed anomaly signatures, the quality control of marketing videos shifts from extensive to intensive management, improving the accuracy, relevance, and efficiency of video quality assessment, and providing a solid technical foundation for the automated production and optimization of marketing videos.

[0045] See Figure 2 In one embodiment, the method for quality detection of a shot segment includes, but is not limited to, steps S201 to S203.

[0046] Step S201: Determine the detection computation cost of each detection item in the quality gating configuration.

[0047] The computational cost of detection can be determined in advance by performing performance analysis and benchmark testing on each detection algorithm and stored in a preset configuration library; or it can be dynamically evaluated based on the complexity of the detection items and the characteristics of the footage segment to be detected (such as resolution, duration, etc.).

[0048] Step S202: Perform quality checks on the lens segment according to the order of detection computation cost from low to high, and obtain the corresponding individual detection results. When the individual detection result of a detection item indicates that the lens segment does not meet the gate control, terminate the execution of the detection item that is located after the detection item and depends on the individual detection result of the detection item.

[0049] When performing quality checks on shot clips according to the specified check items, all check items to be executed and their corresponding computational costs can be sorted, for example using quicksort or heapsort algorithms; alternatively, check items can be placed in a priority queue and executed sequentially according to cost priority. After each check item is executed, an independent check result is generated. This result can be a Boolean value (pass / fail), a quantified score, or a detailed anomaly report, used to characterize the degree to which the shot clip conforms to that specific quality dimension.

[0050] When the result of a single inspection item clearly indicates that a shot segment fails to meet quality gating requirements, it is determined whether there are subsequent inspection items that depend on the result of the current inspection item, or whether there are inspection items that are meaningless to execute under the current failure condition. If so, the execution of these subsequent inspection items is immediately stopped.

[0051] Step S203: Generate the test results based on the individual test results.

[0052] The test results may include a list of all passed and failed tests, the corresponding abnormality type, the degree of abnormality, and other relevant information.

[0053] This application's solution improves the efficiency of marketing video quality gating by introducing an assessment of the computational cost of detection items and cost-based optimization of execution order. Specifically, when performing quality inspection on shot clips, the computational cost of each detection item included in the quality gating configuration is first quantified, thereby identifying which detection items are lightweight and which are resource-intensive. Based on this cost information, the detection items are intelligently reordered, prioritizing those with lower computational costs. When the result of a single detection item clearly indicates that a shot clip does not meet the preset quality gating requirements, the execution of subsequent detection items that depend on the result of the current detection item or are no longer meaningful in the current failure scenario is terminated. Finally, all executed individual detection results are integrated to generate a comprehensive detection result, providing an accurate basis for subsequent anomaly signature generation.

[0054] The following example illustrates this. Suppose a marketing video clip requires quality gating, and its quality gating configuration includes the following detection items and their estimated computational costs (in relative values): file integrity check (cost 1), screen resolution check (cost 2), audio track presence check (cost 3), brand logo recognition (cost 10), facial expression analysis (cost 50), and background semantic analysis (cost 100). First, the computational costs of these detection items are determined. Then, in order of increasing cost, the execution order of the detection items is determined as follows: file integrity check, screen resolution check, audio track presence check, brand logo recognition, facial expression analysis, and background semantic analysis. In the actual detection process, file integrity check is executed first, resulting in a "pass" result. Next, screen resolution check is executed, resulting in a "fail" result. Since brand logo recognition, facial expression analysis, and background semantic analysis depend on screen resolution check, these processes are terminated if the screen resolution does not meet the requirements. Finally, audio track presence check is executed, resulting in a "pass" result. Finally, based on the individual test results of the executed test items (file integrity, screen resolution, and audio track existence), the final test results are generated, which clearly point out the anomaly in "screen resolution detection".

[0055] In some embodiments, the threshold includes hard thresholds and soft thresholds. When the individual detection result of a detection item indicates that the shot segment does not meet the hard threshold, the execution of the detection item that follows the detection item and depends on the individual detection result of the detection item is terminated; otherwise, an individual detection result is generated based on the degree of deviation between the shot segment and the soft threshold.

[0056] A hard threshold refers to a condition in the quality inspection process where, if the inspection result does not meet the threshold requirements, the footage is considered to have a serious anomaly and cannot pass the quality control, requiring immediate termination of the subsequent inspection process. This threshold is typically applied to critical inspection items that have a decisive impact on video quality, such as file corruption, severe screen tearing, or missing brand logos. These anomalies are usually unacceptable and difficult to repair. It can be implemented through a preset Boolean value judgment, i.e., whether the inspection result fully meets a certain strict condition; if not, it is directly judged as unacceptable; or by setting an extremely low tolerance range, exceeding this range triggers a hard rejection judgment.

[0057] A soft threshold refers to a quality inspection process where, even if the test result does not fully meet the threshold requirements, the subsequent inspection process is not immediately terminated. Instead, a corresponding individual test result is generated based on the degree of deviation. This type of threshold is typically applied to test items that affect video quality but are not fatal, such as slight image jitter, minor text layout flaws, or slight deviations in color saturation. These anomalies may be adjustable in post-production or acceptable to a certain extent. A soft threshold can be implemented by setting a numerical range within which the test result is considered acceptable, and by scoring based on the deviation amount if the result exceeds the range but does not reach the hard threshold; or by defining a fuzzy logic judgment that assigns different confidence scores based on how close the test indicator is to the threshold.

[0058] This application's solution further optimizes the flexibility and granularity of the detection process in marketing video quality gating methods by introducing a distinction between hard and soft thresholds. When performing quality checks on a shot segment according to preset configurations, the threshold type corresponding to the current detection item is first identified. If the detection item is configured as a hard threshold, and its single-item detection result clearly indicates that the shot segment does not meet the hard threshold requirements—for example, detecting video file corruption or severe missing key brand logos—the execution of all subsequent detection items that depend on the result of this detection item will be immediately terminated. On the other hand, if the current detection item is configured as a soft threshold, the detection process will not be immediately terminated even if the detection result does not fully meet the threshold. Instead, it will generate a quantitative single-item detection result based on the degree of deviation between the actual performance of the shot segment and the soft threshold. For example, for color saturation detection in the aesthetic dimension of the image, if the saturation is slightly lower than the recommended value, its deviation degree is calculated and a corresponding score is generated, rather than simply judging it as unqualified. By combining hard and soft thresholds, the proposed solution maintains detection efficiency while greatly enhancing the comprehensiveness and flexibility of quality assessment. It can more intelligently handle quality problems of varying severity, providing more precise guidance for subsequent video optimization or rework.

[0059] The following example illustrates this. Suppose that when performing quality gating on a clip from a marketing video, the clip is identified as a "product close-up." According to the preset configuration library, the quality gating configuration for this clip type includes checks such as "File Integrity Detection," "Brand Identity Detection," and "Image Shake Detection." "File Integrity Detection" and "Brand Identity Detection" are set as hard thresholds, while "Image Shake Detection" is set as a soft threshold. During the detection process, "File Integrity Detection" is performed first. If the detection finds that the video file is corrupted, such as missing keyframes, it does not meet the hard threshold, and the execution of all subsequent checks will be immediately terminated, generating a corresponding abnormal signature indicating the serious problem of file corruption. Conversely, if "File Integrity Detection" passes, "Brand Identity Detection" continues. If the detection finds that the brand identity is blurry or misplaced, and this check is set as a hard threshold, the subsequent "Image Shake Detection" is immediately terminated, and an abnormal signature is generated indicating a brand identity compliance issue. If "Brand Identity Detection" also passes, "Image Shake Detection" continues. Assume the soft threshold for this detection item requires the image shake amplitude to be less than 2 pixels. If the detection result shows an image shake amplitude of 3 pixels, a single detection result is generated based on the deviation of 1 pixel, such as "slight image shake, low deviation". Finally, all the single detection results that were not terminated are combined to generate a complete detection result for the shot segment, and an anomaly signature containing information on the anomaly type and anomaly degree is generated based on these results.

[0060] In some embodiments, shot types include product close-ups, voiceover shots, and transition shots.

[0061] In the quality gating configuration for close-up product shots, the thresholds for product geometry detection, brand logo detection, and text detection are set as hard thresholds. For close-up product shots, the core objective is to clearly and accurately display the product's details, appearance, and material characteristics, aiming to highlight its visual appeal. Therefore, in the quality gating configuration for close-up product shots, this application sets the thresholds for product geometry detection, brand logo detection, and text detection as hard thresholds. Product geometry detection aims to detect the shape, proportion, integrity, and presence of geometric anomalies such as deformation or occlusion of the product in the image. Image recognition technology can be used to identify the product outline and compare it with a standard geometric model, or deep learning models can be used for segmentation and pose estimation. Brand logo detection aims to detect the presence of brand logos, trademarks, brand colors, and other identifying elements in the image, and to assess their clarity, compliance with regulations, and absence of tampering or blurring. Feature-matching-based image processing algorithms or object detection models can be used for identification and evaluation. Text detection aims to identify text content appearing in the frame and assess its readability, spelling accuracy, font and size compliance, and whether it contains sensitive information. This can be achieved through optical character recognition (OCR) technology to extract and analyze text, or by combining it with deep learning text detection models. Setting these thresholds as hard thresholds means that if any anomalies are found in the product's geometry, brand logo, or embedded text, the shot will be immediately deemed unqualified, and subsequent detection will be terminated. This ensures strict control over the most critical quality elements of close-up shots of products.

[0062] In the quality gating configuration for voice-over footage, identity detection and lip-sync detection are included as detection items. The main purpose of voice-over footage is to convey information through the speaker's explanation, typically used to communicate, build trust, or enhance rapport. In the quality gating configuration for voice-over footage, this application includes identity detection and lip-sync detection as the detection items. Identity detection aims to verify whether the identity of the person appearing in the voice-over footage meets preset requirements, such as whether they are an authorized spokesperson. This can be achieved by comparing the face in the frame with a preset database of persons, or by combining voiceprint recognition technology for verification. Lip-sync detection aims to verify whether the changes in the speaker's lip movements are synchronized with the accompanying audio content. This can be achieved by analyzing the correlation between the movement trajectory of key facial points and audio waveform features, or by using a deep learning model to determine the matching degree between lip movements and speech. The introduction of these two detection items allows the quality gating of voice-over footage to focus on the core elements of the speaker's expression, ensuring the compliance of the person appearing on screen and the synchronization of audio and video in the explanation.

[0063] In the quality gating configuration for transition shots, the threshold for semantic relevance detection is lowered while the threshold for motion continuity detection is increased. Transition shots primarily function as transitional shots connecting two different scenes or content segments, typically used to smoothly guide the viewer's eye and enhance the video's smoothness and rhythm. Semantic relevance detection aims to assess the semantic connection between the video segment's content and the marketing theme or preceding and following segments, ensuring content consistency and logic. This can be achieved by extracting semantic features using video content understanding techniques and comparing them, or by using multimodal learning models for comprehensive analysis. Motion continuity detection aims to assess the smoothness and coherence of object or camera movement in a video segment, avoiding issues such as stuttering, frame skipping, and shakiness that affect the viewing experience. This can be achieved by analyzing optical flow information or motion vectors, or by using deep learning models for motion feature extraction. Lowering the threshold for semantic relevance detection means that the strict requirements for semantic matching of transition shots can be appropriately relaxed, while raising the threshold for motion continuity detection emphasizes the smoothness and seamlessness of transition effects, ensuring a smooth overall viewing experience for the video.

[0064] Through the aforementioned customized quality gating configuration, the proposed solution can intelligently adjust the detection focus and stringency according to the characteristics of different shot types. This differentiated strategy enables the quality gating system to more accurately identify key anomalies in various shot types, avoiding unnecessary waste of detection resources and significantly improving the overall quality and production efficiency of marketing videos. For example, in a specific implementation, when a segment is identified as a close-up shot of a product, it loads the quality gating configuration for that close-up shot. If the product geometry detection finds that a corner of the product packaging box is cropped out of the frame, the single detection result of this item will indicate that it does not meet the hard threshold, terminating the subsequent detection of the close-up shot of the product and generating an anomaly signature indicating a product geometry anomaly. When a person's voiceover shot is identified, the corresponding configuration is loaded. If the identity detection finds that the person on screen is not an authorized spokesperson, or the lip-sync detection finds that the person's lip movements and speech are significantly out of sync, a corresponding anomaly signature is generated. When a transition shot is identified, the corresponding configuration is loaded, and motion continuity is evaluated with a higher standard. If obvious screen stuttering or severe shaking is detected during the transition, an anomaly signature is generated because the motion continuity does not meet the standard. This targeted optimization makes the quality gating process more efficient and effective, enabling the rapid identification and resolution of key quality issues in specific shot types, thereby comprehensively improving the production quality and user experience of marketing videos.

[0065] See Figure 3 In one embodiment, the method for generating anomaly signatures for shot segments includes, but is not limited to, steps S301 to S304.

[0066] Step S301: Identify each abnormal item and its abnormal duration from the detection results, and merge the abnormal durations of similar abnormal items that are adjacent or overlapping in time to obtain the abnormal merged duration.

[0067] By analyzing the detection results, specific anomalies that do not meet the quality gating standards within each frame or short time window of a video segment can be identified, and their start and end times can be recorded to form anomaly duration periods. These temporally adjacent or overlapping anomaly duration periods are then merged to obtain a more representative and complete merged anomaly duration period. For example, a time threshold can be set; if the interval between two similar anomaly duration periods is less than the threshold, they are merged; or overlapping similar anomaly duration periods can be directly merged.

[0068] Step S302: Based on the confidence level of the detector, the candidate anomaly type information output by each detector is weighted and voted, and the anomaly type information corresponding to the anomaly merging period is determined based on the voting results.

[0069] During quality inspection, an anomaly merging period may be detected by multiple different detectors, or different detectors may provide different candidate types for the same anomaly. Due to differences in their algorithm principles, training data, or applicable scenarios, the reliability (i.e., confidence level) of each detector's output may vary. By assigning a confidence weight to each detector, when multiple candidate anomaly types exist, the type information output by each detector is weighted with its confidence level. Then, a voting mechanism (e.g., selecting the type with the highest weighted score) is used to determine the final anomaly type information for that anomaly merging period.

[0070] Step S303: Determine the degree of abnormality based on the deviation between the detection score of the detection item corresponding to the abnormal merging period and the threshold, the proportion of the abnormal merging period to the total duration of the shot segment, and the proportion of the abnormal item to the screen area of ​​the shot segment.

[0071] The assessment of anomaly severity considers three key factors. First, the deviation between the detection score and the threshold directly reflects the severity of the anomaly; the greater the deviation, the more severe the anomaly. Second, the proportion of the period during which the anomaly is merged into the total duration of the shot reflects the persistence of the anomaly's impact; the longer the duration, the greater the impact. Finally, the proportion of the shot's frame area occupied by the anomaly reflects its visual salience or spatial impact; the larger the proportion, the more pronounced the visual impact. By comprehensively calculating information from these three dimensions (e.g., through weighted summation, hierarchical mapping, or machine learning models), a more comprehensive and accurate anomaly severity assessment can be obtained.

[0072] Step S304: Generate an exception signature based on the exception type information and exception severity information.

[0073] After obtaining precise anomaly merging time periods, anomaly type information determined through weighted voting, and multi-dimensional quantified anomaly severity information, this key information is integrated to form a structured anomaly signature. This anomaly signature includes not only the nature (type) of the anomaly but also its severity (degree), providing a clear and actionable basis for subsequent anomaly reporting, problem tracing, remediation strategy formulation, and video quality management. For example, the anomaly signature can be a JSON object containing "type" and "severity" fields, or a string with specific encoding rules.

[0074] In some embodiments, the anomaly signature includes: anomaly type, signature version, lens identifier, normalized anomaly code, start and end time, object region or trajectory, detection value, threshold value, deviation, detector and model version, gating configuration version, evidence digest, responsibility node, anomaly degree, and confidence level.

[0075] This application's solution optimizes the quality gating method for marketing videos. After obtaining preliminary detection results, these results are first refined. Specifically, each specific anomaly and its duration on the timeline are identified from the detection results. To avoid redundant processing of fragmented or repetitive anomalies, the durations of similar anomalies that are temporally adjacent or overlapping are merged to obtain a unified anomaly merging period. Subsequently, for this anomaly merging period, candidate anomaly type information from multiple detectors is comprehensively utilized. Considering that different detectors may have different reliability, a detector confidence level is introduced, and the candidate anomaly type information output by each detector is weighted and voted on based on this. On this basis, to comprehensively assess the actual impact of the anomaly, the deviation between the detection score of the detection item corresponding to the anomaly merging period and a preset threshold is calculated to reflect the inherent severity of the anomaly. The proportion of the anomaly merging period to the total duration of the entire shot segment is also considered to assess the persistence of the anomaly. The proportion of the frame area occupied by the anomaly in the shot segment is also taken into account to measure the visual salience of the anomaly. By integrating information from these three dimensions, a more refined and objective anomaly severity information can be generated. Finally, the anomaly type and severity information obtained through the above refined processing are integrated to generate the final anomaly signature. This anomaly signature not only indicates the nature of the anomaly but also quantifies its severity, providing a clear and actionable basis for subsequent automated remediation, manual intervention, or quality decisions.

[0076] The following example illustrates this. Suppose that during quality inspection of a person's spoken-word footage, multiple anomalies are identified. For instance, from time point T1 to T2, the face recognition detector reports "blurred face," from time point T1.5 to T2.5, the lip-sync detector reports "abnormal lip movement," and from time point T2.3 to T3, the background detector reports "background jitter." First, these anomalies and their respective durations are identified. Next, for temporally adjacent or overlapping anomalies of the same type, such as "blurred face" and "abnormal lip movement," their durations are merged to obtain a longer merged anomaly period, for example, from T1 to T2.5. When determining the anomaly type, assume the confidence level of the face recognition detector is 0.8 and the confidence level of the lip-sync detector is 0.7. When determining the anomaly type for the merged anomaly duration, a weighted vote is performed based on the confidence levels of the two detectors. If the weighted voting result leans towards "blurred face," then the anomaly type information for that merged anomaly period is determined to be "blurred face." Multiple factors are considered when determining the degree of anomaly. For example, if the detection score for "blurred face" deviates from the threshold by 30%, the merged anomaly period (T1-T2.5) accounts for 25% of the total duration of the entire shot, and the blurred face area occupies 15% of the screen area, these values ​​are input into a pre-defined evaluation model, such as a rule-based scoring system or a simple weighted average model, to calculate a specific degree of anomaly information, such as "moderate blur." Finally, the anomaly type information for "blurred face" and the anomaly degree information for "moderate blur" are integrated to generate an anomaly signature for the shot. This anomaly signature can be stored as structured data, such as {"anomaly_type": "blurred face", "severity": "moderate blur", "duration_start": "T1", "duration_end": "T2.5"}, for subsequent automated processing or manual review.

[0077] In some embodiments, the anomaly signature also includes responsible node information, which indicates the responsible node in the video generation chain that caused the anomaly. See also Figure 4 In one embodiment, the method following the generation of the abnormal signature of the shot segment includes, but is not limited to, steps S401 to S404.

[0078] Step S401: Based on the information of the responsible node and the duration of the anomaly, and combined with the dependency relationship between the shot clip and the auxiliary media data, determine the minimum set of points that needs to be updated, as the rework closure.

[0079] Responsibility node information identifies the specific link or component in the video generation chain that caused the video anomaly. Its purpose is to accurately trace the source of the anomaly, providing target location for subsequent automated repair. This information can be represented as a unique identifier for a specific service module in the video production process, an ID for a specific material in a content management system, or a marker for a certain processing stage in the video rendering pipeline. For example, when a brand logo compliance issue is detected, the responsibility node information might point to the module responsible for reviewing or embedding brand materials; when a motion smoothness issue is detected, it might point to the module responsible for video stabilization or transition effect generation.

[0080] Determining the minimum set of points that need updating can be achieved by analyzing the video's editing history, source code relationships, or rendering graph to construct a dependency graph. Then, the smallest set of nodes directly or indirectly related to the anomaly can be identified from this graph. Alternatively, a version control system or asset management system can be used to track the modification history and source code relationships of video clips and their associated materials (such as audio, subtitles, and effects layers), thereby identifying the smallest repairable unit affected by the anomaly as the rework closure.

[0081] Step S402: In the isolated storage area, according to the rework strategy corresponding to the anomaly type information, the nodes in the rework closure are locally regenerated to obtain candidate video segments.

[0082] An isolated storage area is an environment independent of the main production environment used for partial regeneration and testing of video clips. Its purpose is to ensure that the repair process does not affect the running production system and allows for multiple attempts and verifications in a secure environment. An isolated storage area can be a separate sandbox environment, a virtualized container cluster, or a dedicated group of test servers.

[0083] A rework strategy refers to pre-defined repair rules or operational procedures for different types of anomalies. Its purpose is to guide the local regeneration process, ensuring the effectiveness and specificity of the repair. For example, for the "blurred image" anomaly, a rework strategy might include trying different deblurring algorithms, adjusting rendering parameters, or replacing source material with higher quality material; for the "audio-visual desynchronization" anomaly, a rework strategy might involve adjusting the start time offset of the audio or video.

[0084] Partial regeneration refers to reprocessing or generating only the nodes within the rework closure in an isolated storage area, according to the rework strategy. Its purpose is to efficiently repair video anomalies and avoid unnecessary re-rendering or re-encoding of the entire video. For example, if the rework closure is a specific effects layer, only that effects layer is re-rendered; if it's a subtitle error, only the subtitle track is regenerated.

[0085] Step S403: Perform quality re-verification on the candidate video segments and perform boundary consistency verification on the splicing boundary between the candidate video segments and the retained segments in the shot segments.

[0086] Quality re-verification refers to performing quality checks again on the candidate video segments obtained from the partial regeneration. Its purpose is to verify whether the repaired video segments meet quality gating requirements and ensure that anomalies have been effectively eliminated. This re-verification process can either repeat the initial quality gating detection procedure or employ more refined detection algorithms tailored to specific anomaly types.

[0087] Boundary consistency verification refers to checking the splicing points between candidate video clips and the retained segments in the original footage. Its purpose is to ensure that the repaired clip seamlessly integrates with the rest of the video, avoiding visual or auditory abruptness. For example, it can check whether the color, brightness, and contrast of the image at the splicing point transition smoothly, or whether the volume and pitch of the audio are continuous. Another approach is to analyze the motion vectors or optical flow information at the splicing point to determine if there are any unnatural jumps.

[0088] Step S404: When both the quality re-verification and boundary consistency verification pass, replace the node corresponding to the rework closure with the node corresponding to the candidate video segment using atomic operations.

[0089] An atomic operation is the process of replacing the node corresponding to a rework closure with the node corresponding to a candidate video segment. This process either succeeds completely or fails completely; there are no intermediate states. For example, it can be implemented through database transaction mechanisms, file system snapshots, or commit operations in version control systems.

[0090] This application introduces an intelligent anomaly repair mechanism. When quality inspection identifies anomalies in a shot and generates an anomaly signature, this signature includes not only anomaly type and severity information but also information about the responsible node. This responsible node information precisely identifies the specific link in the video generation chain that caused the anomaly. Based on this, a minimal set of points requiring updating, i.e., a rework closure, can be determined by combining the duration of the anomaly and the dependency between the shot and its associated media data. To ensure the stability and security of the repair process, in an isolated storage area, local regeneration operations are performed on the nodes in the rework closure according to a preset rework strategy matching the anomaly type information, resulting in a candidate video shot. To ensure the repair effect, the generated candidate video shot undergoes rigorous quality re-verification, and boundary consistency verification is performed on its splicing boundary with the retained portion of the original shot to ensure that the repaired shot meets the requirements in terms of quality and seamlessness. Only when both the quality re-verification and boundary consistency verification are successful will the original node corresponding to the rework closure be replaced with the node corresponding to the newly generated candidate video shot in an atomic operation, thereby ensuring the integrity and consistency of data updates. If local regeneration fails to resolve the issue after reaching the preset maximum number of retries, further automated attempts will be terminated, and a detailed anomaly report and alternative degradation solutions will be generated for manual intervention or other remedial measures.

[0091] The following is a concrete example. When the quality gating method of a marketing video detects an anomaly of "blurred brand logo" in a shot, the generated anomaly signature includes information about the anomaly type, the severity of the anomaly, and the responsibility node information indicating that the brand logo was embedded by the "Brand Material Rendering Service." Based on this responsibility node information and the duration of the anomaly, combined with the video's rendering graph, a specific layer or rendering task containing only the blurred brand logo is identified as a rework closure. Subsequently, in an independent containerized sandbox environment, according to the rework strategy corresponding to "blurred brand logo" (e.g., attempting to replace it with a higher resolution brand material, or adjusting rendering parameters to improve clarity), the rework closure is partially regenerated to obtain a candidate video shot containing a clear brand logo. Next, the brand logo clarity detection and compliance detection are re-performed on the candidate video shot, and the smoothness of the transition between it and other background elements in the original shot is checked to ensure there are no abrupt splicing marks. If all verifications pass, a database transaction operation atomically switches the reference in the video asset management system that points to the old blurred brand logo layer to a reference that points to the newly generated clear brand logo layer. If the brand logo still fails to meet the clarity requirements after multiple attempts (e.g., three times), a detailed anomaly report will be generated, indicating that "the quality of the brand material source file is insufficient" or "the rendering parameters cannot meet the requirements," and providing alternative degradation solutions, such as suggesting the use of alternative brand display methods or manual intervention.

[0092] See Figure 5 In one embodiment, the method prior to locally regenerating the nodes in the repair closure includes, but is not limited to, steps S501 to S504.

[0093] Step S501: Based on the exception type information, obtain a set of candidate actions from the preset exception-action mapping table.

[0094] An anomaly-action mapping table is a pre-built knowledge base or database that stores the correspondence between various known anomaly types and the possible remedial actions (i.e., remediation strategies or solutions) that can be taken for these anomaly types. For example, for the anomaly type "blurred image," the candidate action set may include "apply super-resolution algorithm," "re-render the blurred area," and "replace with clear material," etc. This mapping table can be built and updated based on expert experience, historical remediation data, or machine learning models.

[0095] Step S502: Based on the current model performance information, remaining retry budget, and historical failure signatures, sort the candidate actions in the candidate action set to obtain the sorting result.

[0096] Model performance information refers to the real-time running status, processing capacity, success rate, and resource consumption of various models (such as image processing models, AI generation models, rendering engines, etc.) currently used for video processing or restoration.

[0097] The remaining retry budget refers to the number of unused attempts out of the maximum allowed number of repair attempts for the currently detected anomaly.

[0098] Historical failure signatures refer to the exception signatures recorded for past repair failures, which include the exception type that caused the failure, the repair action attempted at the time, the reason for the failure, and the relevant model configuration and parameter combination.

[0099] By comprehensively considering this information, the various repair actions in the candidate action set can be prioritized and ranked. For example, repair actions that have good current model performance, high success rate, low resource consumption, and have never caused failure in the historical record can be selected first to obtain the ranking result.

[0100] Step S503: Determine the rework strategy based on the sorting results.

[0101] A rework strategy refers to determining a specific remediation plan for fixing a current anomaly from a sorted set of candidate actions, based on pre-defined strategy selection rules (e.g., selecting the highest-ranked action or a combination of multiple actions that meet specific conditions). This strategy can be a single remediation action or a remediation process involving multiple steps.

[0102] Step S504: When the historical failure signature and the current abnormal signature contain the same abnormal type information, and the number of consecutive repair failures reaches a preset threshold, the model configuration and parameter combination corresponding to the historical failure signature are excluded from the candidate action set.

[0103] This application's solution optimizes the local regeneration process of abnormal segments in marketing videos by introducing an intelligent rework strategy selection mechanism. Specifically, after identifying the anomaly type information in the marketing video, it first retrieves all candidate actions that can potentially repair that anomaly type from a pre-defined anomaly-action mapping table. To avoid blind attempts and resource waste, these candidate actions are comprehensively evaluated and ranked based on the performance status of the current video processing model, the remaining retry opportunities for that anomaly, and historical failure data (historical failure signatures). For example, priority is given to repair actions with high success rates, low resource consumption, and good performance in similar scenarios. Based on this ranking result, the optimal rework strategy can be determined, thereby guiding subsequent local regeneration operations. Furthermore, to improve repair efficiency and avoid repeated ineffective attempts, if a specific repair action (and its corresponding model configuration and parameter combination) fails consecutively to a preset threshold when facing the same anomaly type, it is excluded from the candidate action set to ensure that this ineffective solution is not repeatedly tried in the future.

[0104] The following example illustrates this. Suppose a marketing video clip is detected as having a "blurred" anomaly. First, the pre-defined anomaly-action mapping table is queried to obtain a set of candidate actions for "blurred" footage, such as "Apply super-resolution algorithm A," "Apply super-resolution algorithm B," "Re-render the clip," and "Replace with a clear backup." Next, the current state is evaluated: for example, model performance information shows that super-resolution algorithm A has a high current server load but a historical success rate of 85%; super-resolution algorithm B has a lower load and a historical success rate of 70%. Meanwhile, the remaining retry budget for this "blurred" anomaly is 3 attempts. Furthermore, checking historical failure signatures reveals two previous attempts to fix the "blurred" issue, both using specific parameter combinations of super-resolution algorithm A, but both failed. Based on this information, the candidate actions are ranked: "Apply super-resolution algorithm B" might be ranked first because it currently has a low load and a certain success rate; second is "Re-render the clip" because it usually solves the blur problem but has a higher computational cost; "Apply super-resolution algorithm A" might be ranked lower because it has failed twice consecutively. If "applying super-resolution algorithm A" fails again, reaching a preset threshold for consecutive failures (e.g., 3 times), then that specific super-resolution algorithm A and its parameter combination will be excluded from the candidate action set, and will no longer be considered for using it to fix this type of "blurry image" problem. Finally, based on the ranking results, "applying super-resolution algorithm B" with the highest ranking is selected as the current rework strategy.

[0105] In some embodiments, replacing the node corresponding to the rework closure with the node corresponding to the candidate video segment using atomic operations includes: locking the rework interval corresponding to the rework closure, and verifying whether the current state of the rework interval is consistent with the previous state when the rework closure was generated. If consistent, the original node reference in the rework interval is switched to a reference to the candidate video segment and its associated media data.

[0106] In some embodiments, verifying whether the current state of the rework interval is consistent with the previous state when the rework closure was generated includes: verifying whether the lease of the rework closure is valid, verifying whether the fence token of the rework closure is equal to the current valid token, verifying whether the baseline version of the rework closure has not changed and whether the quality re-verification and boundary consistency verification have passed.

[0107] This application's solution ensures security and data consistency during atomic replacement operations by introducing locking and state verification mechanisms. Specifically, when an atomic operation is needed to replace the node corresponding to the rework closure with the node corresponding to the candidate video segment, the target rework interval is first locked. This locking operation effectively prevents other concurrent processes from modifying this interval, thus providing a stable operating environment for subsequent state verification and reference switching. After successful locking, the state of the current rework interval is further verified to ensure consistency with the previous state when the rework closure was generated. This verification step is crucial, as it can promptly detect and prevent data drift or inconsistency issues caused by external factors (such as modifications by other users, automatic system updates, etc.). Only when the state verification passes, confirming that the data in the rework interval has not been accidentally modified, is the reference switching operation performed, safely pointing the original node reference to the newly generated candidate video segment and its associated media data. This mechanism ensures the atomicity of the replacement operation not only in terms of its uninterruptibility but also in its strict guarantee of data integrity and timeliness. This avoids potential errors caused by concurrent operations or data expiration, making the quality gating method for marketing videos more robust and reliable when performing partial regeneration and replacement.

[0108] The following is a concrete example. Suppose a clip from a marketing video (e.g., a close-up of a product) is found to have an anomaly in its brand identity compliance dimension, generating a corresponding anomaly signature. Based on the responsible node information and the duration of the anomaly, combined with the dependencies between the clip and its associated media data, the minimum set of points requiring rework, i.e., the rework closure, is determined, and a candidate video clip is successfully regenerated in an isolated storage area. When preparing to replace the original node with this candidate video clip, the first step is to lock the rework interval corresponding to the rework closure. For example, in a database-based video asset management system, a row-level exclusive lock can be applied to the database record storing the clip's metadata (including file path, version number, etc.). Subsequently, the current version number of the rework interval is read and compared with the previous version number recorded when the rework closure was generated. If they match, it indicates that the clip has not been modified by other operations since the rework closure was generated, and a reference switch is performed. Specifically, it updates the file path field of the clip in the database to point to the storage location of the newly generated candidate video clip, and updates the references to its associated media data (such as corresponding audio files and subtitle files). If the version numbers are inconsistent, it means that the range has been modified. In this case, the replacement will not be performed, but a retry mechanism or a warning may be triggered to avoid overwriting valid data.

[0109] This application also provides an electronic device. (See appendix.) Figure 6 , Figure 6 This is a schematic diagram of the main structure of an electronic device according to an embodiment of this application. Figure 6 As shown, the electronic device in this embodiment mainly includes a processor 601 and a memory 602. The memory 602 can be configured to store a program for executing the methods of the above-described method embodiments, and the processor 601 can be configured to execute the program in the memory 602. This program includes, but is not limited to, a program for executing the methods of the above-described method embodiments. For ease of explanation, only the parts related to the embodiments of this application are shown. For specific technical details not disclosed, please refer to the method section of the embodiments of this application.

[0110] In some possible embodiments of this application, the electronic device may include multiple processors 601 and multiple memories 602. The program executing the methods of the above method embodiments may be divided into multiple subroutines, each of which may be loaded and run by a processor 601 to perform different steps of the methods of the above method embodiments. Specifically, each subroutine may be stored in a different memory 602, and each processor 601 may be configured to execute programs in one or more memories 602 to jointly implement the methods of the above method embodiments; that is, each processor 601 executes different steps of the methods of the above method embodiments to jointly implement the methods of the above method embodiments.

[0111] The aforementioned multiple processors 601 can be processors deployed on the same device. For example, the aforementioned electronic device can be a high-performance device composed of multiple processors, and the aforementioned multiple processors 601 can be processors configured on that high-performance device. Alternatively, the aforementioned multiple processors 601 can also be processors deployed on different devices. For example, the aforementioned electronic device can be a server cluster, and the aforementioned multiple processors 601 can be processors on different servers within the server cluster.

[0112] This application also provides a computer-readable storage medium. In one embodiment of the computer-readable storage medium according to this application, the computer-readable storage medium can be configured to store a program that performs the methods of the above-described method embodiments, the program of which can be loaded and run by a processor to implement the above-described methods. For ease of explanation, only the parts related to the embodiments of this application are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of this application. The computer-readable storage medium can be a memory formed by various electronic devices. Optionally, in the embodiments of this application, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0113] The marketing video quality gating method, electronic device, and storage medium provided in this application identify the shot type of the shot segment to be detected in the marketing video and obtain the corresponding quality gating configuration from a preset configuration library. This achieves adaptive quality thresholding, avoiding the application of a uniform and unreasonable standard to all shot types, significantly improving the accuracy, relevance, and efficiency of quality assessment, and reducing unnecessary detection costs. Furthermore, based on the quality detection results, an anomaly signature containing anomaly type and degree information is generated. This overcomes the limitations of a single dimension in quality assessment, clearly distinguishing different types of anomalies and providing precise guidance for subsequent targeted repair or optimization, avoiding blind repair or omission of key anomalies. Therefore, by introducing shot type recognition and adaptive quality gating configuration based on shot type, and generating detailed anomaly signatures, the quality control of marketing videos shifts from extensive to intensive management, improving the accuracy, relevance, and efficiency of video quality assessment, and providing a solid technical foundation for the automated production and optimization of marketing videos.

[0114] Exemplary embodiments of this disclosure have been specifically shown and described above. It should be understood that this disclosure is not limited to the detailed structures, arrangements, or implementations described herein; rather, this disclosure is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims.

Claims

1. A quality gating method for marketing videos, characterized in that, include: Identify the shot segments to be detected in the marketing video and the shot types of those segments; Obtain the quality gating configuration corresponding to the lens type from the preset configuration library; the quality gating configuration includes several quality dimensions used when evaluating the quality of the lens segment, several detection items corresponding to the quality dimensions, and several thresholds corresponding to the detection items. In the quality gating configurations corresponding to different lens types, at least one of the detection items and the thresholds is different. Based on the quality gating configuration, the quality of the shot segment is inspected to obtain the inspection result; Based on the detection results, an anomaly signature is generated for the shot segment; the anomaly signature contains anomaly type information and anomaly degree information of the anomalies present in the shot segment.

2. The quality gating method for marketing videos according to claim 1, characterized in that, The quality detection of the shot segment according to the quality gating configuration includes: Determine the detection computation cost of each detection item in the quality gating configuration; According to the order of detection computation cost from low to high, the quality detection corresponding to the detection item is performed on the lens segment to obtain the corresponding single detection result. When the single detection result of the detection item indicates that the lens segment does not meet the gate control, the execution of the detection item located after the detection item and dependent on the single detection result of the detection item is terminated. The detection results are generated based on the individual detection results.

3. The quality gating method for marketing videos according to claim 2, characterized in that, The threshold includes a hard threshold and a soft threshold. When the single detection result of the detection item indicates that the shot segment does not conform to the hard threshold, the execution of the detection item located after the detection item and dependent on the single detection result of the detection item is terminated. Otherwise, the single detection result is generated according to the degree of deviation between the shot segment and the soft threshold.

4. The quality gating method for marketing videos according to claim 3, characterized in that, The shot types include product close-up shots, voice-over shots, and transition shots. In the quality gating configuration corresponding to the product close-up shots, the thresholds corresponding to product geometry detection, brand logo detection, and text detection are set as hard thresholds. In the quality gating configuration corresponding to the voice-over shots, identity detection and lip-sync detection are set as the detection items. In the quality gating configuration corresponding to the transition shots, the threshold for semantic relevance detection is lowered and the threshold for motion continuity detection is raised.

5. The quality gating method for marketing videos according to claim 1, characterized in that, The step of generating an abnormal signature for the shot segment based on the detection result includes: Each abnormal item and its abnormal duration are identified from the detection results. The abnormal durations of similar abnormal items that are adjacent or overlapping in time are merged to obtain the abnormal merged duration. Based on the confidence level of the detectors, the candidate anomaly type information output by each detector is weighted and voted on, and the anomaly type information corresponding to the anomaly merging period is determined based on the voting results. The degree of abnormality is determined based on the deviation between the detection score of the detection item corresponding to the abnormal merging period and the threshold, the proportion of the abnormal merging period to the total duration of the shot segment, and the proportion of the screen area occupied by the abnormal item in the shot segment. The anomaly signature is generated based on the anomaly type information and the anomaly severity information.

6. The quality gating method for marketing videos according to claim 1, characterized in that, The anomaly signature also includes responsible node information, which is used to indicate the responsible node in the video generation chain that caused the anomaly. The marketing video quality gating method further includes: Based on the information of the responsible nodes and the duration of the anomalies, combined with the dependency relationship between the shot clips and the auxiliary media data, the minimum set of points that need to be updated is determined as the rework closure. In the isolated storage area, according to the rework strategy corresponding to the anomaly type information, the nodes in the rework closure are locally regenerated to obtain candidate video segments; The candidate video segments are re-verified for quality, and boundary consistency verification is performed on the splicing boundary between the candidate video segments and the retained segments in the shot segments. When both the quality re-verification and the boundary consistency verification pass, the node corresponding to the rework closure is replaced with the node corresponding to the candidate video segment using atomic operations.

7. The quality gating method for marketing videos according to claim 6, characterized in that, Before the local regeneration of the nodes in the rework closure, the method further includes: Based on the anomaly type information, a set of candidate actions is obtained from a preset anomaly-action mapping table; Based on the current model performance information, remaining retry budget, and historical failure signatures, the candidate actions in the candidate action set are sorted to obtain the sorting result; Based on the sorting results, the rework strategy is determined; When the historical failure signature and the current abnormal signature contain the same abnormal type information, and the number of consecutive repair failures reaches a preset threshold, the model configuration and parameter combination corresponding to the historical failure signature are excluded from the candidate action set.

8. The quality gating method for marketing videos according to claim 6, characterized in that, The step of replacing the node corresponding to the rework closure with the node corresponding to the candidate video segment using atomic operations includes: Lock the rework interval corresponding to the rework closure, and verify whether the current state of the rework interval is consistent with the previous state when the rework closure was generated. If they are consistent, switch the original node reference in the rework interval to the reference of the candidate video segment and its associated media data.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the quality gating method for marketing videos according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the quality gating method for marketing videos as described in any one of claims 1 to 8.