A teaching video PPT key frame extraction method and system

Abstract

The application discloses a kind of teaching video PPT key frame extraction method and system, the method includes: target video is filtered to scene, obtains PPT demonstration frame;PPT demonstration frame is detected to template page, and effective PPT demonstration frame is screened out;Effective PPT demonstration frame is selected dynamically to anchor point, determine the semantic anchor point of effective PPT demonstration frame, and according to the semantic anchor point of effective PPT demonstration frame, adjacent effective PPT demonstration frame is judged to asymmetric containing relationship, and the page relationship type is determined;According to the page relationship type of adjacent effective PPT demonstration frame, track page state transition, determine key PPT demonstration frame;The present application is with the semantic anchor point of PPT page as identity mark and carries out asymmetric containing relationship identification between continuous frame, and tracks page state transition, can output a main key frame from same page, effectively reduce the information redundancy of PPT key frame extraction, improve the accuracy of PPT key frame extraction.

Description

Technical Field

[0001] This invention relates to the field of video and image processing technology, and in particular to a method and system for extracting keyframes from teaching video PPTs. Background Technology

[0002] With the rapid development of educational informatization and blended learning, a large number of teaching videos are being recorded and stored. The PPT slides in these videos contain significant knowledge value. Therefore, extracting keyframes from the PPT slides in these teaching videos is particularly important for subsequent course indexing and intelligent retrieval of the video content.

[0003] Currently, the mainstream methods for extracting keyframes in PPT presentations mainly include fixed time interval sampling and inter-frame pixel difference methods. Fixed time interval sampling extracts a frame from the video at fixed time intervals. Although simple and easy to implement, it cannot distinguish between animation and page turning. Multiple redundant screenshots can be generated for the same PPT page due to animation effects, resulting in high information redundancy. Inter-frame pixel difference methods determine page changes by calculating the pixel differences between adjacent frames. Although sensitive to changes in lighting and animation effects, it relies solely on pixel-level visual features and lacks semantic understanding of the page content. It cannot identify pages with the same content, resulting in low accuracy in extracting keyframes from PPT presentations. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method and system for extracting keyframes from teaching video PPTs, which can effectively reduce the information redundancy in PPT keyframe extraction while improving the accuracy of PPT keyframe extraction.

[0005] In a first aspect, embodiments of the present invention provide a method for extracting keyframes from a teaching video PPT, including: Scene filtering is performed on the target video to obtain several PPT presentation frames from the target video; Template page detection is performed on each of the PPT presentation frames to filter out a number of valid PPT presentation frames; Dynamic anchor point selection is performed on each of the valid PPT presentation frames to determine the semantic anchor points in each of the valid PPT presentation frames. Based on the semantic anchor points in each of the valid PPT presentation frames, an asymmetric inclusion relationship determination is performed on adjacent valid PPT presentation frames to determine the page relationship type of adjacent valid PPT presentation frames. Based on the page relationship type of adjacent valid PPT presentation frames, track page state changes to determine key PPT presentation frames.

[0006] As an improvement to the above solution, scene filtering is performed on the target video to obtain several PPT presentation frames in the target video, including: For each frame of the target video, a pre-trained image classification model is used to classify the scene and determine the scene category and confidence level of each frame. For images categorized as PPT presentation frames, images with a confidence level higher than a preset confidence threshold are output as the first candidate PPT presentation frames. For images with a confidence level lower than the confidence threshold, the OCR layout analysis model is used to match and identify screen elements to select the first candidate PPT presentation frame. Boundary scene verification is performed on the first candidate PPT presentation frame, and the first candidate PPT presentation frame that passes the boundary scene verification is selected as the PPT presentation frame.

[0007] As an improvement to the above solution, template page detection is performed on each of the PPT presentation frames to filter out several valid PPT presentation frames, including: Each of the aforementioned PPT presentation frames is divided into several screen areas; Based on the histogram difference between corresponding screen areas in two adjacent PPT presentation frames, the PPT presentation frames are filtered to obtain a second candidate PPT presentation frame. Divide the PPT pages in the second candidate PPT presentation frame into several page areas; Element recognition is performed on each page area in the second candidate PPT presentation frame to determine whether the PPT page in the second candidate PPT presentation frame is a PPT template page; If so, skip the second candidate PPT presentation frame; If not, output the second candidate PPT presentation frame as the valid PPT presentation frame.

[0008] As an improvement to the above scheme, dynamic anchor point selection is performed on each of the valid PPT presentation frames to determine the semantic anchor points in each of the valid PPT presentation frames, including: According to the preset anchor point selection priority rules, select the semantic anchor points in each of the effective PPT presentation frames; The semantic anchor points include the title on the PPT page, the main visual element with the largest area on the PPT page, and the main text area with the most text on the PPT page; the selection priority of the title is greater than the selection priority of the main visual element, and the selection priority of the main visual element is greater than the selection priority of the main text area.

[0009] As an improvement to the above scheme, based on the semantic anchor points in each of the valid PPT presentation frames, an asymmetric inclusion relationship determination is performed on adjacent valid PPT presentation frames to determine the page relationship type of adjacent valid PPT presentation frames, including: Determine whether the similarity between the semantic anchor point of the current valid PPT presentation frame and the semantic anchor point of the previous valid PPT presentation frame is greater than a preset similarity threshold. If so, determine that the PPT pages in the current valid PPT presentation frame and the previous valid PPT presentation frame are the same page, and determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame based on the element retention rate, new element ratio and area stability of the current valid PPT presentation frame. If not, determine the content evolution of the currently valid PPT presentation frame.

[0010] As an improvement to the above solution, the content evolution of the currently valid PPT presentation frames is determined, including: When determining whether the similarity between the semantic anchor point of the next valid PPT presentation frame and the semantic anchor point of the current valid PPT presentation frame is not greater than the similarity threshold; If so, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame as page turning; If not, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame based on the element retention rate, new element ratio, and area stability of the current valid PPT presentation frame.

[0011] As an improvement to the above scheme, based on the element retention rate, new element ratio, and area stability of the current valid PPT presentation frame, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined, including: The element retention rate of the current valid PPT presentation frame is calculated based on the first element that matches the elements of the current valid PPT presentation frame and the total elements of the previous valid PPT presentation frame. Calculate the proportion of new elements in the current valid PPT presentation frame based on the second element in the current valid PPT presentation frame that does not match the element in the previous valid PPT presentation frame, and the total number of elements in the current valid PPT presentation frame. The area stability of the current valid PPT presentation frame is calculated based on the area ratio of the matched element in the current valid PPT presentation frame and the total area of ​​all elements in the previous valid PPT presentation frame. The element retention rate, the proportion of new elements, and the area stability are weighted and summed to calculate the inclusion score of the current valid PPT presentation frame; Based on the included score, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame.

[0012] As an improvement to the above scheme, based on the included scoring, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined, including: When the included score is greater than the preset first score threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be animation evolution; When the score is less than the preset second score threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be page turning; When the included score is less than or equal to the first score threshold and greater than or equal to the second score threshold, the layout change parameters between the current valid PPT presentation frame and the previous valid PPT presentation frame are calculated based on the third element in the previous valid PPT presentation frame that does not match the element of the current valid PPT presentation frame and the total elements of the previous valid PPT presentation frame. Based on the KL divergence between the element type distribution of the current valid PPT presentation frame and the element type distribution of the previous valid PPT presentation frame, calculate the parameter of change in element type distribution between the current valid PPT presentation frame and the previous valid PPT presentation frame. When the layout change parameter between the current valid PPT presentation frame and the previous valid PPT presentation frame is greater than the preset first parameter threshold or the element type distribution change parameter is greater than the preset second parameter threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be page turning. If the layout change parameter between the current valid PPT presentation frame and the previous valid PPT presentation frame is not greater than the first parameter threshold and the element type distribution change parameter is not greater than the second parameter threshold, then the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be animation evolution.

[0013] As an improvement to the above scheme, the key PPT presentation frames are determined by tracking page state changes based on the page relationship type of adjacent valid PPT presentation frames, including: For a valid PPT presentation frame with a page-turning relationship type, the current valid PPT presentation frame is used as the initial page frame of the current PPT page, and the page state change sequence of the current PPT page is initialized; wherein, the previous valid PPT presentation frame is used as the end page frame of the page state change sequence of the previous PPT page. For valid PPT presentation frames with page relationship type of animation evolution, add them to the page state change sequence of the corresponding PPT page; For the valid PPT presentation frames in the page state change sequence, select the valid PPT presentation frame with the largest amount of information and output it as the key PPT presentation frame of the corresponding PPT page; The valid PPT presentation frames in the page state change sequence are fused to generate a fused PPT presentation frame, and the fused PPT presentation frame is output as the key PPT presentation frame of the corresponding PPT page.

[0014] Secondly, embodiments of the present invention provide a keyframe extraction system for teaching video PPTs, comprising: The scene filtering module is used to perform scene filtering on the target video to obtain several PPT presentation frames in the target video. The template page detection module is used to perform template page detection on each of the PPT presentation frames and filter out a number of valid PPT presentation frames. The dynamic anchor point selection module is used to dynamically select anchor points for each of the valid PPT presentation frames and determine the semantic anchor points in each of the valid PPT presentation frames. The asymmetric inclusion relationship determination module is used to determine the asymmetric inclusion relationship between adjacent valid PPT presentation frames based on the semantic anchor points in each of the valid PPT presentation frames, and to determine the page relationship type of adjacent valid PPT presentation frames. The page state tracking module is used to track page state changes based on the page relationship type of adjacent valid PPT presentation frames, and to identify key PPT presentation frames.

[0015] Compared to existing technologies, this invention provides a method and system for extracting keyframes from teaching video PPTs. The method involves scene filtering of a target video to obtain several PPT presentation frames; template page detection of each PPT presentation frame to select several valid PPT presentation frames; dynamic anchor point selection for each valid PPT presentation frame to determine its semantic anchor points; and asymmetric inclusion relationship determination of adjacent valid PPT presentation frames based on their semantic anchor points to determine their page relationship type. Finally, tracking page state changes based on the page relationship type of adjacent valid PPT presentation frames helps identify keyframes. This invention, by identifying asymmetric inclusion relationships between consecutive frames using the semantic anchor points of PPT pages as identifiers and tracking PPT page state changes, can output keyframes from the same PPT page, effectively reducing information redundancy in keyframe extraction and improving its accuracy. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a method for extracting keyframes from a teaching video PPT provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the nine-square grid division provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of state transitions provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the overall PPT keyframe extraction method provided in this embodiment of the invention; Figure 5 This is a structural block diagram of a teaching video PPT keyframe extraction system provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It is understood that the various numerical designations used in the embodiments of this invention are merely for descriptive convenience and are not intended to limit the scope of this application. The order of the process numbers does not imply the order of execution; the execution order of each process should be determined by its function and internal logic.

[0020] In embodiments of the invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, without necessarily requiring or implying any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising" does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element. The term "a plurality or several" refers to two or more.

[0021] See Figure 1 , Figure 1 This is a flowchart illustrating a method for extracting keyframes from a teaching video PPT provided in an embodiment of the present invention. The method for extracting keyframes from a teaching video PPT specifically includes: S11: Perform scene filtering on the target video to obtain several PPT presentation frames in the target video; S12: Perform template page detection on each of the PPT presentation frames and filter out a number of valid PPT presentation frames; S13: Dynamically select anchor points for each of the valid PPT presentation frames to determine the semantic anchor points in each of the valid PPT presentation frames; S14: Based on the semantic anchor points in each of the valid PPT presentation frames, perform asymmetric inclusion relationship determination on adjacent valid PPT presentation frames to determine the page relationship type of adjacent valid PPT presentation frames; S15: Track page state changes based on the page relationship type of adjacent valid PPT presentation frames to determine key PPT presentation frames.

[0022] The target video can be a recorded teaching video containing PPT presentations, an online teaching video, a courseware explanation video, a work report video, etc., and is not limited in this embodiment of the invention.

[0023] It should be noted that the method for extracting keyframes from teaching video PPTs described in this embodiment of the invention can be executed by terminal devices such as servers and computers. Specifically, the target video is first filtered to remove noisy content, such as camera frames containing only live narration, excessively black screen frames, blank frames, and other image frames without PPT pages. This initially filters out image frames containing PPT pages as PPT presentation frames. Then, template page detection is performed on the initially filtered PPT presentation frames to filter out template pages or blank pages with no valid content, and select valid PPT presentation frames containing valid content. After that, dynamic anchor point selection is performed on the selected valid PPT presentation frames to select semantic anchor points in each valid PPT presentation frame. Based on the semantic anchor points in each valid PPT presentation frame, the asymmetric inclusion relationship between adjacent frames is determined to determine the page relationship type between adjacent frames. The PPT page state changes are tracked, and the corresponding key PPT presentation frames are output for the same PPT page. This can effectively reduce the information redundancy of PPT keyframe extraction and improve the accuracy of PPT keyframe extraction.

[0024] In one optional embodiment, scene filtering is performed on the target video to obtain several PPT presentation frames in the target video, including: For each frame of the target video, a pre-trained image classification model is used to classify the scene and determine the scene category and confidence level of each frame. For images categorized as PPT presentation frames, images with a confidence level higher than a preset confidence threshold are output as the first candidate PPT presentation frames. For images with a confidence level lower than the confidence threshold, the OCR layout analysis model is used to match and identify screen elements to select the first candidate PPT presentation frame. Boundary scene verification is performed on the first candidate PPT presentation frame, and the first candidate PPT presentation frame that passes the boundary scene verification is selected as the PPT presentation frame.

[0025] It should be noted that the embodiments of the present invention do not limit the network structure of the image classification model. For example, image convolutional neural networks such as ResNet-18 and MobileNet-V3 can be used to construct the image classification model. Image classification models are existing technologies and will not be described in detail here. For example, training samples with pre-labeled image categories can be used to iteratively train the image classification model until the model converges to obtain a trained image classification model.

[0026] The image categories include the following: PPT presentation screen: refers to the normal PPT presentation mode, where the image screen is recognized as full-screen PPT content without any editing interface elements; the PPT presentation screen belongs to the normal PPT frame, and subsequent three-level judgments are performed, such as change trigger judgment, semantic inclusion judgment, and structural auxiliary judgment, to participate in keyframe extraction. PPT editing mode screen: refers to the screen in the PPT editing interface where the toolbar, menu bar, and slide thumbnails are visible; the PPT editing mode screen does not participate in keyframe extraction. Desktop screen: The desktop, browser, and other applications are identified in the image screen, which are not PPT feature elements; the desktop content is meaningless and is not included in keyframe extraction. Player video frame: Embedded video segments are identified in the image frame, such as the video player UI or dynamic video content; the player video frame is not part of the PPT content and is not included in keyframe extraction; Word document screenshot: The image screen recognizes non-PPT content such as Word and PDF documents, such as the Word menu bar and document layout features; Word and PDF documents are not part of PPT and are not included in keyframe extraction. Black and white screen: refers to a transitional black screen or blank frame, such as when the overall brightness of the image is low and the elements are empty; a black and white screen has no content and is not included in keyframe extraction; Camera footage: refers to footage of a live presenter, such as when facial features are recognized but no PPT elements are identified; camera footage of the teacher's presentation is not part of the PPT content and is not included in keyframe extraction.

[0027] In this embodiment of the invention, each frame of the target video is sequentially output to a pre-trained image classification model for scene classification, and the scene category and confidence level of each frame are identified. If the confidence level of the image category to which a certain frame belongs is higher than the preset confidence level threshold, such as 0.7, then the final image category of that frame is directly determined to be the image category output by the image classification model, and the image with the image category of PPT presentation screen is selected and output as the first candidate PPT presentation frame; images corresponding to other image categories are determined to be non-PPT presentation screens and are directly skipped. In this embodiment of the invention, the PPT editing screen is not included in keyframe extraction; similarly, Word documents and embedded videos are not part of the PPT content; the camera footage is part of the teacher's presentation and is also not part of the PPT content; and the black and white screen has no content and the desktop content is meaningless. Therefore, images that are not part of the PPT presentation screen, such as the PPT editing screen, desktop screen, Word document screen, player video screen, black and white screen screen, and camera footage, are skipped and not included in the subsequent OCR layout analysis.

[0028] If the confidence level of the image category to which a certain frame belongs is not higher than 0.7, then the image frame is further output to the OCR layout analysis model for image element matching and recognition; the image element matching and recognition process is as follows: First, perform OCR recognition on the image frame to identify the text elements in the image frame and the coordinates of the bounding boxes of each text element; If the image frame does not contain text elements, the image frame is determined to be a black and white screen; since a black and white screen has no content, the image corresponding to a black and white screen is skipped. If the image frame contains text elements, a toolbar area is defined within the image frame, for example, the top 10% of the image height. Each text element is then iterated through, and those with bounding box coordinates falling within this toolbar area are selected. The selected text elements are then matched against PowerPoint-specific editing toolbar keywords, such as "Start," "Insert," "Design," "Transitions," "Animations," "Slides," "View," "Review," and "Formatting." If any text element matches any of these keywords, the image frame is determined to be a PowerPoint editing screen and is skipped. Otherwise, it is further determined whether the image frame is a Word document screen. Furthermore, considering that the recording process may briefly switch to editing mode, i.e., modifying the PPT content, in this embodiment of the invention, when the current frame image is identified as a PPT editing mode image, it is determined whether the current frame image is a single frame editing mode, i.e., the next frame image is not a PPT editing mode image. If so, the current frame image is skipped directly; if not, it is further determined whether the N consecutive frames starting with the current frame image are all PPT editing mode images. If so, it indicates that the PPT is being modified, and the subsequent images of the target video are sampled according to a preset time interval until the next frame image that is not a PPT editing mode image is sampled. All images from the current frame image of the PPT editing mode image to the next frame image that is not a PPT editing mode image are skipped, and the subsequent recognition process continues starting with the next frame image that is not a PPT editing mode image.

[0029] All identified text elements are matched with keywords from Word's exclusive menu bar, such as File, Home, References, Mail, Revisions, Comments, Header, and Footer. If any text element matches any of the above keywords, it is determined whether the frame image is a Word document and is skipped; otherwise, it is further determined whether the frame image is a video player screen. All identified text elements are matched with player-specific control text, such as play, pause, progress, full screen, speed up, bullet comments, volume, episode selection, etc. If any text element matches the above text, the image frame is determined to be a player video and is skipped directly; otherwise, it is further determined whether the image frame is a camera image. Perform face detection on the frame image and identify the face region; if a face region is identified, determine that the frame image is a camera view and skip it directly; otherwise, treat the frame image as a PPT presentation screen. The image corresponding to the PPT presentation screen identified above is output as the first candidate PPT presentation frame.

[0030] Considering that video playback may be embedded within PPT pages in lecture videos, and some lecture videos use split-screen mode, such as PPT + teacher's camera, misjudgment of PPT presentation screens may occur. Therefore, this embodiment of the invention further performs boundary scene verification on the first candidate PPT presentation frames output by the above image classification model and OCR layout analysis model that belong to the PPT presentation screen, as follows: For a mixed PPT and video scenario where video playback may be embedded within a PPT page, calculate the first area ratio of the video player in the first candidate PPT presentation frame. If the first area proportion is greater than the preset area proportion threshold, it means that the embedded video in the picture is the main element and fails the boundary scene verification. Therefore, the first candidate PPT presentation frame is determined to be not a PPT presentation picture and is skipped. If the first area proportion is not greater than the preset area proportion threshold, it means that the PPT in the picture is the main element. Through boundary scene verification, it is determined that the first candidate PPT presentation frame is the PPT presentation picture, and the first candidate PPT presentation frame is output as the final PPT presentation frame. For some teaching videos that use split-screen mode with the camera and PPT side by side, calculate the second area ratio of the PPT page in the first candidate PPT presentation frame. If the second area ratio is greater than the preset area ratio threshold, it means that the PPT in the picture is the main element. Through boundary scene verification, it is determined that the first candidate PPT presentation frame is the PPT presentation picture, and the first candidate PPT presentation frame is output as the final PPT presentation frame. If the first area proportion is not greater than the preset area proportion threshold, it means that the teacher video in the picture is the main element and fails the boundary scene verification. Therefore, the first candidate PPT presentation frame is determined to be not a PPT presentation picture and is skipped.

[0031] Since lecture videos often contain non-PPT content, such as live footage of the teacher, PPT editing interfaces, and embedded video clips, this invention combines image classification and OCR layout analysis for image category classification. This approach balances single-frame inference speed with the reliability of image subject recognition. By relying on visual features to drive the judgment and supplementing it with textual information, it effectively mitigates recognition bias caused by visual blur while ensuring processing latency, thus improving the accuracy of image classification. Based on the image classification results, it can filter out noisy content in the lecture video, such as PPT editing mode, desktop, video, and camera footage, ensuring that subsequent judgments are only performed on PPT presentation frames. This avoids noisy frames interfering with the accuracy of key frame extraction and reduces unnecessary computational overhead.

[0032] Furthermore, based on OCR layout analysis, this embodiment of the invention defines a unified frame data structure, including: title field and element field; The title field includes: the title text content; the title bounding box coordinates, such as (x1, y1, x2, y2), representing the coordinates of the four vertex corners of the title bounding box; and the OCR confidence score corresponding to the title, which is the confidence score of the title text output by OCR recognition. The element fields include: element type, such as text, table, formula, illustration; element bounding box coordinates; element text, if it is not a text element, the element text is empty; and the OCR confidence score corresponding to the element, which is the confidence score of the element text output based on OCR recognition.

[0033] This invention, through a standardized frame data structure, can complete the layout structure parsing of a single PPT presentation frame, uniformly carrying the identification of the position and content of all titles and multiple types of elements on the PPT page, providing a standardized data source for subsequent keyframe extraction, and improving processing efficiency.

[0034] In one optional embodiment, template page detection is performed on each of the PPT presentation frames to filter out a number of valid PPT presentation frames, including: Each of the aforementioned PPT presentation frames is divided into several screen areas; Based on the histogram difference between corresponding screen areas in two adjacent PPT presentation frames, the PPT presentation frames are filtered to obtain a second candidate PPT presentation frame. Divide the PPT pages in the second candidate PPT presentation frame into several page areas; Element recognition is performed on each page area in the second candidate PPT presentation frame to determine whether the PPT page in the second candidate PPT presentation frame is a PPT template page; If so, skip the second candidate PPT presentation frame; If not, output the second candidate PPT presentation frame as the valid PPT presentation frame.

[0035] For example, firstly, each of the initially selected PPT presentation frames undergoes a change trigger determination. Specifically, the PPT presentation frame is divided into a nine-grid layout, creating a 3×3 screen area. From left to right and top to bottom, this includes nine screen areas: R1, R2, R3, R4, R5, R6, R7, R8, and R9. Figure 2 As shown in the figure. Among them, R1, R2, and R3 are the top regions, R4, R5, and R6 are the middle regions, and R7, R8, and R9 are the bottom regions.

[0036] For the current PPT presentation frame and the previous PPT presentation frame, calculate the HSV histogram difference for each screen region, as follows: (1); in, This represents the chi-square distance calculation function. This indicates the current PPT presentation frame number. HSV histogram of the image area This indicates the previous PPT presentation frame number. HSV histogram of the image area This indicates the current PPT presentation frame number. The screen area is the same as the previous PPT presentation frame. The HSV histogram differences of the screen regions, i=1, 2, ..., 9, correspond to the 9 screen regions R1, R2, ..., R9 respectively, and t represents time.

[0037] We calculate the total histogram difference D between the current PPT presentation frame and the previous PPT presentation frame by weighted summation of all screen areas. The specific calculation is as follows: (2); Where D represents the total histogram difference between the current PPT presentation frame and the previous PPT presentation frame; This represents the first frame of the PPT presentation. Weight of the screen area.

[0038] For example, the top areas R1, R2, and R3, where titles are usually located, are sensitive to changes: set a higher weight, for example, a weight of 0.25 for the top areas; the middle areas R4, R5, and R6, which are the main content areas: set a medium weight, for example, a weight of 0.15 for the middle areas; the bottom areas R7, R8, and R9, which are often notes, blank areas, or animations: set a lower weight, for example, a weight of 0.05 for the bottom areas.

[0039] Then, a threshold judgment is made on the calculated histogram difference D. If the current PPT presentation frame is selected, it will be output as the second candidate PPT presentation frame, and the process will proceed to the subsequent semantic inclusion determination stage; otherwise, the current PPT presentation frame will be ignored and the system will remain silent. This represents the preset histogram difference threshold, for example... .

[0040] In the semantic inclusion determination stage, considering that a PPT page may only have content in the corners, such as page numbers, logos, or decorative elements, these pages are usually template pages or blank pages with no valid information and require further filtering. For the above-mentioned... The current PPT presentation frame is first divided into 9 page areas. For example, the width and height of the top corner page area can be divided into 9 page areas: P1, P2, P3, P4, P5, P6, P7, P8, and P9.

[0041] Then, the valid content of the nine page areas P1-P9 is determined. Specifically, based on the frame structure data of the PPT presentation frame, the coordinates of the element bounding box of each element are obtained, and the coordinates of the center point of the element are calculated. Then, it is determined whether the center point coordinates of all elements in the PPT presentation frame fall in the four corner areas: the top left page area P1, the top right page area P3, the bottom left page area P7, and the bottom right page area P9. If so, it means that there is only decorative content such as watermarks, page numbers, and logos in the corners, and there is no main text information on the page. The PPT presentation frame is determined to have no valid content and is a PPT template page. The PPT presentation frame is skipped and kept silent. Otherwise, the PPT presentation frame is output as a valid PPT presentation frame.

[0042] This invention, through frame-to-frame histogram difference analysis and page element position identification, can quickly filter frames with no obvious changes and blank template pages containing only page numbers, logos, and corner decorations, thereby reducing subsequent computational overhead and improving overall processing efficiency.

[0043] In an optional embodiment, dynamic anchor point selection is performed on each of the valid PPT presentation frames to determine the semantic anchor points in each of the valid PPT presentation frames, including: According to the preset anchor point selection priority rules, select the semantic anchor points in each of the effective PPT presentation frames; The semantic anchor points include the title on the PPT page, the main visual element with the largest area on the PPT page, and the main text area with the most text on the PPT page; the selection priority of the title is greater than the selection priority of the main visual element, and the selection priority of the main visual element is greater than the selection priority of the main text area.

[0044] In this embodiment of the invention, the priority rules for anchor point selection are as follows: when a title exists, the title is given priority as the semantic anchor point; when there is no title, the main visual element corresponding to the largest area, such as a table element or an illustration element, is used as the semantic anchor point; when there is no title and no obvious visual element with the largest area, the main text area corresponding to the largest amount of text, i.e., the text element, is used as the semantic anchor point.

[0045] Specifically, if the title field in the frame structure data corresponding to a valid PPT presentation frame is empty or the corresponding title OCR confidence is lower than a set threshold, it is determined to be a no-title scenario.

[0046] For scenarios without a title, invalid decorative elements at the corners of the PPT page are removed. For example, elements located in the four top-left corner areas P1, P3, P7, and P9 are removed. Then, the area of ​​the bounding box corresponding to each remaining element is calculated. The element with the largest area (greater than a preset area threshold) is selected as the main visual element for matching semantic anchors. If no element with an area greater than the preset area threshold exists, the scenario is determined to be without a clearly defined maximum area visual element.

[0047] By using main visual elements to match semantic anchors, it can adapt to scenarios where PPT pages contain large images, large tables, and large charts. It can use the main block with the highest visual proportion as a benchmark to quickly anchor the main content of the page.

[0048] For scenarios without titles and without obvious maximum visual elements, only elements of type text are selected, i.e., text elements. The length of the characters in each text element is counted, and the area containing the text element with the longest character length is selected as the main text area for matching semantic anchors.

[0049] By using semantic anchor points matched with the main text area, each valid PPT presentation frame has a clear page identity regardless of whether it has a title. At the same time, it can automatically select the largest and most prominent element as the anchor point based on the page content, improving the coverage and accuracy of page recognition, and making the selected anchor points highly resistant to interference and stable.

[0050] In one optional embodiment, based on the semantic anchor points in each of the valid PPT presentation frames, an asymmetric inclusion relationship determination is performed on adjacent valid PPT presentation frames to determine the page relationship type of adjacent valid PPT presentation frames, including: Determine whether the similarity between the semantic anchor point of the current valid PPT presentation frame and the semantic anchor point of the previous valid PPT presentation frame is greater than a preset similarity threshold. If so, determine that the PPT pages in the current valid PPT presentation frame and the previous valid PPT presentation frame are the same page, and determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame based on the element retention rate, new element ratio and area stability of the current valid PPT presentation frame. If not, determine the content evolution of the currently valid PPT presentation frame.

[0051] Specifically, the content evolution determination for the currently valid PPT presentation frame includes: When determining whether the similarity between the semantic anchor point of the next valid PPT presentation frame and the semantic anchor point of the current valid PPT presentation frame is not greater than the similarity threshold; If so, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame as page turning; If not, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame based on the element retention rate, new element ratio, and area stability of the current valid PPT presentation frame.

[0052] In one optional embodiment, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined based on the element retention rate, new element ratio, and area stability of the current valid PPT presentation frame, including: The element retention rate of the current valid PPT presentation frame is calculated based on the first element that matches the elements of the current valid PPT presentation frame and the total elements of the previous valid PPT presentation frame. Calculate the proportion of new elements in the current valid PPT presentation frame based on the second element in the current valid PPT presentation frame that does not match the element in the previous valid PPT presentation frame, and the total number of elements in the current valid PPT presentation frame. The area stability of the current valid PPT presentation frame is calculated based on the area ratio of the matched element in the current valid PPT presentation frame and the total area of ​​all elements in the previous valid PPT presentation frame. The element retention rate, the proportion of new elements, and the area stability are weighted and summed to calculate the inclusion score of the current valid PPT presentation frame; Based on the included score, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame.

[0053] In an optional embodiment, determining the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame based on the included score includes: When the included score is greater than the preset first score threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be animation evolution; When the score is less than the preset second score threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be page turning; When the included score is less than or equal to the first score threshold and greater than or equal to the second score threshold, the layout change parameters between the current valid PPT presentation frame and the previous valid PPT presentation frame are calculated based on the third element in the previous valid PPT presentation frame that does not match the element of the current valid PPT presentation frame and the total elements of the previous valid PPT presentation frame. Based on the KL divergence between the element type distribution of the current valid PPT presentation frame and the element type distribution of the previous valid PPT presentation frame, calculate the parameter of change in element type distribution between the current valid PPT presentation frame and the previous valid PPT presentation frame. When the layout change parameter between the current valid PPT presentation frame and the previous valid PPT presentation frame is greater than the preset first parameter threshold or the element type distribution change parameter is greater than the preset second parameter threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be page turning. If the layout change parameter between the current valid PPT presentation frame and the previous valid PPT presentation frame is not greater than the first parameter threshold and the element type distribution change parameter is not greater than the second parameter threshold, then the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be animation evolution.

[0054] In this embodiment of the invention, after selecting the semantic anchor point of a valid PPT presentation frame, the selected semantic anchor point is used as the page identity identifier of the valid PPT presentation frame to determine the asymmetric inclusion relationship between frames, so as to determine whether the current valid PPT presentation frame is a page turn or an animation evolution relative to the previous valid PPT presentation frame. The specific asymmetric inclusion relationship determination process is as follows: First, calculate the similarity between the semantic anchor point of the current valid PPT presentation frame and the semantic anchor point of the previous valid PPT presentation frame.

[0055] For scenarios where the title serves as the semantic anchor, the text similarity of the title is calculated using edit distance, as follows: (3); in, Indicates the text similarity of the titles; Indicates the edit distance function; The title text content of the currently active PPT presentation frame; This indicates the title text content of the previous valid PPT presentation frame; These represent the character lengths of the title text content of the current valid PPT presentation frame and the previous valid PPT presentation frame, respectively; max represents the maximum value function.

[0056] The positional similarity of the headings is calculated using IoU (Intersection over Union), as follows: (4); in, Indicates the similarity in the position of the titles. Represents the intersection-union ratio function; This represents the area of ​​the title bounding box of the currently active PPT presentation frame; This represents the area corresponding to the title bounding box of the previous valid PPT presentation frame.

[0057] Text similarity of titles and positional similarity The final title similarity score is obtained by performing a weighted summation. Similarity between semantic anchors.

[0058] (5); in, Represents the weighting coefficient, for example .

[0059] If title similarity Less than the preset similarity threshold, such as If the condition is met, it is determined to be a page turn, and the page turn confirmation state is entered directly for page settlement; otherwise, the current valid PPT presentation frame and the previous valid PPT presentation frame are determined to be on the same page, and the layout analysis stage is entered.

[0060] For scenarios where main visual elements, such as images and tables, serve as semantic anchors, the similarity of main visual elements is calculated using IoU as the similarity between semantic anchors. The specific calculation is described in formula (4) above and will not be repeated here. Similarly, if the similarity of main visual elements is less than the preset similarity threshold, it is determined to be a page turn and directly enters the page turn confirmation state for page settlement; otherwise, the current valid PPT presentation frame and the previous valid PPT presentation frame are determined to be on the same page, and the layout analysis stage is entered.

[0061] For scenarios where the main text region is used as the semantic anchor point, the text similarity of the main text region is calculated using edit distance, and the positional similarity of the main text region is calculated using IoU. Then, the weighted sum is used to calculate the total similarity of the main text region, which is used as the similarity between semantic anchor points. For specific calculations, please refer to formulas (3), (4), and (5) above, which will not be repeated here. Similarly, if the similarity of the main text region is less than the preset similarity threshold, it is determined to be a page turn, and the page turn confirmation state is directly entered for page settlement; otherwise, the current valid PPT presentation frame and the previous valid PPT presentation frame are determined to be on the same page, and the layout analysis stage is entered.

[0062] To eliminate pixel-level drift in OCR recognition, the bounding box coordinates of elements are discretized: (6); in, Indicates the floor function; Indicates the grid size, for example ; Represents the discretized coordinates. Represents the coordinates before discretization.

[0063] By converting continuous coordinates to grid coordinates, spatial tolerance can be achieved, eliminating the impact of recognition drift of 1-2 pixels. A smaller g-value improves matching accuracy and is suitable for handling animation scenes with small element displacements.

[0064] Next, element matching is performed to establish the correspondence between elements in consecutive frames, setting the following two matching conditions: (7); (8); in, The element representing the previous valid PowerPoint presentation frame. Elements of the currently active PowerPoint presentation frame The intersection-union ratio of the bounding boxes of the elements; Represents the Euclidean distance between the center points of the elements; This represents the center distance threshold.

[0065] When any one of the above matching conditions is met, then the element With elements Matching is considered as the same element being represented in consecutive frames.

[0066] Animation progression: Indicates that most of the content from the previous frame is retained in the next frame, such as when content is added. Page turning: Indicates that most of the content from the previous frame disappears, such as when content is replaced. To determine the asymmetric inclusion relationship between consecutive frames and to judge whether the consecutive frames are animation progression or page turning, this embodiment of the invention introduces element retention rate, new element ratio, and area stability to calculate the inclusion score of consecutive frames, which is used to determine whether the consecutive frames are animation progression or page turning.

[0067] Element retention rate : Calculate how many elements from the previous valid PPT presentation frame are retained in the current valid PPT presentation frame; (9); in, This indicates the number of elements in the previous valid PPT presentation frame that can be matched with the current valid PPT presentation frame. For specific element matching methods, please refer to the above description. This indicates the total number of elements in the previous valid PPT presentation frame.

[0068] New element ratio : Calculate how many of the currently active PPT presentation frames are newly appearing elements; (10); in, This indicates the number of elements in the current valid PPT presentation frame that cannot be matched with the previous valid PPT presentation frame; This indicates the total number of elements in the currently active PPT presentation frame.

[0069] Area stability : Calculate how much of the element area from the previous valid PPT presentation frame was retained; (11); in, Indicates the area of ​​the matched elements; : The total area of ​​elements in the previous valid PPT presentation frame.

[0070] The element retention rate, the proportion of new elements, and the area stability are weighted and summed to calculate the score. The details are as follows: (12); in, , , These represent the retention rate weight, the new element suppression weight, and the area stability weight, respectively. .

[0071] The animation evolution process has the following characteristics: most existing elements are retained, such as height. New elements are gradually emerging, but their proportions are controllable, such as low... The positions of elements are basically stable, such as high .

[0072] The page-turning process has the following characteristics: a large number of elements from the previous frame disappear, such as low... A large number of new elements have emerged, such as high... Layout and structural changes, such as low .

[0073] Based on the animation evolution process and its characteristics, a first scoring threshold and a second scoring threshold are set to determine the inclusion score between consecutive frames. For example, the first scoring threshold is set to 0.75 and the second scoring threshold to 0.35. The logic for determining asymmetric inclusion relationships is as follows: If the animation is determined to be progressing, the current valid PPT presentation frame is updated to the optimal frame candidate set, and the page animation tracking status is performed. : Determined as a page-turning candidate, entering the waiting page-turning confirmation state; If the area is identified as ambiguous, proceed to the third stage of structure-assisted determination.

[0074] Furthermore, PPT animations come in various types, such as forward animation, reverse animation, and motion animation, among other animation evolution types. Different behaviors require different approaches.

[0075] For example, for positive animation, which is a scenario where the current valid PPT presentation frame has more content than the previous valid PPT presentation frame, the determination logic is shown in the table below: therefore, If the value is ≥0.75, the current valid PPT presentation frame can be directly determined to be an animation progression, and the page animation tracking state will be maintained; otherwise, it will be determined as a page turning candidate and enter the waiting page turning confirmation state. For reverse animation, which is a scenario where the content of the current valid PPT presentation frame is reduced compared to the previous valid PPT presentation frame, the determination logic is as follows: At this point, a semantic anchor point-first determination can be adopted, as follows: When the inclusion score of the previous and next frames is less than 0.35, it indicates that a large amount of content has been removed from the next frame compared to the previous frame. At this time, the similarity between the semantic anchors of the two frames is checked. If the similarity between the semantic anchors is greater than or equal to 0.7, it means that the core content of the page has not changed, and only some content has been reduced. It is judged as a reverse animation and the page animation tracking state is maintained. If the similarity between the semantic anchors is less than 0.7, the next frame is judged as a page-turning candidate and enters the waiting page-turning confirmation state. For motion animations, which are scenarios where the position of elements changes relative to the previous valid PPT presentation frame, the determination logic is as follows: At this point, we can supplement the matching based on the center distance between the center points of the element's bounding box, as follows: When the inclusion score of the previous and subsequent frames is less than 0.35, it indicates that the proportion of elements retained in the previous frame is low. In this case, the Euclidean distance between the center points of the bounding boxes of the same element in the previous and subsequent frames is compared, i.e., the element center distance. If the element center distance is less than a threshold... If the element's center distance exceeds the threshold, it indicates that the element is only slightly moved, and the animation is considered to be a small translational movement of the page element, maintaining the page animation tracking state; If the element has moved significantly or disappeared, it is considered a candidate for page turning and enters the waiting state for page turning confirmation. Based on the types of PPT animations, the logic for determining asymmetric inclusion relationships is as follows: : Directly determine as positive animation and proceed to the positive animation judgment; The similarity between semantic anchors and the distance between element centers are checked sequentially. If the similarity between semantic anchors is greater than 0.7, it is determined to be a reverse animation, and the reverse animation judgment is initiated; otherwise, it is checked whether the distance between element centers is less than a threshold. If so, it is determined to be a moving animation, and the moving animation judgment is entered; otherwise, the current frame is directly marked as a page-turning candidate and enters the waiting page-turning confirmation state.

[0076] It is directly identified as a blurry area and enters the third stage of structure-assisted judgment to avoid misjudgment caused by a single threshold in a single frame.

[0077] Furthermore, a single PPT slide may have multiple independent animation areas triggered simultaneously or sequentially. Therefore, it's also possible to determine animations by area, for example, dividing the PPT slide into a title area, a body text area, and an illustration area; where the body text area is the aggregated area of ​​the bounding boxes of page text, tables, and other elements. The comprehensive multi-animation determination logic is shown in the table below: For example, elements from the title area, body text area, and illustration area of ​​the preceding and following frames can be extracted separately, and the content of each area containing a score can be calculated independently. The animation and page-turning were completed area by area for preliminary assessment. Please refer to the above frame-by-frame inclusion score for details. The determination process will not be repeated here; If the title area, body text area, and illustration area are all determined to be animation, then the entire frame is determined to be animation. If the title area, body text area, and illustration area are all determined to be page turns, then the entire frame is determined to be a page turn. If the region determination results are mixed, such as some regions being determined as animation and some regions being determined as page turning, then they are classified as ambiguous regions and enter the third stage for structure-assisted determination.

[0078] In other embodiments, weight coefficients can be set for each region, for example, the weight coefficients for the title region, body text region, and illustration region are 0.3, 0.5, and 0.2, respectively; the inclusion scores for the title region, body text region, and illustration region of the preceding and following frames are calculated respectively, and then the inclusion scores corresponding to the title region, body text region, and illustration region of the preceding and following frames are weighted and summed to obtain the inclusion score of the entire frame for preliminary determination of animation, page turning, and blurred areas, as detailed above regarding the inclusion score of the entire frame. The determination process will not be repeated here.

[0079] For valid PPT presentation frames in the aforementioned ambiguous areas, structural auxiliary determination is required. This embodiment of the invention introduces layout change parameters and element type distribution change parameters to assist in determining whether a valid PPT presentation frame is a page turn or an animation progression.

[0080] Layout change parameters The area percentage of unmatched elements is calculated as follows: (13); in, This represents the total area of ​​unmatched elements in the currently valid PPT presentation frame. See above for specific element matching methods. This indicates the total page area of ​​the currently active PPT presentation frame.

[0081] Parameters of element type distribution variation The KL divergence was used to calculate the changes in element type distribution. The specific calculation is as follows: (14); in, This indicates the element type distribution of the previous valid PPT presentation frame, such as the proportion of text, tables, formulas, illustrations, etc. This indicates the element type distribution of the currently active PPT presentation frame; This represents the KL divergence function.

[0082] When it's impossible to determine whether a page turn or an animation is occurring based on the included rating, layout parameters can be changed. Parameters of element type distribution variation Auxiliary judgment criteria are specifically designed for scenarios with ambiguous conditions. For example, in animation progression scenarios, the type distribution usually remains stable, such as text pages displaying line by line with unchanged type proportions, resulting in low Key-Like Divergence (KL) divergence. In page-turning scenarios, the type distribution may change drastically, such as a text page transitioning to a chart page, resulting in high KL divergence. In animation progression scenarios, only a small number of elements are added or removed, resulting in minimal layout changes. In page-turning scenarios, a large number of existing elements have no matching items, while a large area of ​​unfamiliar elements is added, resulting in significant layout changes. Based on this, the following auxiliary judgment criteria are set: (15); (16); in, This indicates the preset threshold for element type distribution changes. When any auxiliary judgment condition is met, the corresponding valid PPT presentation frame is determined to be a page-turning candidate and enters the waiting page-turning confirmation state; otherwise, it is determined to be an animation progression and enters the page animation tracking state.

[0083] Furthermore, after the valid PPT presentation frames in the fuzzy area have undergone structure-assisted determination, a state transition can be performed. The state transition rules are as follows: This invention employs a multi-level judgment mechanism comprised of change-triggered judgment, semantic inclusion judgment, and structure-assisted judgment to refine the verification logic layer by layer: The first layer, change-triggered judgment, relies on regional histogram differences to quickly filter out a large number of redundant frames with no changes, significantly reducing the amount of subsequent complex matching operations and improving overall processing efficiency; the second layer, semantic inclusion judgment, integrates page semantic anchor points to determine asymmetric inclusion relationships between frames, initially classifying animation evolution and page-turning scenarios to complete basic frame type identification; the third layer, structure-assisted judgment, specifically targets fuzzy intervals, combining layout changes and element type distribution for secondary verification to determine whether it is animation evolution or page turning, significantly improving discrimination accuracy. Through the above multi-level judgment mechanism, valid PPT presentation frames can be stably and accurately distinguished as in-page animation evolution or page-turning transitions, balancing processing efficiency and recognition accuracy, achieving accurate differentiation of various animations and page turns, and providing reliable data support for achieving optimal keyframe extraction per page and the extraction of the most complete keyframe information.

[0084] In one optional embodiment, key PPT presentation frames are determined by tracking page state changes based on the page relationship type of adjacent valid PPT presentation frames, including: For a valid PPT presentation frame with a page-turning relationship type, the current valid PPT presentation frame is used as the initial page frame of the current PPT page, and the page state change sequence of the current PPT page is initialized; wherein, the previous valid PPT presentation frame is used as the end page frame of the page state change sequence of the previous PPT page. For valid PPT presentation frames with page relationship type of animation evolution, add them to the page state change sequence of the corresponding PPT page; For the valid PPT presentation frames in the page state change sequence, select the valid PPT presentation frame with the largest amount of information and output it as the key PPT presentation frame of the corresponding PPT page; The valid PPT presentation frames in the page state change sequence are fused to generate a fused PPT presentation frame, and the fused PPT presentation frame is output as the key PPT presentation frame of the corresponding PPT page.

[0085] In this embodiment of the invention, for valid PPT presentation frames of the same PPT page, a page state change sequence is constructed. For example, a valid PPT presentation frame whose page relationship type is determined to be page turning is the initial page frame of its PPT page. Subsequent valid PPT presentation frames whose page relationship type is determined to be animation evolution are added to the page state change sequence of the corresponding PPT page until the next valid PPT presentation frame whose page relationship type is determined to be page turning is determined again. The next valid PPT presentation frame is the initial page frame of the next PPT page.

[0086] For valid PPT presentation frames in the page state change sequence, two methods are used to extract key PPT presentation frames: the valid PPT presentation frame with the largest amount of information in a single frame and the integrated PPT presentation frame with the most complete information. The specific process is as follows: The selection mechanism for the most informative and effective PPT presentation frame, i.e., the optimal frame, is as follows: The single frame with the highest overall score from the animation process is selected as the key PPT presentation frame for output. The calculation process for the information score S is as follows: (17); in, These represent element weight, area weight, table weight, formula weight, and text weight, respectively. , , , , ; These represent the normalized number of page elements, element coverage area, number of tables, number of formulas, and text length in a valid PPT presentation frame, respectively.

[0087] The higher the element retention rate, the closer the animation is to completion. Therefore, the element retention rate can be used to... As a completion indicator, a completion score is obtained. ; (18); Based on information score S and completion score Calculate the overall score : (19); in, Represents the balance coefficient, such as .

[0088] The first frame of the page state change sequence is taken as the optimal frame, and its comprehensive score is the highest comprehensive score. It is then determined whether the comprehensive score of the next frame is greater than the sum of the highest comprehensive score and the preset significant advantage threshold. If so, it means that the next frame is significantly better, and the next frame is updated to the optimal frame, with its comprehensive score updated to the highest comprehensive score. If not, it is determined whether the difference between the comprehensive score of the next frame and the highest comprehensive score is less than the preset similar score threshold. If so, it means that the two frames are similar, and the later frame is retained, i.e., the next frame is updated to the optimal frame, with its comprehensive score updated to the highest comprehensive score. Otherwise, by traversing all frames of the page state change sequence using the current optimal frame and the highest comprehensive score, the optimal frame can be obtained and output as the key PPT presentation frame of this PPT page.

[0089] It should be noted that the embodiments of the present invention do not limit the values ​​of the significant advantage threshold and the score similarity threshold. For example, the significant advantage threshold can be set to 0.05 and the score similarity threshold can be set to 0.02.

[0090] The most comprehensive fusion PPT presentation frame, i.e., the fusion mechanism of the fusion frame: First, element collection is performed, collecting all elements within all frames during the animation process, i.e., the page animation tracking state, to form an element pool. : (20); in, Representing a frame All elements, This represents the union operation. These represent the start and end frames in the page animation tracking state for the same PPT page, respectively.

[0091] Each element in each element pool is discretized by coordinates, and elements at the same position are grouped together to achieve coordinate aggregation.

[0092] For the same group of elements, duplicate elements are removed and selected. The optimal element is selected from each group. The specific selection strategy is shown in the table below: Simultaneously, corner element filtering is performed to exclude elements located in the four vertex regions.

[0093] Finally, the selected elements from each group are merged to generate a merged PPT presentation frame containing complete PPT page information, and output as the key PPT presentation frame.

[0094] Furthermore, this embodiment of the invention introduces a state machine mechanism to track page state changes for valid PPT presentation frames of the animation evolution type, to confirm consecutive frames for valid PPT presentation frames of the page-turning type, and to aggregate all valid PPT presentation frames from different PPT pages. For example, the various states of the state machine are defined as shown in the table below: The state transition process of an effective PPT presentation frame is as follows: Figure 3As shown, the PPT presentation frame sequence, composed of PPT presentation frames, is used as the judgment target. Initialization is performed based on the first frame of the sequence, and the aforementioned multi-level judgment mechanism is executed. If it is a template page, the current frame enters a silent state, is skipped, and the judgment proceeds to the next frame. If it is not a template page, the inclusion score is calculated to determine if it is an animation progression. If the current frame is an animation progression, such as an inclusion score greater than 0.75, the current frame enters the page animation tracking state, further determining whether three consecutive frames are stable. If three consecutive frames are stable, such as a similarity of semantic anchor points greater than 0.7, the consecutive frames enter a silent state and are skipped. The process continues until three consecutive unstable frames are detected. If the similarity of the semantic anchor points in these three frames is no greater than 0.7, the system further checks if the included score is less than 0.35. If so, the frame is considered a candidate for page turning, and enters a waiting state for page turning confirmation. The system then checks if the included score in the next frame is less than 0.35. If not, the frame returns to the page animation tracking state. Otherwise, the frame is confirmed as a page turn, enters the page turning confirmation state, and the calculation output is executed, thus obtaining all valid PPT presentation frames for the corresponding PPT page. The state machine is then reset, and the next frame is initialized, repeating the above process to obtain the key PPT presentation frames for the next PPT page. It can be understood that each frame entering different states based on change-triggered judgment, semantic inclusion judgment, and structure-assisted judgment is processed according to the corresponding state in the state machine.

[0095] The overall PPT keyframe extraction process is as follows: Figure 4 As shown, the embodiments of the present invention can effectively distinguish between animation evolution and page turning, achieve simplified output of one frame per page, and ensure the most complete key frame information extracted. At the same time, through the dual output strategy of optimal frame and fused frame, the optimal frame has the characteristic of low storage requirements, which meets the requirements of fast browsing, deduplication, and fingerprint matching, while the fused frame has the characteristic of high information completeness, which meets the requirements of content retrieval and information integrity. Users can choose the output type according to the application scenario to achieve a balance between information integrity and storage cost.

[0096] The following section explains how to extract keyframes from PPT slides, based on several typical scenarios.

[0097] (1) The teacher uses PPT to explain the knowledge points. After clicking the mouse, the text appears line by line. The target video is a teaching video with line-by-line text animation. The processing process is as follows: First frame: Title "Chapter 1 Overview", text area is blank, set title as page semantic anchor; Second frame: The title remains the same, the first line of text appears, containing the rating. = 0.85, determined to be animation, update the optimal frame; Third frame: The title remains the same, but the second line of text appears, containing the rating. = 0.82, determined to be animation, update the optimal frame; Frame 4: The title remains the same, all text has appeared, including the rating. = 0.95, determined to be animation, update the optimal frame; Frame 5: The title changes to "Chapter Two" and includes a rating. = 0.2, indicating a page turn, output the best frame of the previous page, i.e., the fourth frame; Output results: The fourth frame is output as the optimal key PPT presentation frame. At the same time, a merged PPT presentation frame is generated based on the first to fourth frames as the key PPT presentation frame for complete content retrieval.

[0098] (2) The PPT page contains data charts, and the chart data is presented step by step. The target video is a teaching video with progressive animation of the charts. The processing procedure is as follows: First frame: Title "Data Analysis", chart frame present but data blank; Second frame: The title remains unchanged, the first group of data in the chart is displayed, it is determined to be an animation, and the optimal frame is updated; Third frame: The title remains unchanged. The second set of data in the chart shows that it is an animation, so the optimal frame is updated. Fourth frame: The title remains unchanged, the chart is fully displayed, the animation is considered complete, the overall score is the highest, and it is updated as the optimal frame; Fifth frame: The title remains unchanged, and the frame remains stable for three consecutive frames. The state then transitions to a static state, maintaining the optimal frame. Output results: The fourth frame of the complete chart is output as the optimal key PPT presentation frame. At the same time, a merged PPT presentation frame is generated based on the first to fourth frames as the key PPT presentation frame for subsequent retrieval.

[0099] (3) The PPT page contains complex tables and mathematical formulas. It is necessary to select the frame with the most complete information. The target video is a teaching video with PPT pages containing tables and formulas. The processing procedure is as follows: First frame: Title "Formula Derivation", formula framework exists; Second frame: Title remains unchanged, first step of formula is displayed, determined to be animation, update optimal frame; Third frame: The title remains unchanged, the table data is filled in, this is determined to be an animation, and the optimal frame is updated; Fourth frame: The title remains the same, the formula and table are complete, and the overall score is the highest, so it is updated as the best frame; Output results: The fourth frame of the complete output information is used as the optimal key PPT presentation frame. At the same time, a fused PPT presentation frame is generated based on the first to fourth frames and used as the key PPT presentation frame for subsequent retrieval.

[0100] (4) The video sampling rate is low, which may cause the entire animation process to be missed. The target video is a low sampling rate video, such as a teaching video with a sampling rate of 30 seconds / frame. The processing procedure is as follows: First frame: Title "Course Introduction", the initial state of the page, with the title set as the semantic anchor of the page; The second frame, 30 seconds later: The title remains the same, but some content is displayed, including the rating. = 0.6, which is considered a fuzzy region; Entering the third stage of structure-assisted judgment: the layout changes are not significant, so it is judged as an animation, and the optimal frame is updated; The third frame, 60 seconds later: The title remains the same, but the content is more complete, including a rating. = 0.75, determined to be animation, update the optimal frame; Fourth frame after 90 seconds: Title changes, includes rating. = 0.1, indicating a page turn, output the best frame of the previous page, i.e., the third frame; Output: The third frame is output as the optimal key PPT presentation frame. Simultaneously, a merged PPT presentation frame is generated based on the first to third frames and used as the key PPT presentation frame for complete content retrieval. This invention can still correctly determine animations and page turns even in low sampling rate scenarios.

[0101] (5) The PPT page title itself has animation effects, such as fade-in, zoom-in, etc. The target video is a tutorial video on title animation. The processing procedure is as follows: First frame: Because the title position exists in the title animation but the text is blurry, dynamically select the table with the largest area, illustration element or text element with the most information as the semantic anchor point of the page. Second frame: The title is clearly displayed, and the title position has a high similarity (IoU>0.8), indicating that it is an animation from the same page; Third frame: The title is stable, the main text begins to display, it is determined to be an animation, and the optimal frame is updated; Fourth frame: All content is complete, the animation is considered finished, and it is updated to the optimal frame; Output: The fourth frame is output as the optimal key PPT presentation frame. Simultaneously, a merged PPT presentation frame is generated based on frames one through four as the key PPT presentation frame for complete content retrieval. This invention uses high title position similarity as an auxiliary criterion to avoid misjudgments caused by title animation.

[0102] Compared with the prior art, the beneficial effects of the embodiments of the present invention are as follows: By adopting a multi-level progressive judgment mechanism combined with a state machine-driven keyframe extraction process, it can effectively resist the screen disturbances caused by step-by-step animations in PPT, reducing the misjudgment rate of animation as page turning by more than 80%; at the same time, it can output only one optimal main keyframe per page and optionally equip complete fusion frames, reducing redundant frames in the middle of the animation by more than 90%. By calculating a comprehensive score through a multi-feature weighted scoring mechanism, the representative frame with the highest content completeness can be selected, improving the completeness of key frame information by more than 50%. At the same time, it can be compatible with low sampling rate video input of 10 seconds per frame, and the recognition accuracy is stably maintained at more than 90%. By relying on a multi-level filtering mechanism, noise content and template pages in the video are filtered out, and the amount of computation is reduced layer by layer, resulting in an overall reduction of more than 60% in computing overhead.

[0103] The embodiments of this invention can be applied to various scenarios such as teaching video processing, courseware content indexing, online education platforms, teaching knowledge base construction, and teaching quality content evaluation. It can significantly reduce the redundancy of PPT keyframe extraction, improve retrieval efficiency, and extract keyframes with complete information, thereby improving index quality. In addition, it can support the automated processing of large-scale video content, providing a high-quality information source for courseware knowledge bases.

[0104] See Figure 5 , Figure 5 This invention provides a structural block diagram of a keyframe extraction system for teaching video PPTs, which includes: Scene filtering module 11 is used to perform scene filtering on the target video to obtain several PPT presentation frames in the target video; The template page detection module 12 is used to perform template page detection on each of the PPT presentation frames and filter out a number of valid PPT presentation frames. The dynamic anchor point selection module 13 is used to dynamically select anchor points for each of the valid PPT presentation frames and determine the semantic anchor points in each of the valid PPT presentation frames. The asymmetric inclusion relationship determination module 14 is used to determine the asymmetric inclusion relationship of adjacent valid PPT presentation frames based on the semantic anchor points in each of the valid PPT presentation frames, and to determine the page relationship type of adjacent valid PPT presentation frames. The page state tracking module 15 is used to track page state changes based on the page relationship type of adjacent valid PPT presentation frames and to determine key PPT presentation frames.

[0105] In one optional embodiment, the scene filtering module 11 includes: The scene classification unit is used to classify the scene of each frame of the target video using a pre-trained image classification model, and determine the scene category and confidence level of each frame. The first frame image filtering unit is used to output images with a confidence level higher than a preset confidence level threshold as the first candidate PPT presentation frames for images whose image category is PPT presentation screen. The second frame image filtering unit is used to filter out the first candidate PPT presentation frame by matching and recognizing screen elements through the OCR layout analysis model for images with a confidence level lower than the confidence threshold. The first determination unit is used to determine the area ratio threshold of key elements in the first candidate PPT presentation frame; wherein, the key elements include a video player and a PPT page; The third frame image filtering unit is used to select the first candidate PPT presentation frame as the PPT presentation frame, where the area ratio of the video player is less than the preset area ratio threshold or the area ratio of the PPT page is greater than the area ratio threshold.

[0106] In an optional embodiment, the template page detection module 12 includes: A screen division unit is used to divide each of the PPT presentation frames into several screen areas; The fourth frame image filtering unit is used to filter the PPT presentation frames based on the histogram difference between corresponding screen areas in two adjacent PPT presentation frames to obtain a second candidate PPT presentation frame. A page division unit is used to divide the PPT page in the second candidate PPT presentation frame into several page areas; The fifth frame image filtering unit is used to identify elements in each page area of ​​the second candidate PPT presentation frame and determine whether the PPT page in the second candidate PPT presentation frame is a PPT template page; if yes, skip the second candidate PPT presentation frame; if no, output the second candidate PPT presentation frame as a valid PPT presentation frame.

[0107] In an optional embodiment, the dynamic anchor point selection module 13 includes: The semantic anchor point selection unit is used to select semantic anchor points in each of the valid PPT presentation frames according to the preset anchor point selection priority rules. The semantic anchor points include the title on the PPT page, the main visual element with the largest area on the PPT page, and the main text area with the most text on the PPT page; the selection priority of the title is greater than the selection priority of the main visual element, and the selection priority of the main visual element is greater than the selection priority of the main text area.

[0108] In an optional embodiment, the asymmetric inclusion relationship determination module 14 includes: The similarity determination unit is used to determine whether the similarity between the semantic anchor point of the current valid PPT presentation frame and the semantic anchor point of the previous valid PPT presentation frame is greater than a preset similarity threshold. The page relationship type determination unit is used to determine if the PPT pages in the current valid PPT presentation frame and the previous valid PPT presentation frame are the same page, and to determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame based on the element retention rate, new element ratio and area stability of the current valid PPT presentation frame. The content evolution determination unit is used to determine the content evolution of the currently valid PPT presentation frame if no.

[0109] In one optional embodiment, the content evolution determination unit includes: The inter-frame similarity determination subunit is used to determine whether the similarity between the semantic anchor point of the next valid PPT presentation frame and the semantic anchor point of the current valid PPT presentation frame is not greater than the similarity threshold. The first page relationship type determines the sub-unit. If so, it determines that the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is page turning. The second page relationship type determines the sub-unit. If not, it determines the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame based on the element retention rate, new element ratio, and area stability of the current valid PPT presentation frame.

[0110] In one optional embodiment, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined based on the element retention rate, new element ratio, and area stability of the current valid PPT presentation frame, including: The element retention rate of the current valid PPT presentation frame is calculated based on the first element that matches the elements of the current valid PPT presentation frame and the total elements of the previous valid PPT presentation frame. Calculate the proportion of new elements in the current valid PPT presentation frame based on the second element in the current valid PPT presentation frame that does not match the element in the previous valid PPT presentation frame, and the total number of elements in the current valid PPT presentation frame. The area stability of the current valid PPT presentation frame is calculated based on the area ratio of the matched element in the current valid PPT presentation frame and the total area of ​​all elements in the previous valid PPT presentation frame. The element retention rate, the proportion of new elements, and the area stability are weighted and summed to calculate the inclusion score of the current valid PPT presentation frame; Based on the included score, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame.

[0111] In an optional embodiment, determining the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame based on the included score includes: When the included score is greater than the preset first score threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be animation evolution; When the score is less than the preset second score threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be page turning; When the included score is less than or equal to the first score threshold and greater than or equal to the second score threshold, the layout change parameters between the current valid PPT presentation frame and the previous valid PPT presentation frame are calculated based on the third element in the previous valid PPT presentation frame that does not match the element of the current valid PPT presentation frame and the total elements of the previous valid PPT presentation frame. Based on the KL divergence between the element type distribution of the current valid PPT presentation frame and the element type distribution of the previous valid PPT presentation frame, calculate the parameter of change in element type distribution between the current valid PPT presentation frame and the previous valid PPT presentation frame. When the layout change parameter between the current valid PPT presentation frame and the previous valid PPT presentation frame is greater than the preset first parameter threshold or the element type distribution change parameter is greater than the preset second parameter threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be page turning. If the layout change parameter between the current valid PPT presentation frame and the previous valid PPT presentation frame is not greater than the first parameter threshold and the element type distribution change parameter is not greater than the second parameter threshold, then the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be animation evolution.

[0112] In one optional embodiment, the page state tracking module 15 includes: The page state change sequence initialization unit is used to initialize the page state change sequence of the current PPT page for the valid PPT presentation frame corresponding to the page turn type, with the current valid PPT presentation frame as the initial page frame of the current PPT page; wherein, the previous valid PPT presentation frame is used as the end page frame of the page state change sequence of the previous PPT page. The presentation frame aggregation unit is used to add valid PPT presentation frames corresponding to the page relationship type of animation evolution to the page state change sequence of the corresponding PPT page; The first keyframe output unit is used to select the valid PPT presentation frame with the largest amount of information from the valid PPT presentation frames in the page state change sequence, and output it as the key PPT presentation frame of the corresponding PPT page. The second keyframe output unit is used to fuse elements of valid PPT presentation frames in the page state change sequence to generate a fused PPT presentation frame, and output the fused PPT presentation frame as the key PPT presentation frame of the corresponding PPT page.

[0113] It should be noted that the working process of each module in the teaching video PPT keyframe extraction system described in the embodiments of the present invention can refer to the working process of the teaching video PPT keyframe extraction method described in the above embodiments, and the technical effect achieved is the same as that of the teaching video PPT keyframe extraction method described in the above embodiments, so it will not be repeated here.

[0114] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0115] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for extracting keyframes from teaching video PPTs, characterized in that, include: Scene filtering is performed on the target video to obtain several PPT presentation frames from the target video; Template page detection is performed on each of the PPT presentation frames to filter out a number of valid PPT presentation frames; Dynamic anchor point selection is performed on each of the valid PPT presentation frames to determine the semantic anchor points in each of the valid PPT presentation frames. Based on the semantic anchor points in each of the valid PPT presentation frames, an asymmetric inclusion relationship determination is performed on adjacent valid PPT presentation frames to determine the page relationship type of adjacent valid PPT presentation frames. Based on the page relationship type of adjacent valid PPT presentation frames, track page state changes to determine key PPT presentation frames.

2. The method for extracting keyframes from teaching video PPTs as described in claim 1, characterized in that, Scene filtering is performed on the target video to obtain several PPT presentation frames from the target video, including: For each frame of the target video, a pre-trained image classification model is used to classify the scene and determine the scene category and confidence level of each frame. For images categorized as PPT presentation frames, images with a confidence level higher than a preset confidence threshold are output as the first candidate PPT presentation frames. For images with a confidence level lower than the confidence threshold, the OCR layout analysis model is used to match and identify screen elements to select the first candidate PPT presentation frame. Boundary scene verification is performed on the first candidate PPT presentation frame, and the first candidate PPT presentation frame that passes the boundary scene verification is selected as the PPT presentation frame.

3. The method for extracting keyframes from teaching video PPTs as described in claim 1, characterized in that, Template page detection is performed on each of the aforementioned PPT presentation frames to filter out several valid PPT presentation frames, including: Each of the aforementioned PPT presentation frames is divided into several screen areas; Based on the histogram difference between corresponding screen areas in two adjacent PPT presentation frames, the PPT presentation frames are filtered to obtain a second candidate PPT presentation frame. Divide the PPT pages in the second candidate PPT presentation frame into several page areas; Element recognition is performed on each page area in the second candidate PPT presentation frame to determine whether the PPT page in the second candidate PPT presentation frame is a PPT template page; If so, skip the second candidate PPT presentation frame; If not, output the second candidate PPT presentation frame as the valid PPT presentation frame.

4. The method for extracting keyframes from teaching video PPTs as described in claim 1, characterized in that, Dynamic anchor point selection is performed on each of the valid PPT presentation frames to determine the semantic anchor points in each valid PPT presentation frame, including: According to the preset anchor point selection priority rules, select the semantic anchor points in each of the effective PPT presentation frames; The semantic anchor points include the title on the PPT page, the main visual element with the largest area on the PPT page, and the main text area with the most text on the PPT page; the selection priority of the title is greater than the selection priority of the main visual element, and the selection priority of the main visual element is greater than the selection priority of the main text area.

5. The method for extracting keyframes from teaching video PPTs as described in claim 1, characterized in that, Based on the semantic anchor points in each of the valid PPT presentation frames, an asymmetric inclusion relationship determination is performed on adjacent valid PPT presentation frames to determine the page relationship type of adjacent valid PPT presentation frames, including: Determine whether the similarity between the semantic anchor point of the current valid PPT presentation frame and the semantic anchor point of the previous valid PPT presentation frame is greater than a preset similarity threshold. If so, determine that the PPT pages in the current valid PPT presentation frame and the previous valid PPT presentation frame are the same page, and determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame based on the element retention rate, new element ratio and area stability of the current valid PPT presentation frame. If not, determine the content evolution of the currently valid PPT presentation frame.

6. The method for extracting keyframes from teaching video PPTs as described in claim 5, characterized in that, Determine the content evolution of the currently valid PPT presentation frame, including: When determining whether the similarity between the semantic anchor point of the next valid PPT presentation frame and the semantic anchor point of the current valid PPT presentation frame is not greater than the similarity threshold; If so, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame as page turning; If not, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame based on the element retention rate, new element ratio, and area stability of the current valid PPT presentation frame.

7. The method for extracting keyframes from teaching video PPTs as described in claim 5 or 6, characterized in that, Based on the element retention rate, new element ratio, and area stability of the current valid PPT presentation frame, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame, including: The element retention rate of the current valid PPT presentation frame is calculated based on the first element that matches the elements of the current valid PPT presentation frame and the total elements of the previous valid PPT presentation frame. Calculate the proportion of new elements in the current valid PPT presentation frame based on the second element in the current valid PPT presentation frame that does not match the element in the previous valid PPT presentation frame, and the total number of elements in the current valid PPT presentation frame. The area stability of the current valid PPT presentation frame is calculated based on the area ratio of the matched element in the current valid PPT presentation frame and the total area of ​​all elements in the previous valid PPT presentation frame. The element retention rate, the proportion of new elements, and the area stability are weighted and summed to calculate the inclusion score of the current valid PPT presentation frame; Based on the included score, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame.

8. The method for extracting keyframes from teaching video PPTs as described in claim 7, characterized in that, Based on the included score, determine the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame, including: When the included score is greater than the preset first score threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be animation evolution; When the score is less than the preset second score threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be page turning; When the included score is less than or equal to the first score threshold and greater than or equal to the second score threshold, the layout change parameters between the current valid PPT presentation frame and the previous valid PPT presentation frame are calculated based on the third element in the previous valid PPT presentation frame that does not match the element of the current valid PPT presentation frame and the total elements of the previous valid PPT presentation frame. Based on the KL divergence between the element type distribution of the current valid PPT presentation frame and the element type distribution of the previous valid PPT presentation frame, calculate the parameter of change in element type distribution between the current valid PPT presentation frame and the previous valid PPT presentation frame. When the layout change parameter between the current valid PPT presentation frame and the previous valid PPT presentation frame is greater than the preset first parameter threshold or the element type distribution change parameter is greater than the preset second parameter threshold, the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be page turning. If the layout change parameter between the current valid PPT presentation frame and the previous valid PPT presentation frame is not greater than the first parameter threshold and the element type distribution change parameter is not greater than the second parameter threshold, then the page relationship type between the current valid PPT presentation frame and the previous valid PPT presentation frame is determined to be animation evolution.

9. The method for extracting keyframes from teaching video PPTs as described in claim 8, characterized in that, Based on the page relationship type of adjacent valid PPT presentation frames, track page state changes to determine key PPT presentation frames, including: For a valid PPT presentation frame with a page-turning relationship type, the current valid PPT presentation frame is used as the initial page frame of the current PPT page, and the page state change sequence of the current PPT page is initialized; wherein, the previous valid PPT presentation frame is used as the end page frame of the page state change sequence of the previous PPT page. For valid PPT presentation frames with page relationship type of animation evolution, add them to the page state change sequence of the corresponding PPT page; For the valid PPT presentation frames in the page state change sequence, select the valid PPT presentation frame with the largest amount of information and output it as the key PPT presentation frame of the corresponding PPT page; The valid PPT presentation frames in the page state change sequence are fused to generate a fused PPT presentation frame, and the fused PPT presentation frame is output as the key PPT presentation frame of the corresponding PPT page.

10. A keyframe extraction system for teaching video PPTs, characterized in that, include: The scene filtering module is used to perform scene filtering on the target video to obtain several PPT presentation frames in the target video. The template page detection module is used to perform template page detection on each of the PPT presentation frames and filter out a number of valid PPT presentation frames. The dynamic anchor point selection module is used to dynamically select anchor points for each of the valid PPT presentation frames and determine the semantic anchor points in each of the valid PPT presentation frames. The asymmetric inclusion relationship determination module is used to determine the asymmetric inclusion relationship between adjacent valid PPT presentation frames based on the semantic anchor points in each of the valid PPT presentation frames, and to determine the page relationship type of adjacent valid PPT presentation frames. The page state tracking module is used to track page state changes based on the page relationship type of adjacent valid PPT presentation frames, and to identify key PPT presentation frames.

Images