An artificial intelligence-based cross-platform video promotion system

Abstract

This invention belongs to the field of video promotion and mainly relates to an AI-based cross-platform video promotion system. This system uses material version records and rule base version records as deterministic benchmarks to support versioned management and reproducible execution when the same promotional campaign is launched on multiple platforms in parallel. It performs unpacking and parsing of original materials, extracts text, and solidifies content fingerprints; it structurally merges the review constraints of each platform and generates a rule base version; it generates release versions for each platform according to the rule base version and performs multimodal compliance review; after approval, it submits to the platform and solidifies the release receipt to form a release ledger; it reads back the actual presented copy on the platform and verifies its consistency with the released version, supporting random checks, dispute evidence collection, and audit archiving; it monitors changes in laws and regulations and updates the system rule base in real time; and it achieves unified versioned management of multi-version materials across platforms, consistency verification of actual presentation after release, and traceable retention of evidence materials.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of video promotion, and mainly relates to a cross-platform video promotion system based on artificial intelligence. BACKGROUND

[0002] In the medical health, financial planning, insurance and other strictly regulated industries, cross-platform video content promotion has become a crucial marketing and user education tool. However, this process faces dual regulatory challenges: it must comply with national and industry laws and regulations, and it must also adapt to the different content review standards of different publishing platforms (such as short videos, social media, and information flow platforms).

[0003] In Chinese application No. CN202511447816.3, a large-screen video content review method, device, and storage medium are disclosed, which belong to the field of Internet security technology. The method includes: obtaining a video to be reviewed, decoupling the video to be reviewed to obtain an audio stream and multiple video frames in the video to be reviewed, and converting the audio stream into audio text; performing a text compliance detection operation on the audio text to obtain a first text review result corresponding to the audio text; performing an image compliance detection operation on the video frames to obtain an image review result corresponding to the video frames; if any one of the first text review result and the image review result is non-compliant, determining that the video review result of the video to be reviewed is non-compliant, and replacing the video to be reviewed with a preset compliant video to enable the preset compliant video to be played on a large-screen display device, thereby solving the technical problem of low recognition accuracy of single-modal detection in complex scenarios and improving the accuracy and reliability of video content review.

[0004] The aforementioned patents have provided a relatively complete video review technology, but their technical focus remains on performing compliance checks on the audio text and video frames of a single video to be reviewed and providing a compliance result. Furthermore, their handling method is mainly limited to replacing the video with a pre-set compliant one when non-compliant is determined, to meet the display requirements of large-screen playback scenarios. However, in cross-platform video promotion scenarios in strictly regulated industries such as pharmaceuticals, healthcare, finance, and insurance, the same promotional activity typically needs to be launched in parallel on multiple platforms, generating multiple release versions tailored to the screen ratio, subtitle style, risk warning presentation position, and display duration requirements of each platform. Existing technologies struggle to provide unified version management for the multi-version generation process and the multi-platform release process. Simultaneously, the cross-platform release chain usually introduces platform... Processes such as transcoding and compression, cropping and adaptation, automatic watermarking, and subtitle rearrangement can lead to discrepancies between the actual published content and the submitted content. Existing technologies lack mechanisms for reviewing and verifying the published copy and for maintaining traceability. Furthermore, regulated industries require verifiable supporting documentation to demonstrate consistency between the reviewed and published content in scenarios involving random checks, complaints, and regulatory inspections. Existing technologies typically only output review conclusions or log records, making it difficult to create a structured evidence package. Moreover, when laws, regulations, or platform review standards are updated, existing published materials need rapid batch review and replacement. Existing technologies still rely on manual execution for rule version locking, existing material location, and closed-loop anomaly handling, making it difficult to mitigate risks quickly.

[0005] To address the aforementioned issues, this invention proposes an AI-based cross-platform video promotion system. This system uses material version records and rule base version records as two deterministic main lines, solidifying the processes of "multi-platform release version generation, pre-release multimodal compliance review, release receipt solidification, post-release copy verification, structured indexing of evidence packages, and batch review and handling of existing materials triggered by rule updates" into a reproducible workflow. This enables consistency verification and traceability of the versions released on each platform with the actual versions presented on the platform within the same promotional campaign, and supports rapid risk mitigation for existing released materials after rule changes. Summary of the Invention

[0006] This invention provides an AI-based cross-platform video promotion system, which aims to solve the problems of difficulty in unified management of multiple versions of materials in cross-platform video promotion in regulated industries, inconsistencies between the presented content and the approved content due to secondary processing by the platform after release, difficulty in tracing and proving the inconsistencies, and difficulty in quickly locating, reviewing and closing the loop on existing materials when regulations or platform specifications are updated.

[0007] To solve the above problems, the present invention employs the following technology:

[0008] An AI-based cross-platform video promotion system includes:

[0009] Registration and Locking Module: Records the version of promotional materials for the registration campaign; analyzes the restrictions of each platform to form a rule base for the entire platform system;

[0010] Generate review module: Generate release versions for each platform based on the material version; combine the rule base of the entire platform system and use multimodal compliance reasoning measures to review the release versions;

[0011] Release receipt module: For versions that have passed the review and are released, release orchestration measures are adopted to submit the released versions to various platforms; the platform return information is solidified into receipts, release receipt records are generated and written into the release ledger;

[0012] The evidence review module reads back the presented copy of the release receipt record; compares the released version with the presented copy, generates a consistency verification record and a list of deviation items, records relevant information, and forms an index record for the evidence material package;

[0013] Existing data handling module: The module employs update monitoring measures for the legal provisions library and platform specification library, and updates the system rule library; based on the new system rule library, it reviews the published presentation copies, and takes further action on the published videos based on the review results.

[0014] In a preferred embodiment, the registration and locking module specifically includes:

[0015] Register promotional materials, generate material version records and write them into the material version ledger. The content includes: using media unpacking measures to obtain detailed material information from the original video files, including video track files, audio track files and subtitle track files, and writing video resolution, video frame rate, video duration, video encoding format, audio sampling rate, number of audio channels, audio encoding format and subtitle format type into the material metadata field.

[0016] The audio text, keyframe sequence, and subtitle text sequence are extracted from the material details using a corresponding method, and the version identifier of the extraction method is written in; at the same time, text recognition is used on the keyframe sequence to obtain the on-screen text sequence;

[0017] The module uses a fixed-order concatenation method to generate a seed text string from three types of text: audio text, video text, and subtitle text. It then uses a summary calculation method to generate a content fingerprint field from the seed text string.

[0018] The acquired material metadata fields and the extracted information are written together into the material version record.

[0019] In a preferred embodiment, the registration locking module further includes:

[0020] The restrictions and requirements of each platform are analyzed and processed to generate a system rule base for the entire platform. The content includes: for each publishing platform, reading the platform constraint entries from the relevant acquisition source and recording the detailed configuration information of each platform constraint entry;

[0021] Perform structured mapping processing on the platform constraint entries of the records, and generate rule entry records according to a unified field structure;

[0022] The unified field structure includes constraint category fields and other constraint fields;

[0023] Among them, the constraint category field includes screen specification constraints, subtitle style constraints, risk warning presentation constraints, and content review constraints, the specific contents of which are as follows:

[0024] Image specification constraints: Write the image aspect ratio parameters, target resolution parameters, and safety margin parameters into the rule entry record;

[0025] Subtitle style constraints: Write the subtitle area position parameters, font size range parameters, line spacing range parameters, and occlusion restriction parameters into the rule entry record;

[0026] Risk warning presentation constraints: The risk warning text template identifier, display start and end time points, display duration range, and display area location are written into the rule entry record;

[0027] Content moderation constraints: Write the lists of sensitive words, prohibited expressions, and prohibited image element types into the rule entry record;

[0028] For parameters that need to be expressed in numerical range, the numerical range and the basis for determination are written into the rule entry record simultaneously, and the basis for determination is written into the source field in the form of a number, so as to ensure that the rule entry is verifiable and traceable;

[0029] The obtained rules are stored separately according to the platform index, forming a system rule library for the entire platform.

[0030] As a preferred implementation, the generation and review module specifically includes:

[0031] Retrieves relevant information from the material version history, and retrieves the corresponding screen specification constraints, subtitle style constraints, risk warning presentation constraints, and content review constraints from the system rule base according to the platform to be published, and processes the material content according to the constraints;

[0032] The video track file is cropped and scaled to obtain a target image that meets the target screen ratio and target resolution requirements. The cropping uses a center reference and the available rendering area for subtitles and risk warnings is limited by using area constraints on the safety margin parameters.

[0033] The subtitle text is re-formatted and rearranged to generate a subtitle layer. The re-formatting process includes numerically configuring the subtitle area position, font size and line spacing. The subtitle line width is calculated from the target resolution and the safety margin, and the value range of font size and line spacing is limited to the range registered in the rule base.

[0034] The risk warning presentation constraints are handled by template assembly and time sequence arrangement to obtain the risk warning layer, and the start and end times and duration of the risk warning display are configured to specific values ​​within the rule base registration range.

[0035] The target image, subtitle layer, and risk warning layer are processed by compositing rendering and transcoding to generate a release version file for the target platform. The cropping parameters, subtitle layout parameters, risk warning parameters, and transcoding parameters are written into the release version record.

[0036] In a preferred embodiment, the generation and review module further includes:

[0037] After the release version is generated, the generation review module reviews the release version using multimodal compliance reasoning measures. The review process includes:

[0038] The released version is processed by extracting audio text, keyframes, keyframe screen text, subtitle text, and final risk warning text. The keyframe sequence is processed by screen compliance detection to obtain the prohibited element identification results.

[0039] Visual compliance detection refers to: reading the content review constraints corresponding to the target platform from the system's rule base, mainly disabling visual elements;

[0040] Perform uniform preprocessing on each keyframe, including uniform scaling, pixel normalization, color space conversion, and decompression noise filtering;

[0041] The preprocessed keyframes are input into the object detection model for inference. The object detection model adopts a single-stage object detection network structure, which includes a backbone feature extraction network, a feature pyramid fusion network, and a detection head network. The detection head network outputs classification confidence tensors and bounding box regression tensors on feature maps at multiple scales. The classification confidence tensor gives the category number and confidence value corresponding to each candidate box, and the bounding box regression tensor gives the center coordinates, width, and height of the candidate box and converts them into image region coordinates.

[0042] The model file is fixed through the model registration form. The model registration form contains the model structure identifier, model weight file identifier, model weight file summary value, number of categories and export format identifier. Before inference, the model weight file summary value is checked for consistency according to the model registration form. After passing the check, the model is loaded and the model version is locked to participate in this test.

[0043] The model output is then decoded into a set of candidate detection boxes. Each candidate detection box in the set is written with a category number, category label, image region coordinates and confidence value. The category label is obtained by converting the category number through a preset mapping table, and only the prohibited element types covered by the mapping table are retained.

[0044] Then, overlap suppression processing is performed on the candidate detection box set. The overlap of candidate boxes is calculated according to the overlap suppression threshold and duplicate boxes with overlap exceeding the threshold are removed. Then, low-confidence detection boxes are filtered according to the confidence threshold, and detection boxes with an area smaller than the threshold are removed according to the minimum target size threshold.

[0045] Temporal consistency aggregation processing is performed on the detection results of adjacent keyframes. Temporal consistency aggregation processing merges them according to category consistency and the degree of overlap of the image area, and requires effective identification of the frame continuity and duration of the same category of target.

[0046] The valid entries are output as the disabled element identification results. The disabled element identification results include the element category, first appearance time, last appearance time, representative keyframe index, screen area coordinates, confidence value and model version identifier for each entry.

[0047] The audio text, video text, final subtitle text, and final risk warning text are matched against the text review entries and risk warning entries in the rule base of the entire platform system. Combined with the results of the prohibited element identification, a review conclusion record is generated.

[0048] The review conclusion record includes the release version identifier, target platform identifier, rule base version identifier, and review timestamp. For each trigger result, the trigger entry number, trigger time period, and trigger screen area coordinates are also recorded.

[0049] If the review conclusion does not contain any prohibited elements of the target platform, the review is deemed passed; otherwise, the status is recorded and the application is placed in the queue for manual review, and the processing status is written into the review log.

[0050] As a preferred implementation, the release receipt module specifically includes:

[0051] For approved release versions, the corresponding platform identifier, release version file address, encoding parameters, risk warning rendering parameter summary, and review conclusion identifier are read from the release version record, and the interface configuration items that match the platform identifier are read from the system rule base.

[0052] Perform authentication processing according to the interface configuration items and call the platform publishing interface to submit a publishing request. Listen for platform feedback information and obtain the platform's returned information, which includes at least the publishing link.

[0053] When a release is successful, the platform returns a confirmation message, which is then written into the release confirmation message record. The release confirmation message record is then associated with the release version record by the release version identifier and written into the release ledger. When a release fails, the video is marked as abnormal and transferred to the manual review queue.

[0054] In a preferred embodiment, the certificate retrieval module specifically includes:

[0055] Based on the information in the release receipt record, read the platform identifier, release link, platform-side material identifier and release timestamp, and simultaneously read the release version record associated with the receipt record to obtain the summary value of the release version file, the summary of cropping and subtitle layout parameters, the summary of risk warning parameters and the rule base version identifier;

[0056] The system rule base is used to read back the readback strategy and interface configuration according to the platform identifier. The readback strategy includes the readback channel and the readback window.

[0057] The readback channel is limited to either the page crawling readback method based on the published link or the interface readback method based on the platform material query interface. The readback window is limited to one to thirty minutes after the published timestamp to prevent read failures caused by inconsistent published timestamps.

[0058] Initiate a readback request within the readback window, obtain the video actually presented on the platform, and solidify the obtained result into a presentation copy file, while writing it into the presentation copy record;

[0059] Feature extraction is performed on the presented copy and a consistency comparison is performed with the released version, including extracting keyframes and calculating keyframe summary values, generating keyframe text for the presented screen, parsing the presented subtitle text, and parsing the risk warning area to obtain the risk warning text and risk warning location coordinates.

[0060] In a preferred embodiment, the certificate retrieval module further includes:

[0061] After feature extraction is completed, the certificate recall module performs three types of consistency comparisons based on the corresponding features in the release version record: First, it performs a similarity comparison between the keyframe summary value of the release version and the keyframe summary value of the presentation to obtain the amount of image difference. The similarity comparison refers to: calculating the perceptual hash fingerprints of the two, calculating the Hamming distance between the two perceptual hash fingerprints, normalizing them to a similarity threshold, and judging the similarity compliance status according to the threshold.

[0062] Second, normalized text comparison was performed between the published version of the subtitle text and the presented subtitle text to obtain the subtitle difference.

[0063] Third, compare the risk warning parameters of the released version with the risk warning parsing results to obtain the difference in risk warning display duration and the risk warning position offset;

[0064] Write the differences in image quality, subtitle quality, risk warning display duration, and risk warning position offset into the consistency verification record, and read the corresponding tolerance threshold configuration item from the system rule base, registering the differences that exceed the tolerance as a deviation item list.

[0065] For each deviation, the deviation type, occurrence time period, screen area coordinates, difference value, trigger threshold configuration item identifier, and evidence index information for review are written in a fixed manner. The evidence index information includes representative keyframe index and representative screenshot file identifier.

[0066] After the consistency verification record and deviation item list are generated, the evidence reading module performs structured encapsulation processing on the release version record, audit conclusion record, release receipt record, presentation copy record, consistency verification record and deviation item list, thereby generating evidence material package index record and writing it into the evidence index ledger.

[0067] As a preferred embodiment, the stock disposal module specifically includes:

[0068] Continuously monitor the version number and effective date fields of the legal clause library and the platform specification library. When rule changes or expirations are detected, generate rule update event records and write the rule update event records into the rule update ledger.

[0069] After generating the rule update event record, the updated clause entries and specification entries are structured and mapped and written into the rule entry record, generating a new system rule base version record, and the difference entries between the old and new rule base version records are written into the rule difference list;

[0070] Based on the rule difference list, the affected scope of published materials is determined. The affected scope determination refers to:

[0071] The existing data processing module first extracts the platform identifier, constraint category, and change item identifier involved in this change from the rule difference list, and generates search conditions accordingly. In the release log, candidate released records are filtered by platform identifier and old rule library version identifier. Further filtering of candidate records is performed using risk warning template identifier, risk warning parameter summary, subtitle layout parameter summary, and content review category summary to ensure a matching relationship between candidate records and change items. The filtered results are deduplicated by release receipt identifier and platform-side material identifier, and records in the processed state are removed, thus obtaining the set of affected released records, which serves as input for subsequent batch review tasks.

[0072] Generate review task records for the affected set of published records and write them into the review task table. The review task records include platform identifier, publication receipt identifier, presentation copy identifier, new rule base version identifier, and review time window configuration.

[0073] In a preferred embodiment, the stock disposal module further includes:

[0074] The review task record calls the back-read certificate module to perform a back-read of the presented copy and performs consistency verification based on the new rule base version identifier, thereby generating a review result record; it is then associated with the corresponding published record and written into the existing review ledger.

[0075] After obtaining the review result record, generate a disposal instruction record for the published record marked as unsuccessful in the review conclusion and write it into the disposal task table. The disposal instruction record shall include at least a disposal type field, which includes disposal removal, disposal replacement and republication, and disposal with supplementary risk warning.

[0076] After the disposal is completed, the existing stock disposal module will write the disposal results into the existing stock compliance status ledger and write the disposal completion timestamp and execution receipt identifier to form a traceable existing stock risk convergence closed loop.

[0077] The beneficial effects of this invention are:

[0078] 1. Using the material version and rule base version as a unified benchmark, and running through the generation, review, release and verification process, the cross-platform multi-version deployment has a traceable and reproducible version link, reducing the compliance uncertainty risk caused by rule drift and version drift;

[0079] 2. Using the actual copy presented on the platform as the verification object, and outputting locatable deviation items, the content deviations introduced by the platform's secondary processing are transformed from invisible into quantifiable and accountable difference records, supporting subsequent revisions, reissues, and liability determination;

[0080] 3. Structure and encapsulate key process records into evidence material packages and indexes, so that compliance proof is transformed from conclusions or logs into a collection of materials that can be replayed and verified, thereby improving the certainty of spot checks, complaints and disputes, and audit filing;

[0081] 4. By monitoring rule updates and using difference-driven existing material location verification, a closed-loop system for batch processing of existing materials is supported, shortening the risk exposure window after rule changes and reducing manual processing costs. Attached Figure Description

[0082] Figure 1 This is a system structure diagram of the present invention;

[0083] Figure 2 This is a comparison diagram of the effects of the present invention compared to the traditional method. Detailed Implementation

[0084] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0085] Example 1 Figure 1 As shown in the system structure diagram of this invention, this embodiment provides a cross-platform video promotion system based on artificial intelligence, specifically including the following:

[0086] Registration and Locking Module: Records the version of promotional materials for the registration campaign; analyzes the restrictions of each platform to form a rule base for the entire platform system;

[0087] Generate review module: Generate release versions for each platform based on the material version; combine the rule base of the entire platform system and use multimodal compliance reasoning measures to review the release versions;

[0088] Release receipt module: For versions that have passed the review and are released, release orchestration measures are adopted to submit the released versions to various platforms; the platform return information is solidified into receipts, release receipt records are generated and written into the release ledger;

[0089] The evidence review module reads back the presented copy of the release receipt record; compares the released version with the presented copy, generates a consistency verification record and a list of deviation items, records relevant information, and forms an index record for the evidence material package;

[0090] The existing content handling module: It adopts update monitoring measures for the legal clause library and platform specification library, and updates the system rule library; based on the new system rule library, it reviews the published presentation copies, and takes further action on the published videos based on the review results;

[0091] The specific content of the above modules includes:

[0092] (1) Registration and Locking Module: Records the version of promotional materials; analyzes the restrictions of each platform to form a rule base for the entire platform system;

[0093] Specifically: The promotional materials are versioned and registered, generating material version records and writing them into the material version ledger. The specific process includes: after the module receives the original video file and the activity-related information, it first uses media unpacking measures on the original video file to obtain the video track file, audio track file and subtitle track file, and writes the video resolution, video frame rate, video duration, video encoding format, audio sampling rate, number of audio channels, audio encoding format and subtitle format type into the material metadata field;

[0094] Subsequently, speech transcription is applied to the audio track file to obtain audio text, and the speech transcription model identifier and transcription timestamp are written into the transcription source field; keyframe extraction is applied to the video track file to obtain a keyframe sequence, and text recognition is applied to the keyframe sequence to obtain a screen text sequence. The keyframe extraction time interval is read from the configuration table, and its specific setting can be modified according to the length of the video, preset to a range of 0.5 seconds to 2 seconds; subtitle parsing is applied to the subtitle track file to obtain a subtitle text sequence, and the subtitle parser identifier is written into the parsing source field.

[0095] After obtaining the audio text, video text, and subtitle text, the module uses a fixed-order concatenation method to generate a seed text string. It then uses a digest calculation method on the seed text string to generate a content fingerprint field. The digest algorithm uses the SHA-256 algorithm, and the algorithm identifier is written into the fingerprint algorithm field. The material metadata field, audio text field, video text field, subtitle text field, content fingerprint field, source field, and activity index field are all written into the material version record. This allows subsequent versions released on various platforms to be traced back to the same material version record and to reproduce the material parsing input and fingerprint calculation results.

[0096] By adopting standardized collection and structured merging measures for the restriction requirements of each platform, a rule base for the entire platform system is generated. The specific process is as follows: for each publishing platform, the platform constraint entries are read from the platform specification document archive, platform interface rule configuration or enterprise compliance configuration table, and the source number, effective date and applicable platform identifier of each platform constraint entry are written into the corresponding source field and effective field.

[0097] After collecting constraint items, the platform constraint items are processed using structured mapping, and rule item records are generated according to a unified field structure. The unified field structure includes constraint category field, parameter name field, parameter value field, value type field, and verification caliber field. The constraint category field includes screen specification constraints, subtitle style constraints, risk warning presentation constraints, and content review constraints, the specific contents of which are as follows:

[0098] Image specification constraints: Write the image aspect ratio parameters, target resolution parameters, and safety margin parameters into the rule entry record;

[0099] Subtitle style constraints: Write the subtitle area position parameters, font size range parameters, line spacing range parameters, and occlusion restriction parameters into the rule entry record;

[0100] Risk warning presentation constraints: The risk warning text template identifier, display start and end time points, display duration range, and display area location are written into the rule entry record;

[0101] Content moderation constraints: Write the lists of sensitive words, prohibited expressions, and prohibited image element types into the rule entry record;

[0102] For parameters that need to be expressed in numerical range, the numerical range and the basis for determination are written into the rule entry record simultaneously, and the basis for determination is written into the source field with the specification document number or configuration table number, so as to ensure that the rule entry is verifiable and traceable.

[0103] The obtained rules are stored separately according to the platform index to form a system rule library for the entire platform;

[0104] (2) Generate review module: Generate release versions for each platform based on the material version; combine the rule base of the entire platform system and adopt multimodal compliance reasoning measures to review the release versions;

[0105] Specifically: the video track file, audio track file, subtitle text, material metadata and content fingerprint field are read from the material version record, and the corresponding screen specification constraints, subtitle style constraints, risk warning presentation constraints and content review constraints are read from the system rule base according to the platform to be published, and the material content is processed according to the constraints.

[0106] Subsequently, the video track file is cropped and scaled to obtain a target image that meets the target aspect ratio and target resolution requirements. The cropping uses a center reference, and the available rendering area for subtitles and risk warnings is limited by using area constraints on the safety margin parameters.

[0107] The subtitle text is re-formatted to generate a subtitle layer. The re-formatting process includes numerically configuring the subtitle area position, font size and line spacing. The subtitle line width is calculated from the target resolution and safety margin, and the value range of font size and line spacing is limited to the range registered in the rule base.

[0108] The risk warning presentation constraints are handled by template assembly and time sequence arrangement to obtain the risk warning layer. The start and end times and duration of the risk warning display are configured as specific values ​​within the rule base registration range. The specific values ​​are determined by reading from the enterprise compliance configuration table and written into the parameter source field.

[0109] Finally, the target screen, subtitle layer, and risk warning layer are processed by compositing rendering and transcoding to generate the release version file for the target platform. The cropping parameters, subtitle layout parameters, risk warning parameters, and transcoding parameters are written into the release version record to ensure that the release version can be reproduced according to the recorded parameters.

[0110] After the release version is generated, the generation review module reviews the release version using multimodal compliance reasoning measures. The review process includes:

[0111] First, the audio track file of the release version is processed using speech-to-text to obtain audio text with timestamp alignment; the release version is processed using keyframe extraction at frame intervals to obtain a keyframe sequence, with the extraction rules being consistent with those of the source material version; the keyframe sequence is processed using on-screen text recognition to obtain on-screen text; the subtitle layer and risk warning layer are processed using rendering result parsing to obtain the final subtitle text and final risk warning text; and the keyframe sequence is processed using on-screen compliance detection to obtain the results of prohibited element identification.

[0112] The image compliance detection process involves: retrieving content moderation constraints from the system's rule base, primarily disabling certain image elements; performing unified preprocessing on each keyframe, including uniform scaling, pixel normalization, color space conversion, and noise reduction filtering; and inputting the preprocessed keyframes into the object detection model for inference. The object detection model employs a single-stage network structure, comprising a backbone feature extraction network, a feature pyramid fusion network, and a detection head network. The detection head network outputs classification confidence tensors and bounding box regression tensors on feature maps at multiple scales. The classification confidence tensor provides the value of each candidate box. The corresponding category ID and confidence score are used. The bounding box regression tensor provides the center coordinates, width, and height of the candidate boxes, which are then converted into image region coordinates. The model file is fixed through a model registration form, which includes the model structure identifier, model weight file identifier, model weight file summary value, number of categories, and export format identifier. Before inference, the model weight file summary value is checked for consistency according to the model registration form. After passing the check, the model is loaded and the model version is locked to participate in this detection. The model output is then decoded into a set of candidate detection boxes. Each candidate detection box in the set is written with a category ID, category label, image region coordinates, and confidence score. The degree value is calculated, where the category label is obtained by converting the category number through a preset mapping table, and only the prohibited element types covered by the mapping table are retained. Then, overlapping box suppression processing is performed on the candidate detection box set. The overlap degree of the candidate boxes is calculated according to the overlapping box suppression threshold, and duplicate boxes with an overlap degree exceeding the threshold are removed. Subsequently, low-confidence detection boxes are filtered according to the confidence threshold, and detection boxes with an area smaller than the minimum target size threshold are removed. The overlapping box suppression threshold is limited to the range of 0.3 to 0.7, the confidence threshold is limited to the range of 0.6 to 0.9 and is determined statistically through calibration set, and the minimum target size threshold is limited to the keyframe area. The percentage of each keyframe is determined by the minimum display specifications for platform watermarks and corner marks, ranging from 1% to 5%. Next, temporal consistency aggregation is performed on the detection results of adjacent keyframes. This aggregation is based on category consistency and the degree of overlap between the target and the image area. Targets of the same category are considered valid entries only if they appear in at least three consecutive frames with a coverage time span of at least 0.5 seconds. Finally, the valid entries are output as disabled element identification results. For each entry, the disabled element identification results include the element category, first appearance time, last appearance time, representative keyframe index, image area coordinates, confidence score, and model version identifier.

[0113] The audio text, video text, final subtitle text, and final risk warning text are matched against the text review entries and risk warning entries in the rule base of the entire platform system. Combined with the results of the prohibited element identification, a review conclusion record is generated. The review conclusion record is written with the release version identifier, target platform identifier, rule base version identifier, and review timestamp. For each trigger result, the trigger entry number, trigger time period, and trigger screen area coordinates are written.

[0114] If the review conclusion does not contain any prohibited elements of the target platform, the review is deemed passed; otherwise, the status is recorded and the application is placed in the queue for manual review, and the processing status is written into the review log.

[0115] (3) Release receipt module: For the released version that has passed the review, release arrangement measures are adopted to submit the released version to each platform; the information returned by the platform is solidified into a receipt, release receipt record is generated and written into the release ledger;

[0116] Specifically: For approved release versions, the corresponding platform identifier, release version file address, encoding parameters, risk warning rendering parameter summary, and review conclusion identifier are read from the release version record. The interface configuration items that match the platform identifier are read from the system rule base. The interface configuration items include interface address, authentication method, authentication credential reference identifier, request message field mapping rules, and submission frequency limit parameters.

[0117] According to the field mapping rules, the release version file and accompanying metadata are encapsulated in the request. The accompanying metadata must include at least the material title, placement tag, industry classification identifier, risk warning text version identifier and review conclusion number. The encapsulated request is then written into the release request record.

[0118] Then, according to the interface configuration items, the authentication process is performed and the platform publishing interface is called to submit the publishing request. The platform feedback information is listened for and the platform return information is obtained, including publishing status, presented copy, published video number, platform side review receipt number, publishing timestamp, publishing link, etc.

[0119] Upon successful publication, the platform's returned information is processed by receipt solidification. This process writes the returned information into the publication receipt record and associates the publication receipt record with the publication version record according to the publication version identifier, writing it into the publication ledger. This ensures that the subsequent evidence reading module can read back and present a copy based on the publication link or platform-side material identifier and complete consistency verification and evidence encapsulation. If publication fails, the video is marked as abnormal and transferred to the manual review queue.

[0120] (4) Evidence Retrieval Module: Retrieve the presented copy of the release receipt record; compare the release version and the presented copy, generate a consistency verification record and a list of deviation items, record relevant information, and form an index record of evidence material package;

[0121] Specifically: Based on the information in the release receipt record, read the platform identifier, release link, platform-side material identifier and release timestamp, and simultaneously read the release version record associated with the receipt record to obtain the summary value of the release version file, the summary of cropping and subtitle layout parameters, the summary of risk warning parameters and the rule base version identifier;

[0122] The system rule base reads the readback strategy and interface configuration based on the platform identifier. The readback strategy limits the readback channel and readback window. The readback channel is limited to either the page crawling readback method based on the published link or the interface readback method based on the platform material query interface. The readback window is limited to one to thirty minutes after the published timestamp to prevent read failures caused by inconsistent published timestamps.

[0123] Initiate a readback request within the readback window, obtain the media file or presentation stream actually presented on the platform, and solidify the obtained result into a presentation copy file, while writing it to the presentation copy record;

[0124] After the presentation copy file is solidified, the certificate return module performs feature extraction on the presentation copy and performs consistency comparison with the released version. This includes extracting keyframes from the presentation copy at frame intervals and calculating keyframe summary values. At the same time, it performs on-screen text recognition on the keyframes to generate presentation screen text; it performs subtitle parsing on the presentation copy to generate presentation subtitle text; and it parses the risk warning area in the presentation copy to obtain the risk warning text and risk warning location coordinates.

[0125] After feature extraction is completed, the certificate retrieval module performs three types of consistency comparisons based on the corresponding features in the release version record: First, it performs a similarity comparison between the keyframe summary value of the release version and the keyframe summary value of the presentation to obtain the image difference. The similarity comparison refers to: calculating the perceptual hash fingerprint for the keyframe of the release version and the keyframe of the presentation respectively; then calculating the Hamming distance between the two fingerprints, and normalizing the Hamming distance to convert it into a similarity score; if the similarity score is lower than the threshold configured in the rule base, it is considered that there is a difference in the image and it is recorded as the image difference; the threshold is statistically registered in the rule base according to the calibration set, and is initially set to 0.85 to 0.95;

[0126] Second, normalized text comparison was performed between the published version of the subtitle text and the presented subtitle text to obtain the subtitle difference.

[0127] Third, compare the risk warning parameters of the released version with the risk warning parsing results to obtain the difference in risk warning display duration and the risk warning position offset;

[0128] The verification module writes the differences in screen display, subtitle display, risk warning display duration, and risk warning location offset into the consistency verification record. It also reads the corresponding tolerance threshold configuration items from the system rule base according to the platform identifier, rule base version identifier, and difference indicator type, including the allowed threshold values, threshold units, and source numbers for various offsets. The verification module compares each difference with the corresponding tolerance threshold configuration item and registers the differences that exceed the tolerance as a deviation item list. At the same time, it writes the triggered threshold configuration item identifier and source number into the deviation item to support review and traceability.

[0129] For each deviation, the deviation type, occurrence time period, screen area coordinates, difference value, trigger threshold configuration item identifier, and evidence index information for review are written in a fixed manner. The evidence index information includes representative keyframe index and representative screenshot file identifier.

[0130] After the consistency verification record and deviation item list are generated, the back-reading certificate module performs structured encapsulation processing on the release version record, audit conclusion record, release receipt record, presentation copy record, consistency verification record and deviation item list, thereby generating evidence material package index record and writing it into the evidence index ledger;

[0131] The evidence package index records include: a list of materials, summary values ​​for each material, generation timestamp, activity index key, and platform index key, to support subsequent retrieval and playback by promotional activities and platform dimensions, and to be used for material issuance and consistency verification in scenarios such as random inspection review, complaint disputes, and regulatory inspections.

[0132] (5) Existing data handling module: The system rule base is updated by monitoring the update of the legal clauses library and the platform specification library; the published presentation copies are reviewed according to the new system rule base; and the published videos are further processed according to the review results.

[0133] Specifically: Periodically poll or subscribe to read the version number and effective date fields of the legal clause library and platform specification library. When a version number change or effective date is detected, a rule update event record is generated and written to the rule update ledger.

[0134] After generating the rule update event record, the updated clause entries and specification entries are structured and mapped and written into the rule entry record, thereby generating a new system rule base version record. The difference entries between the old and new rule base version records are written into the rule difference list for subsequent review scope positioning.

[0135] Based on the rule difference list, the affected scope of published materials is located. The affected scope location refers to the following steps: The existing material handling module first extracts the platform identifier, constraint category, and change item identifier involved in this change from the rule difference list, and generates search conditions accordingly. Then, in the publishing log, candidate published records are filtered out by platform identifier and old rule library version identifier. The candidate records are further filtered using risk warning template identifier, risk warning parameter summary, subtitle layout parameter summary, and content review category summary to ensure that the candidate records match the change items. Finally, the filtering results are deduplicated by publishing receipt identifier and platform-side material identifier, and records in the handling status are removed to obtain the set of affected published records, which serves as the input for subsequent batch review tasks.

[0136] Generate review task records for the affected set of published records and write them into the review task table. The review task records include platform identifier, publication receipt identifier, presentation copy identifier, new rule base version identifier, and review time window configuration.

[0137] Subsequently, the review task record calls the back-read certificate module to perform back-read on the presented copy and perform consistency verification based on the new rule base version identifier, thereby generating a review result record. The review result record includes the review conclusion identifier, the triggered rule entry number, the trigger time period, the coordinates of the trigger screen area, the difference value, etc., and is associated with the corresponding published record and written into the existing review ledger.

[0138] After obtaining the review result record, generate a disposal instruction record for the published record marked as unsuccessful in the review conclusion and write it into the disposal task table. The disposal instruction record includes disposal type field, disposal target identifier field, disposal parameter field, etc. The disposal type field includes removal disposal, replacement and re-release disposal, and supplementary risk warning re-release disposal.

[0139] Among them, the removal process is completed by calling the platform's removal interface or stopping the release interface and solidifying the platform's receipt; the replacement release process is completed by calling the generation review module to generate a release version that conforms to the new rule library version and calling the release receipt module to complete the re-release and update the release ledger; the supplementary risk warning re-release process is completed by updating the risk warning presentation parameters on the basis of the original material version and regenerating the release version before release.

[0140] After the disposal is completed, the existing stock disposal module will write the disposal results into the existing stock compliance status ledger and write the disposal completion timestamp and execution receipt identifier to form a traceable existing stock risk convergence closed loop.

[0141] like Figure 2The comparison chart of the effects of the present invention and the traditional method is shown in the figure. The black bars in the figure represent the effects of the present invention, and the gray bars represent the effects of the traditional invention. Through automatic compliance processing and automatic review of videos, the efficiency of cross-platform video delivery and review is greatly improved. In addition, the integrity of the evidence chain is effectively improved by recording the entire process. It can also be automatically modified according to rule changes, and can be applied to new rules more quickly.

[0142] Example 2: Taking cross-platform video promotion in the insurance industry as an example, an insurance company launched a "Hospitalization Medical Insurance" promotion campaign on January 20, 2026. The operations team submitted one original landscape video clip, 58 seconds long, with a resolution of 1920×1080, a frame rate of 30 frames per second, an audio sampling rate of 48000 Hz, and a timeline-based subtitle file containing 24 subtitles. The campaign was planned to be simultaneously launched on short video platform A, social media platform B, and news feed platform C. Platform A required a 9:16 aspect ratio, a target resolution of 1080×1920, a safety margin of 64 pixels, and a risk warning displayed continuously from 0 to 6 seconds. Platform B required a 1:1 aspect ratio, a target resolution of 1080×1080, a safety margin of 48 pixels, and a risk warning displayed continuously for at least 8 seconds after the product name appeared. Platform C required a 16:9 aspect ratio, a target resolution of 1920×1080, and a risk warning displayed continuously for at least 6 seconds at the end. The risk warning template text registered in the enterprise compliance configuration form is "Insurance products are underwritten by insurance companies, and the coverage is subject to the insurance terms. Please read the terms before purchasing insurance." At the same time, the list of prohibited expressions includes terms such as "guaranteed claims" and "100% compensation," and the list of prohibited image elements includes categories such as "official endorsement stamp style elements" and "principal guarantee promise corner mark style elements."

[0143] The registration and locking module records the version history of promotional materials and analyzes the restrictions of various platforms to form a rule base for the entire platform system. After receiving the activity identifier, material files, and a list of platforms, the module unpacks the material files to obtain video, audio, and subtitle tracks, and writes the video resolution, frame rate, duration, encoding format, audio sampling rate, number of channels, and subtitle format into the material metadata field. The module performs speech-to-text transcription on the audio track to obtain audio text and writes it into the transcription source field; it extracts a keyframe sequence from the video track at a 1-second interval registered in the configuration table, extracting a total of 58 frames, and performs text recognition on the keyframes to obtain a screen text sequence, writing it into the recognition source field; it performs subtitle parsing on the subtitle track to obtain a subtitle text sequence and writes it into the parsing source field. The module concatenates the audio text, screen text, and subtitle text in a fixed order to generate a seed text string, and performs a secure hash algorithm (256) on the seed text string to calculate a content fingerprint field. Finally, it writes the activity index field, material metadata field, audio text field, screen text field, subtitle text field, and content fingerprint field together into the material version record and material version ledger. The registration and locking module synchronously reads the screen specification constraints, subtitle style constraints, risk warning presentation constraints, and content review constraints of Platform A, Platform B, and Platform C from the platform specification document archive and the enterprise compliance configuration table, and generates rule entry records according to a unified field structure. It also generates a rule library version record and writes it into the system rule library ledger, so that subsequent generation, review, and reread verification of this activity will all reference this rule library version record.

[0144] The Generation and Review Module generates release versions for each platform based on the material version. It combines the platform-wide rule base and employs multimodal compliance reasoning to review the release versions. The Generation and Review Module reads the video track, audio track, subtitle text, and content fingerprint fields from the material version record. It also reads the target screen ratio, target resolution, safety margin, subtitle area position, and risk warning presentation parameters from the system rule base according to the target platform. For Platform A, the Generation and Review Module performs cropping and scaling on the video screen to generate a 1080×1920 vertical target screen, and limits the rendering area of ​​subtitles and risk warnings to a 64-pixel safety margin. It performs layout rearrangement on the subtitle text to generate a subtitle layer and fixes the subtitle layout parameters. It reads the risk warning template text from the enterprise compliance configuration table, sets the risk warning to be continuously displayed from 0 to 6 seconds, and renders it as a risk warning layer. It performs composite rendering and transcoding encapsulation on the target screen, subtitle layer, and risk warning layer to generate the release version file for Platform A, and writes the cropping parameters, subtitle layout parameters, risk warning parameters, and transcoding parameters into the Platform A release version record. The generation and review module uses the same material version record to generate corresponding release version files according to the 1080×1080 screen and risk warning duration constraints of Platform B, and the 1920×1080 screen and end risk warning constraints of Platform C, and writes them into their respective release version records. After the release version is generated, the generation and review module performs multimodal compliance reasoning review on each release version. During the review process, audio text with timestamps is generated for the audio track, screen text is generated for the keyframes, and the subtitle layer and risk warning layer are parsed to obtain the final subtitle text and the final risk warning text. The keyframes are identified for prohibited screen element types according to the screen compliance detection process disclosed in Example 1. The generation and review module performs matching verification between the audio text, screen text, final subtitle text, and final risk warning text and the prohibited expression list and sensitive word list in the system rule base. The prohibited element identification results are also matched and verified against the content review constraints, thereby generating a review conclusion record and writing it into the release version identifier, platform identifier, rule base version identifier, review timestamp, and trigger item number. In this embodiment, none of the release versions of Platform A, Platform B, and Platform C contained any prohibited expressions or were identified as prohibited screen element types, and the review conclusion record was marked as passed.

[0145] Release Confirmation Module: This module employs release orchestration measures for approved release versions, submitting them to various platforms. It solidifies platform feedback information, generating release confirmation records and writing them into the release ledger. The module reads approved release version records, obtaining the platform identifier, release version file address, review conclusion identifier, and accompanying metadata. It also reads platform interface configuration items from the system rule base, writing the interface address, authentication method, authentication credential reference identifier, and request message field mapping rules into these items. The module encapsulates the release version file and accompanying metadata into a request and initiates a platform submission request. Platform A returns the platform-side material identifier "PX-20260120-001", release timestamp "2026-01-20 10:05:12", and release link; Platform B returns the platform-side review confirmation number "RB-778812" and release link; Platform C returns the transcoding task identifier "TC-55031" and release link. The release receipt module will solidify the platform-side material identifier, review receipt number, transcoding task identifier, release timestamp, and release link into the release receipt record. It will also associate the release receipt record with the release version record according to the release version identifier and write it into the release ledger to form a traceable entry point for subsequent review and positioning.

[0146] The evidence review module reads back the published receipt record and its presented copy; compares the published version and the presented copy, generates a consistency verification record and a list of deviation items, records relevant information, and forms an evidence material package index record. The module uses the published receipt record as the entry point, reading the platform identifier, publishing link, platform-side material identifier, and publishing timestamp, and simultaneously reads the published version file summary value, cropping parameter summary, subtitle layout parameter summary, risk warning parameter summary, and rule base version identifier of the corresponding published version record. According to the readback configuration registered in the system rule base, the module initiates a readback request 5 minutes after the publishing timestamp, obtains the actual presented media file on the platform and solidifies it into a presented copy file, and simultaneously writes the presented copy file summary value, presentation resolution, presentation frame rate, and presentation duration to the presented copy record. The module extracts keyframes from the presented copy at a 1-second interval consistent with the review process and calculates the keyframe summary value. It performs text recognition on the keyframes to obtain the presented screen text, performs subtitle parsing on the presented copy to obtain the presented subtitle text, and performs parsing on the risk warning area to obtain the presented risk warning text and presentation location coordinates. The verification module compares the similarity of the keyframe summary values ​​of the released version and the keyframe summary values ​​of the presented version to obtain the screen difference. It performs normalization comparison on the subtitle text of the released version and the subtitle text of the presented version to obtain the subtitle difference. It compares the risk warning parameters of the released version with the parsing results of the presented risk warning to obtain the difference in the display duration and position offset of the risk warning. The above differences are written into the consistency verification record. In this embodiment, the verification result of Platform A shows that the platform automatically adds a watermark, causing the bottom of the risk warning to be obscured. The position offset of the risk warning is 80 pixels, which exceeds the tolerance threshold of 48 pixels registered in the system rule base. The verification module generates a list of deviation entries and writes the deviation type "risk warning obscuration", the occurrence time period "0 seconds to 6 seconds", the representative keyframe index and the representative screenshot file identifier. The evidence recovery module will perform structured encapsulation of the release version record, audit conclusion record, release receipt record, presentation copy record, consistency verification record and deviation item list, generate evidence material package index record and write it into the evidence index ledger. The evidence material package index record will include the material list, the summary value of each material, the generation timestamp, the activity index key and the platform index key to support subsequent random inspection and review and dispute playback.

[0147] The existing content handling module employs update monitoring measures for the regulatory clause library and platform specification library, updating the system rule library. Based on the new system rule library, it reviews the published presentation copies and handles the published videos accordingly. For deviation entries on Platform A, the existing content handling module triggers a handling process. Based on the occlusion position and offset recorded in the deviation entry, it adjusts the risk warning position parameter in Platform A's rule entry to move it upwards by 80 pixels while maintaining the safety margin, and writes the adjustment basis into the handling parameter field. The existing content handling module calls the generation and review module to regenerate the Platform A release version file based on the same material version record, calls the release receipt module to complete the replacement release and solidify the new release receipt record, and then calls the verification readback module to read back the new presentation copy and generate a consistency verification record. After the verification result meets the tolerance threshold, the handling result is written into the existing compliance status ledger and a handling completion timestamp is added. On January 25, 2026, the Existing Material Handling Module detected an update to the Platform B specification library version number and the effective date had arrived. It then generated a rule update event record and wrote the updated minimum display duration requirement for risk warnings into the new system rule library version record. Based on the rule difference list, the Existing Material Handling Module filtered out the set of published records from the publishing ledger that referenced the old rule library version from the Platform B list. It generated a batch review task for this set and called the back-read verification module to perform the review based on the new rule library version. For records that did not meet the new display duration requirement, it generated a "Supplementary Risk Warning Re-publishing Handling" task. After completing the re-publishing and back-read verification, it updated the publishing ledger and the Existing Material Compliance Status Ledger, thus achieving a closed loop of batch review and batch risk convergence for existing materials after the rule update.

[0148] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A cross-platform video promotion system based on artificial intelligence, characterized in that, include: Registration and Locking Module: Records the version of promotional materials for the registration campaign; analyzes the restrictions of each platform to form a rule base for the entire platform system; Generate review module: Generate release versions for each platform based on the material version; combine the rule base of the entire platform system and use multimodal compliance reasoning measures to review the release versions; Release confirmation module: For versions that have passed the review, release orchestration measures are adopted to submit the release version to each platform; The platform's returned information is validated and a release receipt record is generated and written into the release ledger. Retrieval Certificate Module: Retrieve a copy of the published receipt record; Compare the released version with the presented copy, generate a consistency verification record and a list of deviation items, record relevant information, and form an index record of evidence materials package; Existing data handling module: The module employs update monitoring measures for the legal provisions library and platform specification library, and updates the system rule library; based on the new system rule library, it reviews the published presentation copies, and takes further action on the published videos based on the review results.

2. The cross-platform video promotion system based on artificial intelligence according to claim 1, characterized in that: The registration and locking module specifically includes: Register promotional materials, generate material version records and write them into the material version ledger. The content includes: using media unpacking measures to obtain detailed material information from the original video files, including video track files, audio track files and subtitle track files, and writing video resolution, video frame rate, video duration, video encoding format, audio sampling rate, number of audio channels, audio encoding format and subtitle format type into the material metadata field. The audio text, keyframe sequence, and subtitle text sequence are extracted from the material details using a corresponding method, and the version identifier of the extraction method is written in; at the same time, text recognition is used on the keyframe sequence to obtain the on-screen text sequence; The module uses a fixed-order concatenation method to generate a seed text string from three types of text: audio text, video text, and subtitle text. It then uses a summary calculation method to generate a content fingerprint field from the seed text string. The acquired material metadata fields and the extracted information are written together into the material version record.

3. The cross-platform video promotion system based on artificial intelligence according to claim 2, characterized in that: The registration locking module also includes: The restrictions and requirements of each platform are analyzed and processed to generate a system rule base for the entire platform. The content includes: for each publishing platform, reading the platform constraint entries from the relevant acquisition source and recording the detailed configuration information of each platform constraint entry; Perform structured mapping processing on the platform constraint entries of the records, and generate rule entry records according to a unified field structure; The unified field structure includes constraint category fields and other constraint fields; Among them, the constraint category field includes screen specification constraints, subtitle style constraints, risk warning presentation constraints, and content review constraints, the specific contents of which are as follows: Image specification constraints: Write the image aspect ratio parameters, target resolution parameters, and safety margin parameters into the rule entry record; Subtitle style constraints: Write the subtitle area position parameters, font size range parameters, line spacing range parameters, and occlusion restriction parameters into the rule entry record; Risk warning presentation constraints: The risk warning text template identifier, display start and end time points, display duration range, and display area location are written into the rule entry record; Content moderation constraints: Write the lists of sensitive words, prohibited expressions, and prohibited image element types into the rule entry record; For parameters that need to be expressed in numerical range, the numerical range and the basis for determination are written into the rule entry record simultaneously, and the basis for determination is written into the source field in the form of a number, so as to ensure that the rule entry is verifiable and traceable; The obtained rules are stored separately according to the platform index, forming a system rule library for the entire platform.

4. The cross-platform video promotion system based on artificial intelligence according to claim 1, characterized in that: The approval generation module specifically includes: Retrieves relevant information from the material version history, and retrieves the corresponding screen specification constraints, subtitle style constraints, risk warning presentation constraints, and content review constraints from the system rule base according to the platform to be published, and processes the material content according to the constraints; The video track file is cropped and scaled to obtain a target image that meets the target screen ratio and target resolution requirements. The cropping uses a center reference and the available rendering area for subtitles and risk warnings is limited by using area constraints on the safety margin parameters. The subtitle text is re-formatted and rearranged to generate a subtitle layer. The re-formatting process includes numerically configuring the subtitle area position, font size and line spacing. The subtitle line width is calculated from the target resolution and the safety margin, and the value range of font size and line spacing is limited to the range registered in the rule base. The risk warning presentation constraints are handled by template assembly and time sequence arrangement to obtain the risk warning layer, and the start and end times and duration of the risk warning display are configured to specific values ​​within the rule base registration range. The target image, subtitle layer, and risk warning layer are processed by compositing rendering and transcoding to generate a release version file for the target platform. The cropping parameters, subtitle layout parameters, risk warning parameters, and transcoding parameters are written into the release version record.

5. The cross-platform video promotion system based on artificial intelligence according to claim 4, characterized in that: The approval generation module also includes: After the release version is generated, the generation review module reviews the release version using multimodal compliance reasoning measures. The review process includes: The released version is processed by extracting audio text, keyframes, keyframe screen text, subtitle text, and final risk warning text. The keyframe sequence is processed by screen compliance detection to obtain the prohibited element identification results. Visual compliance detection refers to: reading the content review constraints corresponding to the target platform from the system's rule base, mainly disabling visual elements; Perform uniform preprocessing on each keyframe, including uniform scaling, pixel normalization, color space conversion, and decompression noise filtering; The preprocessed keyframes are input into the object detection model for inference. The object detection model adopts a single-stage object detection network structure, which includes a backbone feature extraction network, a feature pyramid fusion network, and a detection head network. The detection head network outputs classification confidence tensors and bounding box regression tensors on feature maps at multiple scales. The classification confidence tensor gives the category number and confidence value corresponding to each candidate box, and the bounding box regression tensor gives the center coordinates, width, and height of the candidate box and converts them into image region coordinates. The model file is fixed through the model registration form. The model registration form contains the model structure identifier, model weight file identifier, model weight file summary value, number of categories and export format identifier. Before inference, the model weight file summary value is checked for consistency according to the model registration form. After passing the check, the model is loaded and the model version is locked to participate in this test. The model output is then decoded into a set of candidate detection boxes. Each candidate detection box in the set is written with a category number, category label, image region coordinates and confidence value. The category label is obtained by converting the category number through a preset mapping table, and only the prohibited element types covered by the mapping table are retained. Then, overlap suppression processing is performed on the candidate detection box set. The overlap of candidate boxes is calculated according to the overlap suppression threshold and duplicate boxes with overlap exceeding the threshold are removed. Then, low-confidence detection boxes are filtered according to the confidence threshold, and detection boxes with an area smaller than the threshold are removed according to the minimum target size threshold. Temporal consistency aggregation processing is performed on the detection results of adjacent keyframes. Temporal consistency aggregation processing merges them according to category consistency and the degree of overlap of the image area, and requires effective identification of the frame continuity and duration of the same category of target. The valid entries are output as the disabled element identification results. The disabled element identification results include the element category, first appearance time, last appearance time, representative keyframe index, screen area coordinates, confidence value and model version identifier for each entry. The audio text, video text, final subtitle text, and final risk warning text are matched against the text review entries and risk warning entries in the rule base of the entire platform system. Combined with the results of the prohibited element identification, a review conclusion record is generated. The review conclusion record includes the release version identifier, target platform identifier, rule base version identifier, and review timestamp. For each trigger result, the trigger entry number, trigger time period, and trigger screen area coordinates are also recorded. If the review conclusion does not contain any prohibited elements of the target platform, the review is deemed passed; otherwise, the status is recorded and the application is placed in the queue for manual review, and the processing status is written into the review log.

6. The cross-platform video promotion system based on artificial intelligence according to claim 1, characterized in that: The release receipt module specifically includes: For approved release versions, the corresponding platform identifier, release version file address, encoding parameters, risk warning rendering parameter summary, and review conclusion identifier are read from the release version record, and the interface configuration items that match the platform identifier are read from the system rule base. Perform authentication processing according to the interface configuration items and call the platform publishing interface to submit a publishing request. Listen for platform feedback information and obtain the platform's returned information, which includes at least the publishing link. When a release is successful, the platform returns a confirmation message, which is then written into the release confirmation message record. The release confirmation message record is then associated with the release version record by the release version identifier and written into the release ledger. When a release fails, the video is marked as abnormal and transferred to the manual review queue.

7. The cross-platform video promotion system based on artificial intelligence according to claim 1, characterized in that: The certificate retrieval module specifically includes: Based on the information in the release receipt record, read the platform identifier, release link, platform-side material identifier and release timestamp, and simultaneously read the release version record associated with the receipt record to obtain the summary value of the release version file, the summary of cropping and subtitle layout parameters, the summary of risk warning parameters and the rule base version identifier; The system rule base is used to read back the readback strategy and interface configuration according to the platform identifier. The readback strategy includes the readback channel and the readback window. The readback channel is limited to either the page crawling readback method based on the published link or the interface readback method based on the platform material query interface. The readback window is limited to one to thirty minutes after the published timestamp to prevent read failures caused by inconsistent published timestamps. Initiate a readback request within the readback window, obtain the video actually presented on the platform, and solidify the obtained result into a presentation copy file, while writing it into the presentation copy record; Feature extraction is performed on the presented copy and a consistency comparison is performed with the released version, including extracting keyframes and calculating keyframe summary values, generating keyframe text for the presented screen, parsing the presented subtitle text, and parsing the risk warning area to obtain the risk warning text and risk warning location coordinates.

8. The cross-platform video promotion system based on artificial intelligence according to claim 7, characterized in that: The certificate retrieval module also includes: After feature extraction is completed, the certificate recall module performs three types of consistency comparisons based on the corresponding features in the release version record: First, it performs a similarity comparison between the keyframe summary value of the release version and the keyframe summary value of the presentation to obtain the amount of image difference. The similarity comparison refers to: calculating the perceptual hash fingerprints of the two, calculating the Hamming distance between the two perceptual hash fingerprints, normalizing them to a similarity threshold, and judging the similarity compliance status according to the threshold. Second, normalized text comparison was performed between the published version of the subtitle text and the presented subtitle text to obtain the subtitle difference. Third, compare the risk warning parameters of the released version with the risk warning parsing results to obtain the difference in risk warning display duration and the risk warning position offset; Write the differences in image quality, subtitle quality, risk warning display duration, and risk warning position offset into the consistency verification record, and read the corresponding tolerance threshold configuration item from the system rule base, registering the differences that exceed the tolerance as a deviation item list. For each deviation, the deviation type, occurrence time period, screen area coordinates, difference value, trigger threshold configuration item identifier, and evidence index information for review are written in a fixed manner. The evidence index information includes representative keyframe index and representative screenshot file identifier. After the consistency verification record and deviation item list are generated, the evidence reading module performs structured encapsulation processing on the release version record, audit conclusion record, release receipt record, presentation copy record, consistency verification record and deviation item list, thereby generating evidence material package index record and writing it into the evidence index ledger.

9. The cross-platform video promotion system based on artificial intelligence according to claim 1, characterized in that: The existing stock disposal module specifically includes: Continuously monitor the version number and effective date fields of the legal clause library and the platform specification library. When rule changes or expirations are detected, generate rule update event records and write the rule update event records into the rule update ledger. After generating the rule update event record, the updated clause entries and specification entries are structured and mapped and written into the rule entry record, generating a new system rule base version record, and the difference entries between the old and new rule base version records are written into the rule difference list; Based on the rule difference list, the affected scope of published materials is determined. The affected scope determination refers to: The existing data processing module first extracts the platform identifier, constraint category, and change item identifier involved in this change from the rule difference list, and generates search conditions accordingly. In the release log, candidate released records are filtered by platform identifier and old rule library version identifier. Further filtering of candidate records is performed using risk warning template identifier, risk warning parameter summary, subtitle layout parameter summary, and content review category summary to ensure a matching relationship between candidate records and change items. The filtered results are deduplicated by release receipt identifier and platform-side material identifier, and records in the processed state are removed, thus obtaining the set of affected released records, which serves as input for subsequent batch review tasks. Generate review task records for the affected set of published records and write them into the review task table. The review task records include platform identifier, publication receipt identifier, presentation copy identifier, new rule base version identifier, and review time window configuration.

10. A cross-platform video promotion system based on artificial intelligence according to claim 9, characterized in that: The existing stock disposal module also includes: The review task record calls the back-read certificate module to perform a back-read of the presented copy and performs consistency verification based on the new rule base version identifier, thereby generating a review result record; it is then associated with the corresponding published record and written into the existing review ledger. After obtaining the review result record, generate a disposal instruction record for the published record marked as unsuccessful in the review conclusion and write it into the disposal task table. The disposal instruction record shall include at least a disposal type field, which includes disposal removal, disposal replacement and republication, and disposal with supplementary risk warning. After the disposal is completed, the existing stock disposal module will write the disposal results into the existing stock compliance status ledger and write the disposal completion timestamp and execution receipt identifier to form a traceable existing stock risk convergence closed loop.

Images