Video intelligent frame extraction system and method for third-party payment merchant submission audit, computer and storage medium

Abstract

The application discloses a kind of video intelligent frame extraction system, method, computer and storage medium for third-party payment merchant application auditing, video information analysis and adaptive module are according to video time length matching scene detection threshold and calculate target sampling frame number under the upper and lower limit constraint of target sampling frame number;Scene mutation detection module detects scene mutation with threshold to form the first group of candidate frames;High-risk area ROI enhancement sampling module independently monitors sensitive screen area and forces high-frequency sampling to form the fourth group of candidate frames;Motion supplementary detection module is scanned with lower threshold again to capture subtle operation and form the second group of candidate frames;Forced uniform bottom frame extraction module is uniformly sampled under interval constraint to form the third group of candidate frames;Perception hash deduplication and output module merges four groups of candidate frames, and after time proximity filtering, perception hash deduplication and still semantic discrimination, output standardized key frame and summary file.The application can reduce artificial auditing and model processing cost.

Description

Technical Field

[0001] This invention relates to third-party payment, merchant application review, and video processing technology, specifically to a video intelligent frame extraction system, method, computer, and storage medium for third-party payment merchant application review. Background Technology

[0002] In the third-party payment industry, merchants applying for onboarding are required to record their business processes as part of their application review materials. The recording method typically involves the merchant recording the entire business operation process (such as browsing products, placing orders, making payments, and booking services) from their own mobile app; or recording the business page operation process in a browser on a computer. After recording, the merchant submits the video along with their application materials. The risk control personnel of the third-party payment institution will use this video to determine whether the merchant's actual business operations are compliant. Key review points usually include: whether the screen recording shows loan processes or information with an annualized interest rate exceeding 24%; whether e-commerce merchants sell prohibited goods such as virtual currency, recharge cards, phone credit, antiques, paintings, precious metals, and jewelry; and whether it displays high-risk function entry points such as recharge, transfer, wallet, earnings, and promotion.

[0003] Merchant onboarding screen recordings differ fundamentally from general video recordings in their visual composition. Video recordings primarily feature camera cuts and scene transitions, with a relatively even temporal distribution. In contrast, merchant onboarding screen recordings are operation-driven UI recordings, where visual changes are entirely triggered by the merchant's demonstration actions, exhibiting a pulse-like distribution of sudden-still-sudden changes. Specifically, this manifests as a prolonged period of stillness when displaying a product details page or business description, followed by a rapid and dramatic change in the visuals when clicking to navigate to the next business step, a pop-up payment confirmation window, or scrolling through a product list. Furthermore, the recording duration varies significantly among different merchants, ranging from a simple product demonstration of ten seconds to a complete business process demonstration of over ten minutes. The density of operations and the frequency of visual changes differ significantly depending on the merchant's business type.

[0004] The existing technology has the following main technical problems when handling the above-mentioned on-site screen recording review:

[0005] (1) The workload of manual frame-by-frame review is huge. The duration of the business screen recordings submitted by merchants ranges from tens of seconds to more than ten minutes. When a batch of merchants join at the same time, a large number of review work orders are backed up, resulting in high labor costs and a longer merchant onboarding cycle.

[0006] (2) General fixed interval and fixed threshold frame extraction tools have poor compatibility with screen recording screens. Merchant business screen recordings involve frequent page switching, up and down scrolling, pop-up windows and other operations. There are many subtle interface changes and they are unevenly distributed. Using fixed interval sampling makes it easy to miss key review screens such as business pop-ups and operation jumps.

[0007] (3) The content displayed in violation of regulations often appears only briefly during a specific operation. For example, a loan interest rate display page that pops up after a merchant clicks on a function entry during a demonstration, or a prohibited product item that flashes by while scrolling through a product list, usually lasts for less than a second. Single-scene detection or fixed-interval frame sampling cannot reliably capture such instantaneous images, posing a risk of compliance oversight.

[0008] (4) The screen recording contains a large number of long static frames. Merchants often display a business page for a long time in order to record it clearly. The general frame extraction solution will output a large number of highly similar screenshots, which can easily cause visual fatigue during manual review and lead to inconsistent review standards. If the complete original video is sent to the multimodal vision model frame by frame, a large number of invalid static frames will generate redundant image token consumption, increase the model inference time and server computing power. At the same time, static frames will interfere with the model's feature extraction and increase the probability of misjudgment.

[0009] (5) The threshold and sampling interval of the existing frame extraction tool are fixed parameters and cannot be flexibly adjusted according to the video length. When a higher threshold is used for short videos, it is easy to lose subtle operation scenes, while when a lower threshold is used for long videos, a large number of noisy frames are generated, making it difficult to meet the dual requirements of "full coverage of business processes" and "low redundancy of output screen" for the submission review at the same time.

[0010] (6) UI rendering artifacts and interference from transition animations. Merchant business screen recordings contain a large number of intermediate frames such as page loading spinning animations, pop-up fade-in and fade-out transitions, and list scrolling afterimages. General scene detection algorithms are prone to misjudging these transition frames with no review value as keyframes, while simply increasing the threshold for filtering will also filter out real business pop-ups, making it impossible to reliably distinguish between animation frames and valid content frames.

[0011] (7) The instantaneous appearance and obscuring of illegal content. Illegal information such as loan interest rate display and prohibited product items may only appear in a very short window (less than 0.5 seconds) after the merchant clicks on a certain function entry point before being covered by subsequent pop-ups or replaced by page jumps. Existing fixed-interval frame sampling or single threshold detection schemes cannot guarantee the capture of effective frames within this very short window, and general motion detection is easily overwhelmed by continuous frame changes generated by sliding operations, making it impossible to specifically focus on high-risk areas.

[0012] (8) Confusion between "invalid stillness" and "valid stillness" in screen recording. A merchant staying on a certain business page for a long time is considered a valid business display, while a long blank page or loading page caused by recording operation error is considered an invalid stillness. The general perceptual hash deduplication scheme treats the two equally, which may excessively remove valid business display pages or retain a large number of uninformative and choppy screens.

[0013] (9) Cross-device / resolution inter-frame difference baseline drift. Screen recording files submitted by different merchants vary greatly in terms of mobile phone model, resolution, and encoding bitrate. The normalized inter-frame difference of the same operation is inconsistent on different devices. The fixed threshold scheme causes false triggering due to encoding noise on some low bitrate devices, and misses subtle operations on high-resolution devices due to the high threshold.

[0014] The aforementioned nine issues, when combined, lead to reduced merchant onboarding efficiency and increased platform compliance risks. Existing manual review methods and general video frame extraction solutions are ineffective in adapting to the specific characteristics of merchant screen recordings, necessitating the design of a dedicated intelligent video frame extraction solution for this business scenario. Summary of the Invention

[0015] The purpose of this invention is to provide a video intelligent frame extraction system, method, computer, and storage medium for third-party payment merchant onboarding review, to solve the technical problems of poor adaptability of existing general frame extraction tools to merchant business screen recording features, easy loss of key review frames, excessive redundant static frames, and inability to balance the adaptation of long and short videos. Furthermore, this invention addresses issues unique to UI screen recording such as rendering artifact interference, instantaneous appearance of illegal content, confusion between invalid and valid static frames, and cross-device benchmark drift by introducing four technical means: UI steady-state verification mechanism, enhanced sampling of high-risk area ROI, static semantic discrimination, and adaptive benchmark calibration, to further improve the accuracy and creativity of the solution for merchant business screen recording.

[0016] The technical solution for achieving the purpose of this invention is: a video intelligent frame extraction system for third-party payment merchant application review, comprising:

[0017] The video information parsing and adaptive module is used to read the duration of the incoming review video, match the scene detection threshold corresponding to the preset duration range according to the video duration, and calculate the target sampling frame number constrained by the upper and lower limits of the target sampling frame number according to the video duration and the preset baseline frame rate.

[0018] The scene change detection module is used to perform frame-by-frame scene change detection on the video based on the current scene detection threshold, identify page switching, pop-up windows and business process jump events, extract the corresponding timestamps to extract key frames to form the first group of candidate frames, and count the number of valid scene frames extracted after the detection is completed. When the number of valid scene frames is lower than the scene frame trigger threshold, a trigger signal is sent to the motion supplementation detection module.

[0019] The high-risk area ROI enhanced sampling module is used to independently monitor predefined sensitive screen areas. When the pixel change rate of the ROI area exceeds the preset multiple of the global inter-frame difference average, high-frequency sampling is forcibly triggered and is not limited by the global scene detection threshold to form a fourth set of candidate frames.

[0020] The motion supplement detection module is used to perform a second scan of the video with a sensitivity threshold lower than the scene detection threshold after receiving a trigger signal, to capture the frames corresponding to page swiping and menu scrolling events, and form a second set of candidate frames;

[0021] The forced uniform bottom-feed frame extraction module is used to dynamically calculate the uniform sampling interval constrained by the shortest interval and the longest interval based on the target sampling frame number, regardless of whether the scene change detection module, the high-risk area ROI enhancement sampling module, and the motion supplementation detection module have extracted candidate frames. It performs uniform sampling on the entire video from beginning to end to form a third set of candidate frames.

[0022] The perceptual hash deduplication and output module is used to merge the candidate frames of the first, second, third and fourth groups, filter frames that are too close in interval according to the time proximity rule, calculate the difference perceptual hash value for each frame and remove visually similar frames according to the Hamming distance, and output a standardized keyframe set and frame extraction summary file according to the playback sequence.

[0023] Furthermore, the video information parsing and adaptive module is also used to read the resolution and frame rate of the video, take the still image at the beginning of the video to calculate the inter-frame noise variance σ as the device reference noise, and dynamically adjust the scene detection threshold θ to θ'=θ+k·σ, where k is a preset calibration coefficient with a value of k=0.5;

[0024] The preset duration intervals include: a first duration interval of less than 30 seconds, a second duration interval of greater than or equal to 30 seconds and less than or equal to 180 seconds, and a third duration interval of greater than 180 seconds; the scene detection threshold θ includes a first threshold of 0.06 corresponding to the first duration interval, a second threshold of 0.08 corresponding to the second duration interval, and a third threshold of 0.12 corresponding to the third duration interval;

[0025] The target number of sampling frames is calculated based on a baseline density of 15 frames per minute, with an upper limit of 300 frames and a lower limit of 20 frames.

[0026] Furthermore, the scene mutation detection module performs UI steady-state verification, takes the change area mask for the detected mutation candidate frame, calculates the inter-frame difference of the next 3 consecutive frames in the mask area, and if all three are less than 0.01, it is confirmed as a valid scene frame; otherwise, it is determined as a transition animation frame and is removed.

[0027] The scene frame trigger threshold for the scene mutation detection module is 10 frames.

[0028] Furthermore, the sensitive screen area of ​​the high-risk ROI enhancement sampling module includes the pop-up high-occurrence area from 50% to 85% down from the top of the screen and the status bar area from 0% to 10% down from the top of the screen. The preset multiplier is 2. The time interval between the extracted ROI frame and the existing candidate frame is checked. If it is less than 0.3 seconds, it is discarded. Otherwise, a fourth group of candidate frames is formed.

[0029] The sensitivity threshold of the motion supplement detection module is 0.05, and it filters frames with an interval of less than 0.5 seconds from existing scene frames.

[0030] Furthermore, the forced uniform bottom-feed frame extraction module checks the time interval between each extracted frame and the first, second, or fourth group of candidate frames during the sampling process. If the time interval is less than the proximity threshold, the frame is discarded immediately, and the frames that pass the check form the third group of candidate frames.

[0031] The minimum interval of the forced uniform bottom-out frame extraction module is 2 seconds, the maximum interval is 5 seconds, and the time proximity threshold is 1 second.

[0032] When the calculated sampling interval is less than the shortest interval, the shortest interval value is used; when it is greater than the longest interval, the longest interval value is used.

[0033] Furthermore, when the perceptual hash deduplication and output module calculates the difference perceptual hash, it scales the image to a 9×8 pixel grayscale image and compares the grayscale values ​​of adjacent pixels in each row to generate a binary hash value.

[0034] The perceptual hash deduplication and output module introduces static semantic discrimination for screen segments with a continuous static duration exceeding a preset threshold. If the screen contains business text information, the deduplication Hamming distance threshold is relaxed to 6; if it does not contain business text, the standard threshold of 4 is applied.

[0035] A video intelligent frame extraction method for third-party payment merchant application review, applied to the aforementioned video intelligent frame extraction system for third-party payment merchant application review, characterized in that the method includes the following steps:

[0036] S1. Video parsing and threshold adaptation: Read the duration of the merchant's application review video, match the scene detection threshold corresponding to the preset duration range according to the duration, and calculate the target sampling frame number constrained by the upper and lower limits of the target sampling frame number according to the video duration and the preset baseline frame rate.

[0037] S2, Scene Change Frame Extraction: Perform frame-by-frame scene change detection on the video using the current scene detection threshold, identify page switching, pop-up windows and business process jump events, extract the corresponding timestamps to extract key frames to form the first group of candidate frames; count the number of valid scene frames, if it is lower than the scene frame number trigger threshold, then trigger step S3.

[0038] S2.a. Enhanced sampling of high-risk ROI areas: This is executed in parallel with step S2. Predefined sensitive screen areas are independently monitored. When the pixel change rate of the ROI area exceeds a preset multiple of the global inter-frame difference average, a frame is forcibly extracted. This is not limited by the global scene detection threshold and forms the fourth group of candidate frames.

[0039] S3, Motion Supplement Frame Extraction: The video is scanned a second time with a sensitivity threshold lower than the scene detection threshold to capture the frames corresponding to page swiping and list scrolling events, forming a second set of candidate frames;

[0040] S4. Forced bottom-line frame extraction: Independent of the results of steps S2, S2.a and S3, the uniform sampling interval is dynamically calculated based on the target number of sampled frames, and uniform sampling is performed on the entire video from beginning to end to form a third group of candidate frames.

[0041] S5. Merging, filtering and deduplication: Merge the candidate frames of the first, second, third and fourth groups, filter frames that are too close in interval according to the time proximity rule, calculate the perceptual hash value of the difference between each frame and remove visually similar frames according to Hamming distance.

[0042] S6. Review Frame Output: Output a standardized set of keyframes and a frame summary file containing timestamps of each frame according to the playback sequence.

[0043] Furthermore, in step S1, the preset duration interval includes a first duration interval of less than 30 seconds, a second duration interval of greater than or equal to 30 seconds and less than or equal to 180 seconds, and a third duration interval of greater than 180 seconds. The scene detection threshold includes a first threshold of 0.06 corresponding to the first duration interval, a second threshold of 0.08 corresponding to the second duration interval, and a third threshold of 0.12 corresponding to the third duration interval.

[0044] The target number of sampling frames is calculated based on a baseline density of 15 frames per minute, with an upper limit of 300 frames and a lower limit of 20 frames.

[0045] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the video intelligent frame extraction method for third-party payment merchant application review.

[0046] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the video intelligent frame extraction method for third-party payment merchant application review.

[0047] Compared with the prior art, the significant advantages of this invention are:

[0048] 1. Specialized adaptations were made for the characteristics of merchant onboarding screen recording operations and the pulse-like distribution of screen changes. The ability to capture fragmented operations such as sliding, pop-ups, and interface jumps is superior to the general fixed-interval frame-skipping solution.

[0049] 2. By adopting a video duration-based adaptive threshold strategy, the complete extraction of key frames is ensured in short video scenarios, while the number of output frames is controlled within a reasonable range in long video scenarios, avoiding the disadvantages of fixed parameter solutions for both long and short videos.

[0050] 3. Scene mutation detection, motion supplementation detection and forced uniform bottom-up frame extraction correspond to event-driven drastic change capture, event-driven subtle change supplementation and time-driven full-process coverage, respectively. The three mechanisms complement each other and work together to ensure accurate extraction of operation-intensive areas and complete coverage of operation-sparse areas.

[0051] 4. Through a dual-layer redundant filtering of temporal proximity filtering and perceptual hash deduplication, visually highly similar static frames are effectively removed, reducing the processing load of downstream manual review or model recognition, and reducing review fatigue and computing power consumption.

[0052] 5. Outputs a standardized set of keyframes and timestamped frame summaries, which can be directly used as input for a general multimodal vision model without requiring specific model fine-tuning to assist in compliance identification.

[0053] 6. To address the unique rendering artifacts and transition animation interference in UI screen recording, a UI steady-state verification mechanism is used to distinguish between motion effect frames and valid content frames, avoiding misjudging intermediate states such as page loading animations and pop-up fade-in / fade-out as keyframes, thereby improving the effectiveness of output frame review.

[0054] 7. By enhancing the sampling mechanism through high-risk area ROI, local high-frequency monitoring is implemented in pop-up high-incidence areas and status bar areas to effectively capture extremely short window images of illegal information that flash for a moment (less than 0.5 seconds), making up for the omission defect of the global threshold solution in detecting instantaneous illegal content;

[0055] 8. By using an adaptive benchmark calibration mechanism, the system estimates the device's background noise using still images at the beginning of the video and dynamically corrects the scene detection threshold, thus solving the threshold failure problem caused by different mobile phone models, resolutions, and encoding bitrates, and ensuring the stable performance of the solution on massive heterogeneous screen recording files. Attached Figure Description

[0056] Figure 1 This is a diagram of the intelligent video frame extraction system architecture of the present invention.

[0057] Figure 2 This is a flowchart illustrating the complete implementation steps of the intelligent video frame extraction method of the present invention. Detailed Implementation

[0058] 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.

[0059] This invention proposes a video intelligent frame extraction method, system, computer, and storage medium for third-party payment merchant application review. It employs a multi-level adaptive threshold strategy for video duration, combining scene change detection, motion supplementation detection, forced uniform bottom-line frame extraction, and perceptual hashing for dual-layer deduplication. Addressing the characteristics of merchant business screen recordings—numerous swiping operations, frequent pop-ups, and fragmented interface changes—it completely captures key business scenes and automatically removes redundant static frames. The output concise keyframe set with temporal summaries can be directly integrated with general multimodal vision models, leveraging temporal context information to reduce model misjudgment rate and computational overhead.

[0060] like Figure 1 As shown, the present invention provides a video intelligent frame extraction system for third-party payment merchant application review. Based on the screen recording characteristics of third-party payment merchant application business, a six-level frame extraction processing system is constructed: First, the basic video parameters are parsed and the scene detection threshold is adaptively configured according to the duration and the device reference noise is calibrated. Then, scene change detection (including UI steady state verification), high-risk area ROI enhancement sampling, motion supplementation detection, and forced uniform bottom-line frame extraction are executed in sequence. Finally, the system outputs a set of key frames that cover the entire process and are free of redundancy through perceptual hash deduplication (including static semantic discrimination).

[0061] Merchant application screen recordings differ from regular videos in their visual characteristics. These recordings are business process demonstration videos created by merchants themselves to demonstrate their application for onboarding with a third-party payment platform. The recording covers the complete process of the merchant's business operations: on mobile devices, it typically starts from the merchant's own app's entry page and demonstrates the business steps sequentially, including product browsing, category filtering, order placement, payment, and order viewing; on computers, it usually records the operation process of the merchant's website or business system in a browser, including page navigation, form filling, and function clicks. After recording, the video is submitted to the payment institution as application material, where risk control personnel review whether the merchant's actual business operations violate regulations.

[0062] Unlike continuously recorded videos such as those from film and television or surveillance cameras, the screen changes in merchant business recordings are entirely triggered by the merchant's demonstration operations, and have the following typical characteristics:

[0063] First, the visual changes are distributed in a pulsed pattern. When a merchant displays a product details page, business description page, or service agreement page, the screen may remain still for tens of seconds or even longer; however, when clicking to jump to the next business step, a payment confirmation window pops up, or a product category page is switched, the screen changes drastically within two or three frames. This pattern of long-term stillness-sudden change-stillness is fundamentally different from the gradual shot transitions in film and video.

[0064] Second, a high proportion of the screen recordings involve subtle operations. Besides page-level navigation, the recordings contain numerous minor interface changes such as scrolling up and down product lists, expanding product details, switching category tags, and entering search keywords. These changes may appear small at the pixel level, but for review purposes, they could contain crucial clues of violations, such as the display of prohibited products, loan interest rate information, or access to high-risk functions.

[0065] Third, the duration varies greatly and the density of operations is uneven. A simple product demonstration screen recording may only be ten seconds long, while a screen recording that includes a complete demonstration of the ordering, payment, and after-sales process can be more than ten minutes long. Short videos have dense operations and frequent changes in the scene, while long videos contain long sections of static page displays. The two have completely opposite requirements for frame-skipping sensitivity.

[0066] Fourth, differences in business types lead to significant variations in screen content. E-commerce merchants' screen recordings primarily feature product images and lists, financial merchants' recordings mainly display forms and interest rates, and service merchants' recordings focus on appointment and confirmation processes. The frequency and magnitude of screen changes vary significantly across different business types, further increasing the difficulty of adapting a universal frame-skipping solution.

[0067] The aforementioned characteristics dictate that a general frame-dropping scheme using fixed thresholds and fixed intervals cannot simultaneously ensure the integrity of key image capture and the controllability of the output frame rate. A dedicated scheme is needed that can dynamically adjust the detection strategy based on video duration and operation density. The module's specific workflow is as follows:

[0068] 1. Video Information Parsing and Adaptive Module

[0069] The module reads basic parameters such as screen recording duration, resolution, and frame rate, and configures differentiated scene detection thresholds based on video duration. It also simultaneously sets minimum and maximum frame rate constraints for single videos to prevent over-sampling of short videos or under-sampling of long videos. The threshold parameters and target sampled frame rate output by this module serve as the basis for the operation of subsequent detection modules. Details are as follows:

[0070] Receive merchant application review video files and read the video duration T (in seconds), resolution (width × height pixels), and frame rate f (in fps). Automatically set a three-level scene detection threshold θ based on the video duration.

[0071] Duration T < 30 seconds: Use threshold ;

[0072] 30 seconds ≤ T ≤ 180 seconds: Use threshold ;

[0073] Duration T > 180 seconds: Use threshold .

[0074] The threshold mentioned above is the normalized inter-frame difference, which is calculated as follows: take two adjacent frames, scale them to a uniform size, calculate the average absolute value of the pixel grayscale difference, and then divide it by the maximum grayscale value of 255 for normalization, resulting in a difference value between 0 and 1. The smaller the threshold, the higher the sensitivity to changes in the image, and the easier it is to trigger frame skipping.

[0075] The three threshold levels are set based on the following: Based on tests and statistics of merchant screen recording samples accumulated in actual business operations, short videos less than 30 seconds in length typically involve intensive operations and frequent screen changes, requiring high sensitivity to capture subtle operations such as pop-ups and swipes. A threshold of 0.06 achieves a good balance between the key image capture rate and the proportion of noisy frames. Medium-length videos between 30 and 180 seconds have moderate operation density, and a threshold of 0.08 balances sensitivity and noise suppression. Long videos exceeding 180 seconds have a significantly increased proportion of static frames. If a low threshold is still used, image encoding noise and slight rendering jitter will generate a large number of invalid triggers. Increasing the threshold to 0.12 effectively suppresses these noisy frames.

[0076] Simultaneously, the first 5 seconds of the video (or the entire video if the video is less than 5 seconds long) are used to calculate the inter-frame noise variance σ as the device's baseline noise. The subsequent scene detection threshold is dynamically adjusted to θ' = θ + k·σ (k is a preset calibration coefficient, preferably k = 0.5) to adapt to baseline drift caused by different devices and encoding parameters. This calibration mechanism ensures that the same operation can be reliably detected under different mobile phone models, resolutions, and bitrates. The target sampling frame count N_target is also calculated: based on a baseline density of 15 frames per minute, i.e., N_target = T / 60 × 15, with a lower limit of N_min = 20 frames and an upper limit of N_max = 300 frames. When the calculated value is lower than the lower limit, the lower limit is used; when it is higher than the upper limit, the upper limit is used. The lower limit is set to ensure sufficient frames to cover the business process even for very short videos, while the upper limit is set to prevent excessive output frames from very long videos, which would increase the downstream processing burden.

[0077] 2. Scene Sudden Change Detection Module

[0078] The module receives the scene detection threshold output by the video information parsing and adaptive module, performs frame-by-frame abrupt scene detection, identifies events such as page switching, pop-up and closing, and business process jumps, extracts corresponding timestamps and keyframes to form the first set of candidate frames. After detection, the module counts the number of extracted scene frames. If the number is lower than the scene frame count trigger threshold, it sends a trigger signal to the motion supplementation detection module. Specifically:

[0079] The video is subjected to frame-by-frame scene change detection using the current scene detection threshold output by the video information parsing and adaptive module. Specifically, adjacent frame pairs (F_t, F_{t+1}) are taken sequentially according to the frame rate f, and the normalized inter-frame difference D_t between them is calculated. When D_t≥θ', a suspected scene change occurs at time t, and the UI steady-state verification process is initiated: it checks whether the changed area of ​​the frame remains pixel-stable in the subsequent 3 consecutive frames (i.e., the inter-frame difference of the changed area is less than 0.01). If it is stable, it is confirmed as a valid scene frame, the timestamp is recorded, and a frame image is extracted at the corresponding position as the first group of candidate frames; if it is unstable, it is determined to be a transition animation or rendering artifact frame, and is discarded and not included in the number of scene frames.

[0080] This module primarily identifies the following types of screen changes: page-level switching and business process jumps (such as going from the product list page to the product details page, jumping from the shopping cart page to the payment confirmation page, and jumping from the homepage to the loan application process page); pop-up windows and their closing (such as payment confirmation pop-ups, coupon redemption pop-ups, risk warning pop-ups, and product specification selection pop-ups); and interface refreshes triggered by clicking on function entry points (such as switching product category tags, refreshing the list after clicking the search button, and jumping to the next page after clicking "view all").

[0081] After the detection is completed, the total number of extracted scene frames N_scene is counted. If N_scene < scene frame trigger threshold (in this embodiment, the threshold is 10 frames), it is determined that the key frame extraction of the current video is insufficient (this trigger threshold is an empirical estimate based on the shortest business process screen recording containing at least 3 to 5 business process pages, each of which may be accompanied by 1 to 2 pop-up windows or jump operations), and a trigger signal is sent to the motion supplementation detection module.

[0082] 3. High-risk area ROI enhanced sampling module

[0083] It runs in parallel with the scene change detection module, independently monitoring predefined sensitive ROI areas in the video, and is specifically adapted to address the characteristic of instantaneous appearance of non-compliant information in financial compliance audits. Details are as follows:

[0084] The Region of Interest (ROI) is defined as: the 50%~85% area from the top of the screen downwards (high-frequency pop-up area) and the 0%~10% area from the top of the screen downwards (status bar / interest rate display area). The pixel change rate within the ROI area is calculated frame by frame. When the pixel change rate of the ROI area exceeds twice the current global average inter-frame difference, a frame is forcibly extracted at that moment, regardless of the global scene detection threshold. The time interval between the extracted ROI frame and existing candidate frames is checked. If it is less than 0.3 seconds, it is considered a related frame of the same operation and is discarded; otherwise, it is retained as the fourth group of candidate frames.

[0085] This module is designed based on the fact that illegal information such as loan interest rates and prohibited goods is often covered within 0.5 seconds of the pop-up window appearing, and the global threshold cannot guarantee its capture. By focusing on high-risk areas to achieve a differentiated sampling strategy that is locally frequent and globally unchanged, it can effectively cover the instantaneous illegal window without significantly increasing the total number of frames.

[0086] 4. Exercise supplementation detection module

[0087] It only activates upon receiving a trigger signal from the scene change detection module. Using a sensitivity threshold lower than the scene detection threshold, it performs a secondary scan of the video to capture images corresponding to subtle interface changes such as page scrolling, menu scrolling, and content loading. Detection results that are too close to the time interval of existing scene frames are filtered out to form a second set of candidate frames. Details are as follows:

[0088] This module only activates upon receiving a trigger signal from the scene change detection module, and is not executed for every video. After activation, it performs a second frame-by-frame scan of the video with a sensitivity threshold θ_motion=0.05. This sensitivity threshold is lower than the scene detection threshold, aiming to capture subtle operations with small changes in the image, including: vertical scrolling and swiping of product lists or business pages; expansion and collapse of category menus and filter dropdowns; asynchronous loading and transition animations of page content; and partial refresh of the list after entering search keywords.

[0089] For each detected motion event, the interval between its timestamp and the most recent timestamp of an existing scene frame is determined. If the interval is less than 0.5 seconds, it is considered a related screen change caused by the same operation and is filtered out; otherwise, the frame is retained and added to the second group of candidate frames. The 0.5-second time window is set based on the fact that, under normal operating rhythm, a page jump and accompanying animation caused by a click operation are usually completed within 0.3 to 0.5 seconds, and screen changes exceeding this interval are likely to be independent new operations.

[0090] 5. Forced uniform bottom-feed frame extraction module

[0091] Independent of the execution results of preceding scene detection and motion supplementation detection, mandatory segmented sampling is always performed on the entire video. The sampling interval is dynamically calculated based on the target sampling frame number output by the video information parsing and adaptive module, and minimum and maximum interval protection values ​​are set. Frames are uniformly extracted from the video from beginning to end, and frames with time intervals too close to existing candidate frames are removed to form a third set of candidate frames, ensuring that there are no time blind spots throughout the entire process. Specifically:

[0092] This module executes independently of the scene change detection and motion supplementation detection results, regardless of how many frames were extracted in the preceding steps. Its purpose is to prevent complete framelessness in certain time periods due to excessively high threshold settings or special video content (such as a merchant recording only a static display of a single page, or a pure browsing screen recording without any navigation). Specifically, it calculates the theoretical sampling interval Δt = T / N_target based on the target number of sampled frames N_target and the total video duration T, and imposes upper and lower limits on Δt: Δt_min = 2 seconds, Δt_max = 5 seconds. When the calculated interval is less than 2 seconds, it is set to 2 seconds; when it is greater than 5 seconds, it is set to 5 seconds. The lower limit of 2 seconds is to avoid overly dense sampling for extremely short videos, and the upper limit of 5 seconds is to ensure that even for very long videos, there will not be a frameless blind spot exceeding 5 seconds.

[0093] Frames are uniformly extracted from the start to the end of the video at constrained intervals Δt to obtain the third group of candidate frames. When extracting each catch-all frame, the time interval between it and the existing frames in the first, second, and fourth groups of candidate frames is checked in real time. If it is less than the time proximity threshold (1 second), the catch-all frame is discarded and not included in the third group of candidate frames to avoid data expansion during subsequent merging and deduplication.

[0094] 6. Perceptual hash deduplication and output module

[0095] The system receives the merged set of the four candidate frames mentioned above. First, it filters frames that are too close together based on temporal proximity. Then, it calculates the difference-aware hash (dHash) for each of the remaining frames, determines image similarity using Hamming distance, and removes visually highly similar redundant frames. Finally, it outputs standardized keyframes according to the playback sequence, synchronously recording the timestamp corresponding to each frame to generate a frame summary file. The details are as follows:

[0096] The candidate frames from the first, second, third, and fourth groups are merged into a unified candidate frame set based on their timestamps, and then a two-layer deduplication filter is performed:

[0097] The first layer is time proximity filtering. Candidate frames are sorted in ascending order of timestamp, and the time interval between adjacent frames is checked sequentially. If the interval is less than 1 second, the frame with the earlier timestamp is kept, and the frame with the later timestamp is discarded. This step is used to eliminate overlapping frames caused by different detection modules repeatedly triggering the same operation event.

[0098] The second layer is perceptual hashing for deduplication. For each frame retained after time filtering, a difference perceptual hash (dHash) feature value is calculated. The dHash calculation process is as follows: the image is scaled to a 9×8 pixel grayscale image. For each row, adjacent pixels are compared in grayscale. If the grayscale value of the left pixel is greater than that of the right pixel, it is recorded as 1; otherwise, it is recorded as 0, thus obtaining a 64-bit binary hash value. All candidate frames are arranged in ascending order of timestamp. For each frame, the Hamming distance (the number of different bits in two 64-bit binary numbers) is compared with each of the subsequent frames one by one. If the Hamming distance is less than 4, the two frames are considered visually highly similar, and the frame with the earlier timestamp is retained, while the frame with the later timestamp is discarded. This step-by-step comparison method can eliminate duplicate frames across time periods when the same business page appears multiple times before and after the video. The Hamming distance threshold of 4 is set based on the following: Under a 9×8 scaling size, the dHash deviation of the same page at different times due to minor differences in rendering anti-aliasing, cursor position, etc. is usually between 1 and 3 bits, while the dHash difference between different business pages is usually more than 10 bits. Taking 4 as the dividing line can remove nearly duplicate frames while avoiding the accidental deletion of frames with different content.

[0099] In the perceptual hashing deduplication process, static semantic discrimination is introduced: for a segment of video with a continuous static duration exceeding 10 seconds, if the segment contains business text information (determined by lightweight OCR to check for product names, prices, agreement terms, etc.), the deduplication Hamming distance threshold for that segment is relaxed from 4 to 6, retaining more valid business display frames; if the segment does not contain business text (blank page, loading page), the deduplication standard of Hamming distance 4 is strictly enforced. This mechanism resolves the confusion between valid static frames (merchants displaying their business on the page) and invalid static frames (blank pages caused by recording errors).

[0100] It should be noted that time proximity filtering is a prerequisite, which aims to reduce the computational complexity of subsequent dHash. If the order is reversed (hashing first and then time filtering), it will cause the hash value to be calculated repeatedly for adjacent frames of the same operation, increasing the overhead of ineffective computing power. Therefore, the order of the two-layer filtering is irreversible.

[0101] After deduplication, the retained frames are arranged in playback order and renamed in the format "frame_serial_timestamps.jpg" (e.g., frame_0001_3.250s.jpg). A frame summary file is generated simultaneously, recording the following information: total video duration, number of scene frames, number of motion supplement frames, number of fallback frames, number of deduplicated frames, final number of retained frames, and average time interval between adjacent frames.

[0102] In a preferred embodiment, the system may further include a visual recognition access point. This access point sends keyframe images and extracted frame summary files to a general multimodal visual model via two standardized output channels. The model then combines the screen content with temporal context information to automate compliance recognition of merchant screen recordings. This access point is an optional downstream application module and does not affect the independent operation of the frame extraction system itself. Details are as follows:

[0103] The final output set of keyframes and the extracted frame summary file can be fed into a general multimodal visual model (such as a vision-language model based on the Transformer architecture). The model combines the frame image content and the temporal information in the summary to automatically identify whether the merchant's business screen recording contains illegal content such as loan process displays, pages with annualized interest rates exceeding 24%, prohibited goods (virtual currency, recharge cards, antiques, precious metals and jewelry, etc.), and high-risk function entry points (recharge, transfer, wallet, earnings, promotion, etc.). This step is a downstream application of the frame extraction system. The frame extraction system itself does not depend on a specific visual model, nor does it require special fine-tuning of the model for merchant risk control scenarios.

[0104] Based on the above frame extraction system, the present invention also provides a corresponding intelligent video frame extraction method. Figure 2 The flowchart below shows the complete implementation steps of the intelligent video frame extraction method of the present invention, illustrating the execution order, conditional branches, and complementary relationship of the three-layer detection mechanism for each step from S1 to S6. The complete implementation steps are as follows:

[0105] Step S1, Video Parsing and Threshold Adaptation: Read the duration T, resolution, and frame rate f of the merchant's application review video. Match the scene detection threshold θ corresponding to the preset duration interval based on the duration (0.06 for T < 30s, 0.08 for 30s ≤ T ≤ 180s, and 0.12 for T > 180s). Simultaneously, take the first 5 seconds of the video (or all frames if the video duration is less than 5 seconds) and calculate the inter-frame noise variance σ as the device's baseline noise. The subsequent scene detection threshold is dynamically adjusted to θ' = θ + k·σ (k is a preset calibration coefficient, preferably k = 0.5). Calculate the target sampling frame number N_target, which is constrained by the upper and lower limits of the target sampling frame number.

[0106] Step S2, Scene Sudden Change Frame Extraction: Using the current scene detection threshold determined in Step S1, perform frame-by-frame scene sudden change detection on the video to identify sudden scene events such as page switching, pop-ups, and business process jumps; perform UI steady state verification on each detected sudden change candidate frame: check whether the changed area of ​​the frame remains pixel-stable in the next 3 consecutive frames (i.e., the inter-frame difference of the changed area is less than 0.01). If it is stable, it is confirmed as a valid scene frame and retained; if it is unstable, it is determined to be a transition animation or rendering artifact frame and is removed; form the first group of candidate frames with the retained valid scene frames; after the detection is completed, count the number of valid scene frames N_scene. If it is lower than the scene frame number trigger threshold, trigger step S3;

[0107] Step S2.a, Enhanced sampling of high-risk ROI areas: This step is executed in parallel with step S2, independently monitoring predefined sensitive ROI areas in the video. ROI areas are defined as the 50%~85% area from the top of the screen downwards (high-incidence pop-up area) and the 0%~10% area from the top of the screen downwards (status bar / interest rate display area). The pixel change rate within the ROI area is calculated frame by frame. When the pixel change rate of the ROI area exceeds twice the current global average inter-frame difference, a frame is forcibly extracted at that moment, without being limited by the global scene detection threshold. The time interval between the extracted ROI frame and existing scene frames is checked. If it is less than 0.3 seconds, it is considered an associated image of the same operation and is removed; otherwise, it is retained as the fourth group of candidate frames.

[0108] Step S3, Motion Supplement Frame Extraction: If N_scene < scene frame trigger threshold, perform a second scan of the video with a sensitivity threshold of 0.05 to capture subtle interface changes such as sliding and list scrolling, filter frames with a time interval of less than 0.5 seconds from the scene frames in the first group of candidate frames, and supplement the second group of candidate frames; if N_scene ≥ 10, skip this step; this step is a condition-triggered supplementary detection, not a necessary step, to avoid unnecessary second full scan of operation-intensive videos.

[0109] Step S4, Forced Cubble Frame Extraction: Independent of the results of steps S2, S2.a, and S3, a uniform sampling interval constrained by upper and lower limits of 2 to 5 seconds is dynamically calculated based on the target number of sampled frames N_target. Frames are extracted uniformly from beginning to end of the entire video. When extracting each catch-all frame, the time interval between it and existing frames in the first, second, and fourth candidate frames is checked in real time. If it is less than 1 second, the catch-all frame is discarded and not included in the third candidate frame set to avoid data expansion during subsequent merging and deduplication. The catch-all frames that pass the real-time rejection check form the third candidate frame set. Step S5, Merging, Filtering, and Deduplication: The first, second, third, and fourth candidate frames are merged into a unified set according to timestamps, sorted in ascending order of time, and filtered for time intervals less than 1 second. The process involves several steps: First, the neighboring frames (keeping the frame with the earlier timestamp and discarding the frame with the later timestamp) are selected. This step is a prerequisite to reduce the complexity of subsequent dHash calculations, and the order is irreversible. Then, the dHash value is calculated for each of the remaining frames, and visually similar frames are discarded according to the rule that the Hamming distance is less than 4 (keeping the frame with the earlier timestamp and discarding the frame with the later timestamp). In the perceptual hash deduplication stage, static semantic discrimination is introduced: for a segment of continuous static content exceeding 10 seconds, if the segment contains business text information, the deduplication Hamming distance threshold is relaxed to 6; otherwise, the standard threshold of 4 is applied. Step S6, frame output review: The retained frames are sorted and renamed according to the playback sequence, and a standardized keyframe set and a frame summary file containing the timestamps of each frame are output. If the number of retained frames after deduplication still exceeds the target sampling frame limit (i.e., N_max = 300 frames calculated in step S1), frames are uniformly extracted from the retained frame sequence at equal intervals to ensure that the final output frame count does not exceed the upper limit.

[0110] The present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the video intelligent frame extraction method for third-party payment merchant application review.

[0111] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the video intelligent frame extraction method for third-party payment merchant application review.

[0112] Example

[0113] To verify the effectiveness of the present invention, the following experiment was conducted.

[0114] Functional testing and verification

[0115] Functional verification was performed on each module of the present invention:

[0116] (1) Video parsing and adaptive module: Input test videos of different durations (10 seconds, 60 seconds, 300 seconds, 600 seconds) to verify the duration reading accuracy, threshold allocation correctness and target sampling frame number calculation results;

[0117] (2) Scene change detection module: Manually insert screen change events such as page switching and pop-up windows at known time points into the test video to verify the detection recall rate and timestamp positioning accuracy;

[0118] (3) High-risk area ROI enhanced sampling module: In the test video, instantaneous violation images of pop-up high-incidence areas and status bar areas are manually inserted to verify the accuracy of ROI area pixel change rate detection and the capture rate of the fourth group of candidate frames.

[0119] (4) Motion supplement detection module: Simulate page sliding, menu expansion and other operations in the test video to verify the capture effect of subtle screen changes and the accuracy of proximity filtering;

[0120] (5) Forced uniform bottom frame extraction module: Test the pure static video with no operation throughout and the short video with dense operation throughout to verify the integrity of the whole process coverage and the effectiveness of the interval constraint.

[0121] (6) Perceptual hash deduplication and output module: The input contains a mixed sequence of a large number of similar static frames and a small number of different content frames. The similar frame removal rate and the retention rate of different content frames are verified. The final number of frames is verified to be within the upper and lower limits, the naming format is standardized, and the fields of the summary file are complete and accurate.

[0122] Validation of the Implementation Effect of the End-to-End Engineering

[0123] To verify the technical effectiveness of this invention in a real-world business environment, the frame extraction processing logic was encapsulated as an independent service component. An end-to-end audit and verification link was established using a low-code workflow orchestration platform. The workflow sequentially connects the frame extraction component with a general multimodal visual model recognition node. The third-party payment merchant application system calls this workflow via a standard API. After completing frame extraction and compliance recognition, the workflow sends the structured risk identification results back to the business system. It should be noted that the aforementioned workflow orchestration platform is merely a verification carrier; the frame extraction processing logic of this invention does not depend on any specific scheduling platform or orchestration framework.

[0124] The controlled experiment was designed as follows: 216 samples were selected from merchant ingestion screen recordings accumulated from actual business operations, and divided into three groups according to video length: short recordings (T < 30s, 72 recordings), medium-long recordings (30s ≤ T ≤ 180s, 78 recordings), and long recordings (T > 180s, 66 recordings). The controlled experiment used fixed-interval frame sampling, with a uniform sampling interval of 3 seconds, without duration adaptation or deduplication. Test metrics included: keyframe capture rate based on manually annotated keyframe time points, the proportion of visually similar redundant frames in the output frames, the average image token consumption after being fed into the multimodal visual model, the model's single inference time, and the average time for manual review of a single video.

[0125] The test results are as follows:

[0126] (1) Short screen recording group: Due to the large fixed interval, the control scheme missed many subtle scenes with a duration of less than one second, such as sliding and pop-ups. The present invention uses a low threshold of 0.06 combined with motion supplement detection, which improves the key scene capture rate by about 35 percentage points compared with the control scheme (96.2% of the present invention and 61.4% of the control scheme); at the same time, because the deduplication process removes duplicate frames in the operation interval, the proportion of redundant frames decreases by about 75%, and the image token consumption fed into the model is reduced by about 68%.

[0127] (2) Medium-length screen recording group: The output frames of the control scheme contain a large number of repeated screenshots during static periods. The present invention uses three-layer detection combined with perceptual hashing to remove duplicates, achieving a visual similarity redundant frame removal rate of about 83%, and reducing the model's single inference time by about 71% compared to the control scheme.

[0128] (3) Long screen recording group: The control scheme generates a large number of frames in long videos due to the fixed interval of 3 seconds, most of which are static and repetitive images. The present invention uses a high threshold of 0.12 to suppress noise triggering, and with the help of bottom frame extraction to ensure full process coverage, the average time for manual review of a single video is reduced by about 57% compared with the control scheme.

[0129] It should be noted that for pure display screen recordings where there is no operation and the screen is completely still, neither the scene change detection nor the motion supplementation detection of this invention will trigger valid frame extraction. In this case, forced fallback frame extraction is mainly used to ensure basic coverage, and its effect is comparable to the fixed interval frame extraction scheme, without any significant advantage. The technical advantage of this invention is mainly reflected in typical input screen recording scenarios with intensive operations and uneven screen changes.

[0130] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0131] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A video intelligent frame extraction system for third-party payment merchant application review, characterized in that, include: The video information parsing and adaptive module is used to read the duration of the incoming review video, match the scene detection threshold corresponding to the preset duration range according to the video duration, and calculate the target sampling frame number constrained by the upper and lower limits of the target sampling frame number according to the video duration and the preset baseline frame rate. The scene change detection module is used to perform frame-by-frame scene change detection on the video based on the current scene detection threshold, identify page switching, pop-up windows and business process jump events, extract the corresponding timestamps to extract key frames to form the first group of candidate frames, and count the number of valid scene frames extracted after the detection is completed. When the number of valid scene frames is lower than the scene frame trigger threshold, a trigger signal is sent to the motion supplementation detection module. The high-risk area ROI enhanced sampling module is used to independently monitor predefined sensitive screen areas. When the pixel change rate of the ROI area exceeds the preset multiple of the global inter-frame difference average, high-frequency sampling is forcibly triggered and is not limited by the global scene detection threshold to form a fourth set of candidate frames. The motion supplement detection module is used to perform a second scan of the video with a sensitivity threshold lower than the scene detection threshold after receiving a trigger signal, to capture the frames corresponding to page swiping and menu scrolling events, and form a second set of candidate frames; The forced uniform bottom-feed frame extraction module is used to dynamically calculate the uniform sampling interval constrained by the shortest interval and the longest interval based on the target sampling frame number, regardless of whether the scene change detection module, the high-risk area ROI enhancement sampling module, and the motion supplementation detection module have extracted candidate frames. It performs uniform sampling on the entire video from beginning to end to form a third set of candidate frames. The perceptual hash deduplication and output module is used to merge the candidate frames of the first, second, third and fourth groups, filter frames that are too close in interval according to the time proximity rule, calculate the difference perceptual hash value for each frame and remove visually similar frames according to the Hamming distance, and output a standardized keyframe set and frame extraction summary file according to the playback sequence.

2. The video intelligent frame extraction system for third-party payment merchant application review according to claim 1, characterized in that, The video information parsing and adaptive module is also used to read the resolution and frame rate of the video, take the still image at the beginning of the video to calculate the inter-frame noise variance σ as the device reference noise, and dynamically adjust the scene detection threshold θ to θ'=θ+k·σ, where k is a preset calibration coefficient with a value of k=0.5; The preset duration intervals include: a first duration interval of less than 30 seconds, a second duration interval of greater than or equal to 30 seconds and less than or equal to 180 seconds, and a third duration interval of greater than 180 seconds; the scene detection threshold θ includes a first threshold of 0.06 corresponding to the first duration interval, a second threshold of 0.08 corresponding to the second duration interval, and a third threshold of 0.12 corresponding to the third duration interval; The target number of sampling frames is calculated based on a baseline density of 15 frames per minute, with an upper limit of 300 frames and a lower limit of 20 frames.

3. The video intelligent frame extraction system for third-party payment merchant application review according to claim 1, characterized in that, The scene mutation detection module performs UI steady-state verification. It takes the change area mask for the detected mutation candidate frame and calculates the inter-frame difference of the next three consecutive frames within the mask area. If all three are less than 0.01, it is confirmed as a valid scene frame; otherwise, it is determined as a transition animation frame and is removed. The scene frame trigger threshold for the scene mutation detection module is 10 frames.

4. The video intelligent frame extraction system for third-party payment merchant application review according to claim 1, characterized in that, The sensitive screen areas of the high-risk ROI enhancement sampling module include the pop-up high-occurrence area from 50% to 85% down from the top of the screen and the status bar area from 0% to 10% down from the top of the screen. The preset multiplier is 2. The time interval between the extracted ROI frame and the existing candidate frame is checked. If it is less than 0.3 seconds, it is discarded. Otherwise, a fourth group of candidate frames is formed. The sensitivity threshold of the motion supplement detection module is 0.05, and it filters frames with an interval of less than 0.5 seconds from existing scene frames.

5. The video intelligent frame extraction system for third-party payment merchant application review according to claim 1, characterized in that, During the sampling process, the forced uniform bottom-feed frame extraction module checks the time interval between each extracted frame and the first, second, or fourth group of candidate frames. If the time interval is less than the proximity threshold, the frame is discarded immediately, and the frames that pass the check form the third group of candidate frames. The minimum interval of the forced uniform bottom-out frame extraction module is 2 seconds, the maximum interval is 5 seconds, and the time proximity threshold is 1 second. When the calculated sampling interval is less than the shortest interval, the shortest interval value is used; when it is greater than the longest interval, the longest interval value is used.

6. The video intelligent frame extraction system for third-party payment merchant application review according to claim 1, characterized in that, When the perceptual hash deduplication and output module calculates the difference perceptual hash, it scales the image to a 9×8 pixel grayscale image and compares the grayscale values ​​of adjacent pixels in each row to generate a binary hash value. The perceptual hash deduplication and output module introduces static semantic discrimination for screen segments with a continuous static duration exceeding a preset threshold. If the screen contains business text information, the deduplication Hamming distance threshold is relaxed to 6; if it does not contain business text, the standard threshold of 4 is applied.

7. A video intelligent frame extraction method for third-party payment merchant application review, applied to the video intelligent frame extraction system for third-party payment merchant application review as described in any one of claims 1-6, characterized in that, The method includes the following steps: S1. Video parsing and threshold adaptation: Read the duration of the merchant's application review video, match the scene detection threshold corresponding to the preset duration range according to the duration, and calculate the target sampling frame number constrained by the upper and lower limits of the target sampling frame number according to the video duration and the preset baseline frame rate. S2, Scene Change Frame Extraction: Perform frame-by-frame scene change detection on the video using the current scene detection threshold, identify page switching, pop-up windows and business process jump events, extract the corresponding timestamps to extract key frames to form the first group of candidate frames; count the number of valid scene frames, if it is lower than the scene frame number trigger threshold, then trigger step S3. S2.a. Enhanced sampling of high-risk ROI areas: This is executed in parallel with step S2. Predefined sensitive screen areas are independently monitored. When the pixel change rate of the ROI area exceeds a preset multiple of the global inter-frame difference average, a frame is forcibly extracted. This is not limited by the global scene detection threshold and forms the fourth group of candidate frames. S3, Motion Supplement Frame Extraction: The video is scanned a second time with a sensitivity threshold lower than the scene detection threshold to capture the frames corresponding to page swiping and list scrolling events, forming a second set of candidate frames; S4. Forced bottom-line frame extraction: Independent of the results of steps S2, S2.a and S3, the uniform sampling interval is dynamically calculated based on the target number of sampled frames, and uniform sampling is performed on the entire video from beginning to end to form a third group of candidate frames. S5. Merging, filtering and deduplication: Merge the candidate frames of the first, second, third and fourth groups, filter frames that are too close in interval according to the time proximity rule, calculate the perceptual hash value of the difference between each frame and remove visually similar frames according to Hamming distance. S6. Review Frame Output: Output a standardized set of keyframes and a frame summary file containing timestamps of each frame according to the playback sequence.

8. The video intelligent frame extraction method for third-party payment merchant application review according to claim 7, characterized in that, In step S1, the preset duration interval includes a first duration interval of less than 30 seconds, a second duration interval of greater than or equal to 30 seconds and less than or equal to 180 seconds, and a third duration interval of greater than 180 seconds. The scene detection threshold includes a first threshold of 0.06 corresponding to the first duration interval, a second threshold of 0.08 corresponding to the second duration interval, and a third threshold of 0.12 corresponding to the third duration interval. The target number of sampling frames is calculated based on a baseline density of 15 frames per minute, with an upper limit of 300 frames and a lower limit of 20 frames.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the video intelligent frame extraction method for third-party payment merchant application review as described in claim 7 or 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the video intelligent frame extraction method for third-party payment merchant application review as described in claim 7 or 8.

Images