A document parsing method, system, device and medium

Abstract

This invention discloses a document parsing method, system, device, and medium, relating to the field of document image parsing technology. The method includes the following steps: obtaining an initial representation of a multi-page document image; performing layout parsing on each page of the document image based on the initial representation, and during layout parsing, using text region detection, table region detection, and image region detection to obtain the position, content, and category of element information within each page of the document image, converting the position, content, and category of the element information into vector representations, reconstructing each page of the document image using vector representations, obtaining the cross-page matching relationship of each reconstructed page of the document image, and integrating the document images based on the cross-page matching relationship to obtain the document image parsing result. This invention alleviates the problem that existing technologies cannot effectively parse multi-column document images and multi-chart document images by reconstructing each page of the document image using vector representations.

Description

Technical Field

[0001] This invention relates to the field of document image parsing technology, and in particular to a document parsing method, system, device and medium. Background Technology

[0002] Document image parsing technology, also known as document image structuring or text mining technology, refers to the process of extracting information from unstructured electronic document images, PDFs, or other file formats. It typically involves computer science fields such as Large Language Modeling (LLM), Computer Vision (CV), and Pattern Recognition. This technology analyzes text content, identifies key data such as headings, paragraphs, tables, links, and image annotations, and transforms them into data structures that machines can understand and manipulate.

[0003] Common applications of document image parsing technology include search engine optimization, data analysis, automated report generation, information extraction, and knowledge graph construction. For example, in financial statements, parsing technology can help automatically identify important figures such as income and expenses; in legal document images, key parts of terms and conditions can be found.

[0004] Existing document image parsing algorithms can effectively recover the layout information of document images with simple layout structures; however, for document images with multiple charts and graphs, existing algorithms generally perform poorly. Although they can extract layout information, the messy layout information and low correlation between the hierarchical structure of the information result in inaccurate relationships between the obtained document elements. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a document parsing method, system, device, and medium to solve the problem that the relationships between document elements are not accurate due to the cluttered layout information and low correlation between hierarchical structures of information in the prior art.

[0006] The present invention specifically provides the following technical solution: A document parsing method includes the following steps: Obtain an initial representation of a multi-page document image; the initial representation is the extracted text information. Based on the initial representation of each page of the document image, a cascaded detection framework is used to perform refined layout analysis on each page. Specifically, during the analysis process, text region detection, table region detection, and image region detection are used to obtain the position, content, and category of each element within each page of the document image. A cross-element relationship perception mechanism is introduced to remove the overlap of elements within the document image, resulting in optimized elements. The position, content, and category information of the elements are encoded into dynamic vector representations, and based on an adaptive mechanism, the dynamic vector representations are dynamically adjusted according to the document structure and content complexity for feature expression adjustment. Based on the dynamically vector representations adjusted for feature expression, the document images are reconstructed page by page for structured analysis and semantic representation. The element categories include text, tables, and images. By combining the dynamic vector representation adjusted by feature expression, a cross-page matching label is constructed. A dynamic positive sample mining mechanism is integrated into the contrastive learning method to distinguish similar elements on different pages from continuous content across pages. Cross-page matching relationships are learned through an end-to-end training framework. Based on the learned cross-page matching relationships, structural integration and semantic alignment are performed on the reconstructed document image of each page, and finally, high-quality and structurally coherent document parsing results are output.

[0007] Preferably, the initial representation of the multi-page document image is obtained by: The multi-page document image is split into pages and each page image is determined to be either an electronic document image or a scanned document image. For electronic document images, the metadata of the document image is read directly, and the text information of the document image is obtained through the metadata; For scanned document images, the text location is located using object detection methods, and the corresponding text information is identified using text recognition algorithms; By integrating text information extracted from electronic document images or scanned document images, a preliminary representation of the document images was constructed.

[0008] Preferably, the text region detection includes: Select text regions at multiple different locations in the document image, identify the text region at each location, and obtain the initial text block at each location; Determine the orientation of the initial text block at each position, and divide the initial text block into individual text lines based on the orientation and line spacing information of the initial text block; A single line of text is merged with adjacent lines of text in the same area to form a final text block.

[0009] Preferably, the table region detection includes: Select table images from multiple different locations in the document image, process the table area at each location, identify the row and column boundaries of the table at each location, and reconstruct the logical structure of the table. Analyze the relative positions of cells to determine which content belongs to the same column or row; Parse the contents of each cell, distinguishing between text, numbers, and dates; If a cell has an incorrect format, a correct format is predicted based on the content of the adjacent cells, and the incorrect format is corrected.

[0010] Preferably, the image region detection includes: Select multiple image regions at different locations within the document image, and process each image region to obtain an image description; the processing procedure is as follows: When there is text in the document image in the image area, OCR technology is used to recognize and extract any text in the document image, and convert the text in the document image into an editable text format; When there is no text in the document image within the image region, a visual language model is used to analyze the content of the document image and generate descriptive text.

[0011] Preferably, this method further includes: Identify whether a document image uses a multi-column layout and save the multi-column layout information; Extract the directory tree structure of the document image, which includes the organization information of chapters, sections and subheadings in the document image, as well as metadata of page numbers, document image titles and authors; The saved multi-column layout information and directory tree structure are used as layout and directory vectors for conversion. These vectors are then combined with the position, content, and category of the element information to reconstruct the document image for each page.

[0012] Preferably, the step of converting the position, content, and category of element information into vector representations, and reconstructing the document image for each page using the vector representations, includes: The position, content, and category of element information, as well as multi-column layout information and directory tree structure, are input into the encoder for processing, and the position, content, and category of element information, as well as multi-column layout information and directory tree structure, are converted into vector representations. The vector representation is input into the layout restoration model. The layout of the document image is reconstructed by comprehensively considering the position, content and category of element information, as well as the relationship between multi-column layout information and directory tree structure. The text, images and element information are correctly arranged and organized in the restored document image.

[0013] This invention provides a document parsing system, comprising: The acquisition module is used to acquire the initial representation of a multi-page document image; the initial representation is the extracted text information. The document image reconstruction module combines the initial representation of each page's document image with a cascaded detection framework to perform refined layout analysis on each page. Specifically, during the analysis process, text region detection, table region detection, and image region detection are used to obtain the position, content, and category of each element within each page's document image. A cross-element relationship awareness mechanism is introduced to remove overlaps between elements within the document image, resulting in optimized elements. The position, content, and category information of the elements are encoded into dynamic vector representations, and based on an adaptive mechanism, the dynamic vector representations are dynamically adjusted according to the document structure and content complexity for feature expression adjustment. Based on the dynamically vector representations adjusted for feature expression, the document images are reconstructed page by page, performing structured analysis and semantic representation. The element categories include text, tables, and images. The document image integration module is used to construct a cross-page matching label by combining the dynamic vector representation adjusted by feature expression, and integrates a dynamic positive sample mining mechanism with the contrastive learning method to distinguish similar elements and continuous content across pages. Through an end-to-end training framework, it learns the cross-page matching relationship. Based on the learned cross-page matching relationship, it performs structural integration and semantic alignment on each page of the reconstructed document image, and finally outputs a high-quality and structurally coherent document parsing result.

[0014] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of the document parsing method described above.

[0015] The present invention provides a storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the document parsing method described above.

[0016] Compared with the prior art, the present invention has the following significant advantages: This invention performs layout analysis on the initial representation of document images to obtain element information within each page of the document image. To improve the relationships between different categories of elements, it uses text region detection, table region detection, and image region detection to obtain the position, content, and category of element information within each page of the document image. The position, content, and category of element information are then converted into vector representations, and each page of the document image is reconstructed using these vector representations. This alleviates the problem that existing technologies cannot effectively parse multi-page document images. Furthermore, by obtaining the cross-page matching relationships for each page and integrating them, the parsing results of the document images are obtained, improving the hierarchical connections between multi-page document images. Attached Figure Description

[0017] Figure 1This is a flowchart illustrating the overall process of a document parsing method according to the present invention. Figure 2 A screenshot of the input PDF; Figure 3 This is the first parsing result of the input PDF; Figure 4 This is the second parsing result of the input PDF; Figure 5 This is the third parsing result of the input PDF. 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] like Figure 1 As shown, the present invention provides a document parsing method, which specifically includes the following steps: Step S1: Obtain the initial representation of the multi-page document image; the initial representation is the extracted text information.

[0020] Obtain an initial representation of the multi-page document image, including: The multi-page document image is split into pages, and each page image is determined to be either an electronic document image or a scanned document image (e.g., ...). Figure 2 The image shown is a single page of a PDF document. Multiple pages of document images are those requiring document image parsing, such as... Figure 2 This is a page of an electronic document image.

[0021] For electronic document images, the metadata of the document image can be read directly to obtain the text information. Specifically, there are many existing methods to recover text information from the metadata of a document image, such as the PyPDF2 library; the existing PyPDF2 library can be used to read the metadata of the document image to obtain the text information (e.g., Figure 2 The image shows the result from Figure 1 (Text information extracted from metadata).

[0022] For scanned document images, the text location is determined using object detection methods, and the corresponding text information is then identified using text recognition algorithms. Specifically, due to the lack of metadata support, text information can only be obtained through image processing methods. Currently, the more widely used method is deep learning-based text detection and recognition algorithms. This approach first locates the text location using object detection methods, and then uses text recognition algorithms to identify the corresponding text information. This type of technology is relatively mature and directly utilizes existing text recognition and detection engines to extract text information from scanned document images.

[0023] By integrating textual information extracted from electronic document images or scanned document images, a preliminary representation of the document images was constructed. This not only encompasses the direct textual description of its content but also incorporates structured data and metadata, laying the foundation for further analysis and understanding.

[0024] Step S2: Combining the initial representation of each page of the document image, a cascaded detection framework is used to perform refined layout analysis on each page of the document image. Specifically, during the analysis process, text region detection, table region detection, and image region detection are used to obtain the position, content, and category of each element within each page of the document image. A cross-element relationship perception mechanism is introduced to remove the overlap of elements within the document image, resulting in optimized elements. This addresses the issues of element overlap and hierarchical interaction, improving the robustness and accuracy of detection. The position, content, and category information of the elements are encoded into dynamic vector representations. Based on an adaptive mechanism, the dynamic vector representations are dynamically adjusted according to the document structure and content complexity to achieve more accurate semantic modeling of document elements. Based on the dynamically vector representations adjusted for feature expression, the document images are reconstructed page by page for structured analysis and semantic representation. The element categories include text, tables, and images.

[0025] Page layout analysis meticulously analyzes the internal structure of each page of document images, identifying and understanding the relationships and layouts of elements such as text, images, and tables, thus providing a precise framework for comprehensive interpretation of document images and information retrieval.

[0026] Obtaining element information, including location, content, and category, within each page of a document image is crucial for detecting layout elements. This involves identifying the elements contained in a single page of a document image, such as text region detection, image region detection, and table region detection. The document image is then input into the layout element detection model to obtain the element information contained in a single page of the document image.

[0027] In one embodiment, a PDF document is used as the document image. During detection, the location, content, and category of each text, table, and image in the PDF document are obtained. The category can be the font, font size, etc. of the text, as well as the type of the header, column, and row in the table, and the shape of the image, etc.

[0028] Text region detection includes: We select text regions at multiple different locations within a document image and use an object detection model to identify the text region at each location, obtaining an initial text block for each location. This facilitates finer-grained segmentation and recognition. Here, each location represents the position of a local region within a PDF document after one element is separated from another.

[0029] Determine the orientation of the initial text block at each position (from left to right or from right to left) to facilitate correct text line segmentation, and divide the initial text block at each position into individual text lines based on the orientation and line spacing information of the initial text block at each position.

[0030] Based on text connectivity, a single text line is merged with adjacent text lines in the same region (i.e., the region where the same position is located) into a final text block.

[0031] Table area detection includes: Select table images from multiple different locations within the document image, use deep learning algorithms to process the table region at each location, identify the row and column boundaries of the table at each location using a table restoration model, and reconstruct the logical structure of the table.

[0032] The table recovery model analyzes the relative positions of cells to determine content belonging to the same column or row, even if this content was incorrectly separated in the original data.

[0033] Based on structure recognition, the content of each cell is parsed, distinguishing between text, numbers, and dates, and attempting to understand the relationships between them.

[0034] If data is missing or formatted incorrectly, the table recovery model automatically fills in or corrects errors using patterns learned during training, ensuring data integrity and accuracy. Specifically, if a cell has an incorrect format, the model predicts a correct format based on the content of adjacent cells, correcting the incorrect format. This process is also labeled; the model calculates a loss function based on the predictions and labels, guiding the network to learn the correct format.

[0035] Finally, the restored table will be presented in a clear and structured manner, facilitating further data analysis or direct use in reports and presentations. This process not only saves a significant amount of manual data processing time but also greatly improves the accuracy and reliability of data processing, laying the foundation for subsequent large-scale input document image analysis models.

[0036] Image region detection includes: Select multiple image regions at different locations within the document image, and process each region to obtain an image description; the processing procedure is as follows: When text exists in different image regions within a document image, Optical Character Recognition (OCR) technology is used to identify and extract any text in the document image, converting the text in the document image into an editable text format, thereby preserving the readability and searchability of the original information.

[0037] When there is no text in each image region at different locations in a document image, a deep learning-based visual language model (such as a description generator based on a convolutional neural network) is used to analyze the content of the document image and generate a descriptive text. This helps to convey the visual information of the image so that the image can be inserted into the corresponding position when restoring the document image layout later.

[0038] After identifying text regions, table regions, and image regions in different categories, the process also includes: This function identifies whether a document image uses a multi-column layout and saves the multi-column information for easy information retrieval later. A multi-column layout is a common format in document image design; it divides page content into multiple vertical columns to improve information organization and readability.

[0039] To effectively restore the hierarchical structure of the document image, a comprehensive approach was adopted to conduct in-depth analysis of the entire document image and extract its directory tree structure. The directory tree not only captures the organizational information of the document image's chapters, sections, and subheadings, but also includes page numbers, document image titles, and author metadata.

[0040] The saved multi-column layout information and directory tree structure are converted into vectors for layout and directory, and combined with the position, content, and category of element information to reconstruct the document image for each page. The position, content, and category of element information are converted into vector representations, and these vector representations are used to reconstruct the document image for each page, including:

[0041] The information obtained in the previous steps is processed using a Transformer-based encoder, which is capable of processing sequence data and capturing long-range dependencies.

[0042] The encoder processes the position, content, and category (text, image description) of elements in the document image, as well as the multi-column layout information and directory tree structure. It then transforms the position, content, and category of elements, as well as the multi-column layout information and directory tree structure, into (high-dimensional) vector representations. These vectors contain deep semantic information of the document image.

[0043] The vector representation is input into the large-scale layout restoration model, which comprehensively considers the position, content, and category of element information, as well as multi-column layout information and directory tree structure to reconstruct the document image's layout. In other words, this model can comprehensively consider all aspects of the document image, including the relationship between text content, image information, layout, and directory structure, to extract deeper information. The large-scale layout restoration model uses this information to reconstruct the document image's layout, correctly arranging and organizing text, images, and element information within the restored document image.

[0044] Finally, as Figure 3 , Figure 4 and Figure 5 As shown, the results of document image parsing were obtained, which are... Figure 2 The analysis results include not only the text content and image description of the document image, but also the restored layout and hierarchical structure. This process preserves the original information of the document image and improves its readability and usability. Whether in digital publishing, document image management, or content analysis, this document image analysis technology can provide users with significant assistance and value.

[0045] Step S3: Combining the dynamically adjusted vector representation with the feature representation, a cross-page matching label is constructed. A dynamic positive sample mining mechanism is integrated with the contrastive learning method to distinguish similar elements across different pages from continuous content across pages. An end-to-end training framework is used to collaboratively optimize the vector representation and the learned cross-page matching relationship. Based on the learned cross-page matching relationship, structural integration and semantic alignment are performed on each reconstructed page document image, ultimately outputting a high-quality, structurally coherent document parsing result. In cross-page matching, a continuous structure modeling and document quality enhancement strategy specifically for cross-page tables is introduced to ensure the integrity of the table content and layout.

[0046] This tag differs from the previous element information tags. This part specifically constructs a cross-page matching tag. This tag can be understood as the correct reading order manually marked between these two pages. Then, the model will also predict a set of orders, and the model can then learn the correct reading order marked by humans.

[0047] By analyzing cross-page matching relationships, the connections and transitions between document images on each page are accurately captured, ensuring the overall coherence of the document images and the integrity of the information, providing key links and context for a deeper understanding of the document image content.

[0048] By comprehensively analyzing and integrating page elements, cross-page relationships, and document image structure, a hierarchical representation of document images is generated. This integration not only restores the physical layout and logical order of document images, but also reveals the hierarchy and depth of information, providing a systematic view and framework for the comprehensive understanding and effective utilization of document images.

[0049] Based on the above method, the present invention provides a document parsing system, including: an acquisition module, a document image reconstruction module, and a document image integration module.

[0050] The document acquisition module is used to obtain the initial representation of multi-page document images; the initial representation is the extracted text information. The document image reconstruction module is used to combine the initial representation of each page of the document image and use a cascaded detection framework to perform refined layout analysis on each page of the document image. Specifically, during the analysis process, the position, content, and category of each element within each page of the document image are obtained through text region detection, table region detection, and image region detection. A cross-element relationship perception mechanism is introduced to remove the overlap of elements within the document image, resulting in optimized elements. The position, content, and category information of the elements are encoded into dynamic vector representations, and based on an adaptive mechanism, the vector representation is dynamically adjusted according to the document structure and content complexity. The dynamic vector representation is adjusted for feature representation; based on the adjusted dynamic vector representation, the document image is reconstructed page by page, and structured parsing and semantic representation are performed; the element categories include text, tables, and images; the document image integration module is used to combine the adjusted dynamic vector representation to construct a cross-page matching label, and integrates a contrastive learning method with a dynamic positive sample mining mechanism to distinguish similar elements and continuous content across pages, and learns the cross-page matching relationship through an end-to-end training framework; based on the learned cross-page matching relationship, the reconstructed document image of each page is structurally integrated and semantically aligned, and finally outputs a high-quality, structurally coherent document parsing result.

[0051] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of a document parsing method.

[0052] According to the disclosed embodiments, the computer device can communicate with one or more external devices (e.g., keyboard, pointing device, Bluetooth communication, etc.) or with any device that enables the computing device to communicate with one or more other computing devices (e.g., router, demodulator, etc.).

[0053] The present invention also provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of a document parsing method.

[0054] According to the disclosed embodiments, the storage medium can be a non-volatile computer-readable storage medium, such as, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, the storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0055] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.

Claims

1. A document parsing method, characterized in that, Includes the following steps: Obtain an initial representation of a multi-page document image; the initial representation is the extracted text information. Based on the initial representation of each page of the document image, a cascaded detection framework is used to perform refined layout analysis on each page. Specifically, during the analysis process, text region detection, table region detection, and image region detection are used to obtain the position, content, and category of each element within each page of the document image. A cross-element relationship perception mechanism is introduced to remove the overlap of elements within the document image, resulting in optimized elements. The position, content, and category information of the elements are encoded into dynamic vector representations, and based on an adaptive mechanism, the dynamic vector representations are dynamically adjusted for feature expression based on the document structure and content complexity. The document image is reconstructed page by page based on the dynamic vector representation adjusted by feature expression, and then structured parsing and semantic representation are performed; wherein the element categories include text, tables and images; By combining the dynamic vector representation adjusted by feature expression, a cross-page matching label is constructed. The contrastive learning method is integrated with a dynamic positive sample mining mechanism to distinguish similar elements on different pages from continuous content across pages. The cross-page matching relationship is learned through an end-to-end training framework. Based on the learned cross-page matching relationships, the reconstructed document images of each page are structurally integrated and semantically aligned, ultimately outputting high-quality, structurally coherent document parsing results.

2. The document parsing method as described in claim 1, characterized in that, The initial representation of the multi-page document image obtained includes: The multi-page document image is split into pages and each page image is determined to be either an electronic document image or a scanned document image. For electronic document images, the metadata of the document image is read directly, and the text information of the document image is obtained through the metadata; For scanned document images, the text location is located using object detection methods, and the corresponding text information is identified using text recognition algorithms; By integrating text information extracted from electronic document images or scanned document images, a preliminary representation of the document images was constructed.

3. The document parsing method as described in claim 1, characterized in that, The text region detection includes: Select text regions at multiple different locations in the document image, identify the text region at each location, and obtain the initial text block at each location; Determine the orientation of the initial text block at each position, and divide the initial text block into individual text lines based on the orientation and line spacing information of the initial text block; A single line of text is merged with adjacent lines of text in the same area to form a final text block.

4. The document parsing method as described in claim 1, characterized in that, The table region detection includes: Select table images from multiple different locations in the document image, process the table area at each location, identify the row and column boundaries of the table at each location, and reconstruct the logical structure of the table. Analyze the relative positions of cells to determine which content belongs to the same column or row; Parse the contents of each cell, distinguishing between text, numbers, and dates; If a cell has an incorrect format, a correct format is predicted based on the content of the adjacent cells, and the incorrect format is corrected.

5. The document parsing method as described in claim 1, characterized in that, The image region detection includes: Select multiple image regions at different locations within the document image, and process each image region to obtain an image description; the processing procedure is as follows: When text exists in the image area of ​​a document image, OCR technology is used to recognize and extract any text in the document image, converting the text in the document image into an editable text format; When there is no text in the image area of ​​a document image, a visual language model is used to analyze the content of the document image and generate descriptive text.

6. The document parsing method as described in claim 1, characterized in that, This method also includes: Identify whether a document image uses a multi-column layout and save the multi-column layout information; Extract the directory tree structure of the document image, which includes the organization information of chapters, sections and subheadings in the document image, as well as metadata of page numbers, document image titles and authors; The saved multi-column layout information and directory tree structure are used as layout and directory vectors for conversion. These vectors are then combined with the position, content, and category of the element information to reconstruct the document image for each page.

7. A document parsing method as described in claim 6, characterized in that, The step of converting the position, content, and category of element information into vector representations, and then using these vector representations to reconstruct the document image for each page, includes: The position, content, and category of element information, as well as multi-column layout information and directory tree structure, are input into the encoder for processing, and the position, content, and category of element information, as well as multi-column layout information and directory tree structure, are converted into vector representations. The vector representation is input into the layout restoration model. The layout of the document image is reconstructed by comprehensively considering the position, content and category of element information, as well as the relationship between multi-column layout information and directory tree structure. The text, images and element information are correctly arranged and organized in the restored document image.

8. A document parsing system, characterized in that, include: The acquisition module is used to obtain the initial representation of multi-page document images; The initial representation is the extracted text information; The document image reconstruction module combines the initial representation of each page's document image with a cascaded detection framework to perform refined layout analysis on each page. Specifically, during the analysis process, text region detection, table region detection, and image region detection are used to obtain the position, content, and category of each element within each page's document image. A cross-element relationship awareness mechanism is introduced to remove the overlap of elements within the document image, resulting in optimized elements. The position, content, and category information of the elements are encoded into dynamic vector representations, and based on an adaptive mechanism, the dynamic vector representations are dynamically adjusted for feature expression based on the document structure and content complexity. The document image is reconstructed page by page based on the dynamic vector representation adjusted by feature expression, and then structured parsing and semantic representation are performed; wherein the element categories include text, tables and images; The document image integration module is used to construct a cross-page matching label by combining the dynamic vector representation adjusted by the feature expression, and to integrate the contrastive learning method with the dynamic positive sample mining mechanism to distinguish similar elements on different pages and continuous content across pages. Through an end-to-end training framework, the cross-page matching relationship is learned. Based on the learned cross-page matching relationships, the reconstructed document images of each page are structurally integrated and semantically aligned, ultimately outputting high-quality, structurally coherent document parsing results.

9. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores a program that, when executed by the processor, causes the processor to perform the steps of a document parsing method as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of a document parsing method according to any one of claims 1 to 7.

Images