PDF extraction method and system based on deep learning and layout analysis

Abstract

The invention relates to a PDF extraction method and system based on deep learning and layout analysis, and the method comprises the steps: obtaining a PDF file, carrying out the element recognition of a scanning version file and a text version file through an OCR technology and a deep learning model, and enabling the elements to comprise a text, a picture, a table and a formula; dividing a page of the PDF file into a plurality of logic blocks according to the identified elements, and determining a hierarchical relationship and a sequence; deleting elements without substantial help from the identified elements, and extracting the content of the document; performing page layout analysis according to the document content, and reordering and organizing each logic block to obtain a layout; document contents are classified and analyzed, and then classified storage is carried out; and according to user requirements, splicing the classified and stored document contents in a corresponding layout to generate a document in a target format. Compared with the prior art, efficient and accurate PDF extraction is achieved, and meanwhile the important structure and content of an original document are reserved.

Description

Technical Field

[0001] The present invention relates to the field of PDF intelligent extraction, and in particular to a PDF extraction method and system based on deep learning and layout analysis. Background Art

[0002] In the field of intelligent PDF extraction, the application of deep learning technology is of great significance. Traditional PDF extraction methods often rely on rule-based or heuristic algorithms, which are incapable of processing PDF documents with complex layouts, diverse text formats, and mixed images and tables. Deep learning technology, especially convolutional neural networks (CNN), recurrent neural networks (RNN), and Transformer models, can automatically learn and understand complex patterns and structures in PDF documents, thereby achieving more accurate and intelligent content extraction.

[0003] Existing deep learning-based PDF extraction solutions have overcome the limitations of traditional methods to a certain extent. They train a large amount of PDF sample data to enable the model to automatically identify and parse various elements in the document, such as text blocks, images, tables, etc. These solutions have shown obvious advantages in processing complex layouts and diverse text formats, and the accuracy and efficiency of extraction have been significantly improved.

[0004] The core technologies of the existing deep learning-based PDF extraction solutions are:

[0005] Layout analysis: Use computer vision technology to analyze the layout of PDF documents and identify the position and size of elements such as text blocks, images, and tables.

[0006] Deep learning model: Use deep learning models such as convolutional neural networks (CNN) or Transformer to further process and understand the results of layout analysis and extract text content, images, and table data in the document.

[0007] Post-processing: Format and organize the extracted content to make it easier for subsequent application and processing.

[0008] Implementation methods include:

[0009] Data preprocessing: Convert PDF documents into sequences of images or text blocks as input to the deep learning model.

[0010] Model training: Use a large amount of labeled data to train the deep learning model so that it can accurately identify and parse various elements in PDF documents.

[0011] Extraction and post-processing: Apply the trained model to new PDF documents, extract the content, and perform necessary post-processing.

[0012] Application scenarios include:

[0013] Academic document extraction: Extract text, charts, and citation information from PDF documents such as academic papers and research reports.

[0014] Enterprise document management: Extract key information from PDF documents such as contracts, invoices, and reports to achieve automated document processing.

[0015] Digital library construction: Convert a large number of PDF book resources into searchable and editable text formats to facilitate the construction and management of digital libraries.

[0016] The disadvantages of the existing deep learning-based PDF extraction solutions are as follows:

[0017] 1. Insufficient segmentation function: When performing OCR recognition on a multi-column PDF, the text lines will horizontally span multiple vertical columns. The open source models in the existing technology do not have the ability to merge paragraphs across pages or columns when extracting PDF content. This results in the content being incorrectly segmented or omitted when extracting multi-page or multi-column documents, affecting the integrity and accuracy of the information.

[0018] 2. Lack of footnote recognition: Existing technologies cannot recognize footnotes in PDF, which may lead to the omission of important information. Footnotes usually contain supplementary explanations or references to the document content and are crucial for understanding the document content.

[0019] 3. Formula detection limitations: Although existing technologies can detect formulas in PDFs to a certain extent, the overall recall rate for long formula detection is low, only 0.6. This means that many important mathematical formulas may be ignored, affecting the completeness and accuracy of the document content.

[0020] 4. Reading order problem: Existing technologies usually process PDF content in a simple splicing manner, and do not support column processing. For example, text-based PDFs account for about 60-70% of the total, and currently only 20% of the total can be processed, which may lead to confusion in the reading order and affect the user experience.

[0021] 5. Document styles and formats vary greatly: papers, magazines, K12 textbooks, scanned versions of ancient texts, comics, and photo albums have different format characteristics.

[0022] 6. Limited multimodal support: Existing technologies mainly support the extraction of text content, and have limited support for multimodal content (such as pictures, tables, etc.). This limits the diversity and richness of information extracted from PDF.

[0023] In summary, the existing PDF intelligent extraction methods based on deep learning and layout analysis have achieved accurate extraction and intelligent understanding of PDF documents to a certain extent. However, these methods still have some limitations when processing certain types of PDF documents, such as the recognition accuracy of scanned PDFs and the ability to parse complex layouts. Summary of the invention

[0024] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a PDF extraction method and system based on deep learning and layout analysis to achieve accurate recognition and structured extraction of PDF document content.

[0025] The purpose of the present invention can be achieved by the following technical solutions:

[0026] A PDF extraction method based on deep learning and layout analysis includes the following steps:

[0027] Obtain a PDF file, determine whether the obtained PDF file is a scanned version or a text version, extract text using OCR technology for scanned version files to perform element recognition; for text version files, perform a comprehensive scan using a deep learning model to identify corresponding elements; the elements include text, images, tables, and formulas;

[0028] According to the identified elements, the pages of the PDF file are divided into multiple logical blocks, and the hierarchical relationship and order between the logical blocks are determined to obtain the division result of the layout;

[0029] Identify and delete the identified elements that are of no substantial help, and then extract information based on the deleted elements to obtain the document content of each logical block in the layout;

[0030] Performing page layout analysis based on the document contents of each logical block in the extracted layout, reordering and organizing each logical block to obtain the layout;

[0031] Classify and parse the document content of each logical block, and classify and store the document content according to the classification results;

[0032] According to the user's information output format requirements, the classified and stored document contents are spliced ​​with corresponding layouts to generate documents in the target format.

[0033] Furthermore, the logic blocks include title blocks, text blocks and chart blocks, and the hierarchical relationship and order between the logic blocks are determined according to the types and positions of the identified logic blocks.

[0034] Furthermore, the insubstantial help elements include advertisements, watermarks, and headers and footers.

[0035] Furthermore, the method also includes performing coordinate repair, image merging and table merging on the elements identified in the PDF file.

[0036] Furthermore, the method further includes identifying the cross-column and cross-page paragraphs of the elements identified in the PDF file, and combining the merging rules to realize the merging of the cross-column and cross-page paragraphs.

[0037] Furthermore, the recognition process of the element is specifically as follows: extracting text information through OCR technology or deep learning model, extracting image information through Markdown embedded image link technology, processing table information through table formatting technology, processing formula information through Latex conversion, and identifying layout information through Layout detection, so as to determine the hierarchical relationship and order between each logical block.

[0038] Furthermore, the page layout analysis is specifically as follows: according to the document content of each logical block in the extracted layout, a simple layout and a complex layout are distinguished. If it is a simple layout, subsequent processing is directly performed; if it is a complex layout, adaptive information extraction is performed to obtain logical and organized layout information.

[0039] Furthermore, the process of classifying the document content of each logical block is specifically as follows: classifying the document content of each logical block according to text, picture, table and formula as categories.

[0040] The present invention also provides a PDF extraction system based on the PDF extraction method based on deep learning and layout analysis as described above, comprising:

[0041] The pre-processing layer is used to extract metadata from the input PDF file, and preliminarily classify the PDF file according to the extracted metadata to identify whether it is a scanned version or a text version;

[0042] The model layer is used to identify elements in PDF files, including text, images, tables, and formulas. Specifically, the scanned files are extracted using OCR technology for element recognition, and the text files are fully scanned using a deep learning model; image information is extracted using Markdown embedded image link technology, table information is processed using table formatting technology, formula information is processed using Latex conversion, and layout information is identified through Layout detection to determine the hierarchical relationship and order between various logical blocks;

[0043] The pipeline layer is used to perform coordinate repair, image merging, and table merging on the element information identified by the model layer, so as to divide the pages of the PDF file into multiple logical blocks according to the final element information, and determine the hierarchical relationship and order between the logical blocks to obtain the layout division result; the identified elements are deleted without substantial help, and the complex page layout is filtered and paragraphs are merged to obtain the final layout and the document content of each corresponding logical block;

[0044] The output layer is used to classify and parse the document content of each logical block, and store the document content in an intermediate state according to the classification results; according to the user's information output format requirements, the classified and stored document contents are spliced ​​in the corresponding layout to generate a document in the target format for visual display;

[0045] The quality inspection layer is used to perform quality inspection and evaluation on the results of the output layer.

[0046] Furthermore, the quality inspection layer includes testing by inputting PDF files of papers, teaching materials, and test papers into the PDF extraction system, thereby performing quality inspection and evaluation.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] (1) The present invention combines the deep learning model and the layout analysis algorithm as a whole to achieve accurate recognition and structured extraction of PDF document content; the deep learning model is used to identify elements such as text, images, and tables in the document, while the layout analysis is responsible for parsing the spatial relationship between these elements to ensure that the converted Markdown document conforms to the human reading order and the original document layout.

[0049] (2) For the common cross-page or cross-column text in PDF documents, the present invention adopts advanced cross-page and cross-column segmentation technology. By combining rules and models, it intelligently identifies and merges paragraphs scattered on different pages or columns, realizes smooth splicing of paragraphs, and improves the accuracy and readability of the conversion results.

[0050] (3) The present invention supports multiple front-end model inputs and can automatically identify and process PDF files from different sources. At the same time, it has the ability to automatically identify and convert garbled PDF files, ensuring the robustness and wide applicability of the conversion process. For special content such as formulas, images, and tables, Latex conversion, Markdown embedded image links, and table formatting are used to maintain the richness and integrity of the original document. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic diagram of a flow chart of a PDF extraction method based on deep learning and layout analysis provided in an embodiment of the present invention;

[0052] Figure 2 A schematic diagram of a processing flow framework of a PDF extraction method based on deep learning and layout analysis provided in an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of the structure of a PDF extraction system based on deep learning and layout analysis provided in an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the layout detection results of a PDF extraction method based on deep learning and layout analysis provided in an embodiment of the present invention. The left side of the figure is the layout detection result, and the right side is the corresponding PDF document. The English fonts therein are only examples and have no effect on the main solution. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0056] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0058] Example 1

[0059] like Figure 1 and Figure 2 As shown, this embodiment provides a PDF extraction method based on deep learning and layout analysis, comprising the following steps:

[0060] S1: Obtain a PDF file, determine whether the obtained PDF file is a scanned version or a text version, extract text from the scanned version using OCR technology to perform element recognition; for the text version, perform a comprehensive scan using a deep learning model to identify the corresponding elements; the elements include text, images, tables, and formulas;

[0061] This means that for scanned documents, OCR technology is needed to extract text, while for text documents, the deep learning model is used to fully scan the PDF document and automatically identify various types of elements such as text, pictures, tables, formulas, etc. This internal self-developed model has been trained with a large number of samples and can accurately identify document content in different fonts, sizes and formats, laying a solid foundation for subsequent processing.

[0062] Specifically, text information is extracted through OCR technology or deep learning models, image information is extracted through Markdown embedded image link technology, table information is processed through table formatting technology, and formula information is processed through Latex conversion. The detection recall rate of long formulas is preferably improved to ensure the complete extraction of key information such as mathematical formulas; layout information is identified through Layout detection to determine the hierarchical relationship and order between various logical blocks;

[0063] Preferably, the above steps further include performing coordinate repair, image merging and table merging on the elements identified in the PDF file.

[0064] Preferably, the footnotes in the PDF file are accurately identified to ensure that important information is not missed. In the extraction process, auxiliary information such as footnotes, captions of figures and tables are specially processed to ensure the logical coherence and reading fluency of the extracted content.

[0065] S2: Divide the pages of the PDF file into multiple logical blocks according to the identified elements, and determine the hierarchical relationship and order between the logical blocks to obtain the division result of the layout;

[0066] This means that based on the identified elements, the system further analyzes the layout of the document, divides the page into multiple logical blocks (such as title blocks, text blocks, chart blocks, etc.), and determines the hierarchical relationship and order between these blocks. This step ensures a structured representation of the document content, which is convenient for subsequent processing and user understanding.

[0067] S3: Identify and delete the identified elements that are of no substantial help, and then extract information based on the deleted elements to obtain the document content of each logical block in the layout; the elements of no substantial help include advertisements, watermarks, and headers and footers;

[0068] The system has a built-in intelligent filtering mechanism that can automatically identify and delete elements that are not of substantial help to reading comprehension, such as advertisements, watermarks, headers and footers, etc. It can parse the pre-processed PDF files and initially extract the text content in the document to purify the document content. At the same time, the system will also adjust the fluency of the text to ensure the coherence of the content.

[0069] Optionally, use the discarded PDF labels to analyze the causes of each category and solve the problems in a targeted manner. Expand the scope of PDF parsing capabilities and increase the amount of corpus.

[0070] Preferably, the method further comprises identifying the cross-column and cross-page paragraphs of the elements identified in the PDF file, and combining the merging rules to realize the merging of the cross-column and cross-page paragraphs.

[0071] This means that for the common cross-page or cross-column text in PDF documents, the present invention adopts advanced cross-page and cross-column segmentation technology. By combining rules and models, it can intelligently identify and merge paragraphs scattered on different pages or columns, achieve smooth splicing of paragraphs, and improve the accuracy and readability of the conversion results.

[0072] That is, the model is used to replace the rule to identify the text order: after the layout detection, the OCR recognition is performed on the text boxes one by one. By combining the model and the rule method, the paragraphs across pages and columns can be merged to solve the problem of confusing the reading order of the text in the multi-column scanned PDF.

[0073] Semantic recognition can be performed for the first and last paragraphs of two adjacent pages or columns. If the paragraphs include mutually connected semantics, the paragraphs that span columns or pages can be merged.

[0074] S4: performing page layout analysis according to the document contents of each logical block in the extracted layout, reordering and organizing each logical block to obtain the layout;

[0075] Specifically, based on the document content of each logical block in the extracted layout, a distinction is made between simple layout and complex layout. If it is a simple layout, subsequent processing is performed directly. If it is a complex layout, adaptive information extraction is performed to obtain logical and organized layout information.

[0076] This is equivalent to analyzing the page layout of the extracted text based on the analysis results of the previous steps, reordering and organizing the document content, and distinguishing between simple and complex layouts. Simple layout pages are directly processed in the next step, while complex layout pages enter the adaptive extraction process. Ensure the logic and orderliness of the information. This step not only improves the readability of the document, but also facilitates subsequent applications such as search, indexing, and summary generation.

[0077] S5: classify and analyze the document contents of each logical block, and classify and store the document contents according to the classification results;

[0078] The process of classifying the document content of each logical block is specifically as follows: classifying the document content of each logical block according to text, picture, table and formula as categories.

[0079] S6: According to the user's information output format requirements, the classified and stored document contents are spliced ​​with corresponding layouts to generate a document in a target format.

[0080] This is equivalent to using technologies such as chart detection models and formula detection models to classify and analyze the extracted information. This ensures that all types of information can be accurately identified and stored in categories. Information fusion and output: The classified information is fused and processed to ensure the integrity and accuracy of the information. Finally, according to user needs, the processed information is spliced ​​in the original layout order to generate a document in the target format.

[0081] Example 2

[0082] like Figure 3 As shown, this embodiment provides a PDF extraction system based on a PDF extraction method based on deep learning and layout analysis as in Embodiment 1. The overall system architecture includes a preprocessing layer, a model layer, a pipeline layer, an output layer, and a quality inspection layer. Each layer carries specific functions and tasks, specifically:

[0083] The pre-processing layer is used to extract metadata from the input PDF file, and preliminarily classify the PDF file according to the extracted metadata to identify whether it is a scanned version or a text version;

[0084] Specifically, the preprocessing layer first extracts metadata to obtain metadata information of PDF documents. Subsequently, the document classification module will perform preliminary classification of documents and identify different types such as disordered and scanned versions, providing a basis for subsequent processing. In addition, the preprocessing layer also includes preprocessing operations such as coordinate repair, image merging, and table merging, aiming to optimize the document structure and improve the efficiency of subsequent processing.

[0085] The model layer is used to identify elements in PDF files, including text, images, tables, and formulas. Specifically, the scanned files are extracted using OCR technology for element recognition, and the text files are fully scanned using a deep learning model; image information is extracted using Markdown embedded image link technology, table information is processed using table formatting technology, formula information is processed using Latex conversion, and layout information is identified through Layout detection to determine the hierarchical relationship and order between various logical blocks;

[0086] Optionally, the model layer can be divided into a layout detection module, a formula detection module, and a text OCR module. The layout detection module is responsible for identifying the layout information in the PDF document, providing a basis for subsequent layout optimization and content extraction. The formula detection module focuses on identifying the formula information in the document to ensure the accurate extraction of these key contents. The text OCR module performs optical character recognition on the text in the document and converts it into an editable text format.

[0087] The pipeline layer is used to perform coordinate repair, image merging, and table merging on the element information identified by the model layer, so as to divide the pages of the PDF file into multiple logical blocks according to the final element information, and determine the hierarchical relationship and order between the logical blocks to obtain the layout division result; the identified elements are deleted without substantial help, and the complex page layout is filtered and paragraphs are merged to obtain the final layout and the document content of each corresponding logical block;

[0088] In other words, the pipeline layer is the data processing center of the system, responsible for deep processing of preprocessed documents. This layer includes multiple submodules such as coordinate repair, image merging, table merging, etc., which are designed to further optimize the document content and improve the accuracy and completeness of extraction. In addition, the pipeline layer also includes processing steps such as formula replacement, useless block removal, complex layout filtering, paragraph merging, etc. to ensure the clarity and readability of the output results.

[0089] The output layer is used to classify and parse the document content of each logical block, and store the document content in an intermediate state according to the classification results; according to the user's information output format requirements, the classified and stored document contents are spliced ​​in the corresponding layout to generate a document in the target format for visual display;

[0090] The intermediate state is output in the format of middle_json, which is convenient for subsequent processing and display. This format is not only easy to understand and use, but also can support a variety of output requirements, such as visualization, markdown, etc.

[0091] The quality inspection layer is used to perform quality inspection and evaluation on the results of the output layer.

[0092] Optionally, the quality inspection layer includes multiple modules, such as paper detection, textbook detection, test paper detection, etc., which can identify the specific type of document and perform corresponding quality assessment on it. In addition, the quality inspection layer also includes modules such as unit testing and visual quality inspection tools to ensure the accuracy and consistency of the output results.

[0093] It should be noted that the specific content and beneficial effects of the system of the present application can be found in the above-mentioned method embodiment, which will not be repeated here.

[0094] This solution can implement a closed-loop iteration strategy, which starts from pre-labeling, goes through manual correction and model update, and then returns to pre-labeling to form a complete closed loop. This strategy enables the layout detection model to significantly improve its detection capabilities through continuous learning and optimization.

[0095] This solution is suitable for English papers, Chinese books, and textbooks above high school, showing wide applicability. The output format supports multi-modal and NLP training formats, and the format can be expanded at will, providing flexible output options.

[0096] The layout detection model capabilities of this solution have been significantly improved, far exceeding the existing open source models and reaching a practical level.

[0097] like Figure 4 The examples of layout detection results demonstrate the excellent performance of the present invention in practical applications.

[0098] The formula detection results of three different models (Baseline, CnSTD paid version and Ours) under two evaluation indicators (AP50 and AR50) on the SCIHub evaluation set and the textbook evaluation set are compared, as shown in Table 1. The results show that the performance of Ours model on both evaluation sets is significantly better than that of Baseline and CnSTD paid version.

[0099] Table 1

[0100]

[0101] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A PDF extraction method based on deep learning and layout analysis, characterized in that: The following steps are involved: Obtain a PDF file, determine whether the obtained PDF file is a scanned version or a text version, extract text using OCR technology for scanned version files to perform element recognition; for text version files, perform a comprehensive scan using a deep learning model to identify corresponding elements; the elements include text, images, tables, and formulas; According to the identified elements, the pages of the PDF file are divided into multiple logical blocks, and the hierarchical relationship and order between the logical blocks are determined to obtain the division result of the layout; Identify and delete the identified elements that are of no substantial help, and then extract information based on the deleted elements to obtain the document content of each logical block in the layout; Performing page layout analysis based on the document contents of each logical block in the extracted layout, reordering and organizing each logical block to obtain the layout; Classify and parse the document content of each logical block, and classify and store the document content according to the classification results; According to the user's information output format requirements, the classified and stored document contents are spliced ​​with corresponding layouts to generate documents in the target format.

2. A PDF extraction method based on deep learning and layout analysis according to claim 1, characterized in that: The logic blocks include title blocks, text blocks and chart blocks. The hierarchical relationship and order between the logic blocks are determined according to the identified types and positions of the logic blocks.

3. A PDF extraction method based on deep learning and layout analysis according to claim 1, characterized in that: The non-essential help elements include advertisements, watermarks, and headers and footers.

4. A PDF extraction method based on deep learning and layout analysis according to claim 1, characterized in that: The method also includes performing coordinate repair, image merging and table merging on the elements identified in the PDF file.

5. The PDF extraction method based on deep learning and layout analysis according to claim 1, characterized in that: The method further includes identifying cross-column and cross-page paragraphs for the elements identified in the PDF file, and combining the merging rules to realize the merging of the cross-column and cross-page paragraphs.

6. A PDF extraction method based on deep learning and layout analysis according to claim 1, characterized in that: The element recognition process is specifically as follows: extracting text information through OCR technology or deep learning model, extracting image information through Markdown embedded image link technology, processing table information through table formatting technology, processing formula information through Latex conversion, and identifying page layout information through Layout detection, so as to determine the hierarchical relationship and order between each logical block.

7. A PDF extraction method based on deep learning and layout analysis according to claim 1, characterized in that: The page layout analysis is specifically as follows: based on the document content of each logical block in the extracted layout, a simple layout and a complex layout are distinguished. If it is a simple layout, subsequent processing is directly performed. If it is a complex layout, adaptive information extraction is performed to obtain layout information that is logical and organized.

8. The PDF extraction method based on deep learning and layout analysis according to claim 1, characterized in that: The process of classifying the document content of each logical block is specifically as follows: classifying the document content of each logical block according to text, picture, table and formula as categories.

9. A PDF extraction system based on a PDF extraction method based on deep learning and layout analysis as claimed in any one of claims 1 to 8, characterized in that: include: The pre-processing layer is used to extract metadata from the input PDF file, and preliminarily classify the PDF file according to the extracted metadata to identify whether it is a scanned version or a text version; The model layer is used to identify elements in PDF files, including text, images, tables, and formulas. Specifically, the scanned files are extracted using OCR technology for element recognition, and the text files are fully scanned using a deep learning model; image information is extracted using Markdown embedded image link technology, table information is processed using table formatting technology, formula information is processed using Latex conversion, and layout information is identified through Layout detection to determine the hierarchical relationship and order between various logical blocks; The pipeline layer is used to perform coordinate repair, image merging, and table merging on the element information identified by the model layer, so as to divide the pages of the PDF file into multiple logical blocks according to the final element information, and determine the hierarchical relationship and order between the logical blocks to obtain the layout division result; the identified elements are deleted without substantial help, and the complex page layout is filtered and paragraphs are merged to obtain the final layout and the document content of each corresponding logical block; The output layer is used to classify and parse the document content of each logical block, and store the document content in an intermediate state according to the classification results; according to the user's information output format requirements, the classified and stored document contents are spliced ​​in the corresponding layout to generate a document in the target format for visual display; The quality inspection layer is used to perform quality inspection and evaluation on the results of the output layer.

10. The system according to claim 9, characterized in that The quality inspection layer includes testing by inputting PDF files of papers, teaching materials, and test papers into the PDF extraction system, thereby performing quality inspection and evaluation.

Images