Multi-information multi-round pdf document paragraph searching and positioning method based on numerical assistance

Abstract

The invention discloses a multi-information multi-round pdf document paragraph searching and positioning method based on numerical assistance. The method aims at solving the problems that efficiency is low and positioning is inaccurate when bidding documents are manually audited in the prior art. The method comprises the following steps: firstly, analyzing a PDF document, and extracting text, table and image information; constructing a candidate page number set by identifying chapter title keywords, and performing interval merging; automatic positioning of search starting and ending page numbers is realized by utilizing a numerical relationship between chapter serial numbers and subsequent serial numbers of the chapter serial numbers; aggregating directory page number information to accurately position a target interval; and finally, image processing and optical character recognition are carried out in the target page, and extraction of specific information such as business licenses and qualification certificates is completed. According to the method, the searching precision and efficiency of the target information in the PDF document are remarkably improved, and the method is particularly suitable for automatic review of bidding and tendering documents.

Description

Technical Field

[0001] This invention belongs to the field of text processing technology, and in particular relates to a method for searching and locating paragraphs in PDF documents based on numerical assistance and multiple information rounds. Background Technology

[0002] In the bidding process, identifying and locating the qualification review attachments of different bidders from the tender documents is an important and complex task. Traditional search and location methods often rely on manual review, which is inefficient, highly subjective, and prone to inconsistencies and inaccuracies between the identified qualification attachments and the actual qualification requirements. Furthermore, with increasing market competition, companies need to identify suitable bidders more quickly and accurately to improve their success rate and reduce risk.

[0003] In practical work, unstructured world knowledge appears more frequently and abundantly than structured world knowledge. Only by effectively parsing this massive amount of unstructured data can large models achieve better performance. PDF documents, as carriers of stored and retrieved knowledge, make the exploration of how to extract effective unstructured information an indispensable part of enhancing large model retrieval. This paper proposes a numerically assisted, multi-information, multi-round PDF paragraph search and localization algorithm to effectively integrate bidding documents, such as bid letter seals and signatures, business licenses, project members, and company performance. Therefore, a numerically assisted, multi-information, multi-round PDF document paragraph search and localization method is needed to solve the above problems. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a numerically assisted, multi-information, multi-round PDF document paragraph search and positioning method, which aims to solve the problems of low efficiency and inaccurate positioning in traditional manual review of tender documents.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A numerically assisted, multi-information, multi-round PDF document paragraph search and location method includes the following steps: S1, PDF document content analysis: Use the pdfminer library to extract text, table, and image attributes from a full PDF document and save the extracted information locally. S2, Construct a candidate page number set based on interval merging: Based on the parsed text information, chapter titles containing predefined keywords are identified, and their corresponding start and end page numbers are extracted to form an initial interval; an interval merging algorithm is performed on the initial interval to obtain the merged page number interval and add it to the candidate page number set; S3, Numerical Assisted Positioning: Based on the chapter number where the target information is located, determine the starting page number of the search; based on the next chapter number of the aforementioned chapter number, determine the ending page number of the search, thereby forming a numerically assisted positioning range; S4, Multi-information Aggregation: Obtain the page number corresponding to the target information in the directory, and aggregate it with the numerical auxiliary positioning range to update the final search page number range; S5, Image Text Recognition: Image processing and optical character recognition are performed on the pages covered by the final search page range and the candidate page set to locate and extract the target information; S6, Information Location and Extraction: First, the text content is extracted using an OCR engine. Then, the semantic entity recognition module obtains the semantic entities in the image. Finally, the relationship extraction module obtains the correspondence between the semantic entities, thereby extracting the key information needed.

[0006] Preferably, step S1 specifically includes: Extracting titles from a PDF: Use regular expressions to restrict text attributes containing Chinese numerals or Arabic numerals, thereby determining whether the current text attribute belongs to the title content; Extracting tables from a PDF: The table extraction function based on pdfminer is used to extract the table of the current page into an array. For tables that span multiple pages, it is analyzed whether the first column of the current page table contains bidding keywords. If it does, it is treated as a new table; if it does not, the current table is merged with the last table of the previous page. Extracting images from a PDF: Use pdfminer to traverse each page element of the PDF document. If the current element's attribute is LTFigure, save it as an image. Identifying signatures and seals in PDFs: The page of the tender document should correspond to the legal representative's signature or the company seal.

[0007] Preferably, the identification of signatures and seals in a PDF specifically includes: First, use the pdf2image library to convert the tender document page into an image: For seal data: Use the seal detection OCR interface to determine whether a company seal exists on the current page; For the company's legal representative's signature: First, use OCR technology and regular expressions to locate the left and right boundaries of the signature keywords; after saving the fixed-boundary rectangle on the page, use an algorithm to calculate the proportion of black pixels in the fixed rectangle. The specific steps are as follows: A threshold is set. If the proportion of black pixels in a fixed rectangular frame exceeds the threshold, it is determined that a legal person's signature exists in the frame. First, the original image is converted to a grayscale image. Then, the grayscale image is debinarized to calculate the proportion of black pixels to the total number of pixels in order to evaluate the pixel ratio.

[0008] Preferably, the interval merging algorithm in step S2 includes: Let the interval array be The merged interval array is ; Sort all initial intervals by their starting page number ; Iterate through the sorted intervals sequentially. If the current interval overlaps with a merged interval, extend the end page number of the merged interval; otherwise, store the merged interval in the result set and start a new merge with the current interval. After the traversal is complete, the final merged interval is added to the candidate page number set.

[0009] Preferably, step S3, which involves determining the start and end page numbers of the search based on the chapter number, specifically includes: Establish a mapping dictionary between Chinese numeric serial numbers and their successor serial numbers; Based on the chapter number, locate the first occurrence of the text information in the parsed text information to determine the starting page number; Based on the mapping dictionary, the next chapter number corresponding to the chapter number is calculated, and the position of the first occurrence of the next chapter number is located in the parsed text information to determine the end page number.

[0010] Preferably, the rule for calculating the next chapter number includes: If the last digit of the current chapter number is "nine", then the next chapter number is the first digit of the current chapter number plus "ten". If the last digit of the current chapter number is "ten", then the next chapter number will be the current chapter number plus "one". If the last digit of the current chapter number is any other number, then the next chapter number is the remaining part of the current chapter number after removing the last digit, plus the successor value of the last digit in the mapping dictionary.

[0011] Preferably, the aggregation operation in step S4 is as follows: taking the maximum value between the directory page number and the starting page number of the numerical auxiliary positioning interval as the starting page number of the final search interval; and taking the ending page number of the numerical auxiliary positioning interval as the ending page number of the final search interval.

[0012] Preferably, the image processing in step S5 includes: separating the color channels of the image, extracting the red channel and performing thresholding to remove stamp interference, and then performing optical character recognition.

[0013] Preferably, a computer device includes a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the numerically assisted multi-information multi-round PDF document paragraph search and location method.

[0014] Preferably, a computer-readable storage medium stores computer instructions that cause a computer to execute the numerically assisted, multi-information, multi-round PDF document paragraph search and location method.

[0015] The beneficial effects of this invention are as follows: 1. Improved search accuracy: Numerical assistance can be used to more accurately match the location of the extracted information in the PDF, reducing interference from irrelevant results.

[0016] 2. Enhanced user experience: Quickly locates relevant paragraphs or sentences, saving users time browsing the entire text and making the information acquisition process more efficient and convenient.

[0017] 3. High adaptability: It works effectively not only with plain text PDF files, but also with academic papers containing charts, formulas, and other similar content. 4. Cross-language capability: In theory, the algorithm parameters can be adjusted to support document search in different language versions, thus expanding the scope of application.

[0018] 5. Good scalability: With the development of technology and the emergence of new demands, this method can be easily expanded and upgraded in terms of functionality, such as integrating a natural language understanding module to improve the level of semantic understanding. Attached Figure Description

[0019] Figure 1 This is a module diagram of the multi-information, multi-round PDF document paragraph search and positioning method based on numerical assistance in this embodiment of the invention; Figure 2 This is a flowchart of the multi-information, multi-round PDF document paragraph search and positioning method based on numerical assistance in an embodiment of the present invention; Figure 3This is an example diagram of the candidate page number set construction algorithm based on interval merging in an embodiment of the present invention. Detailed Implementation

[0020] Example 1: like Figures 1-3 As shown, a numerically assisted, multi-information, multi-round PDF document paragraph search and location method includes the following steps: S1, PDF document content analysis: Use the pdfminer library to extract text, table, and image attributes from a full PDF document and save the extracted information locally. S2, Construct a candidate page number set based on interval merging: Based on the parsed text information, chapter titles containing predefined keywords are identified, and their corresponding start and end page numbers are extracted to form an initial interval; an interval merging algorithm is performed on the initial interval to obtain the merged page number interval and add it to the candidate page number set; S3, Numerical Assisted Positioning: Based on the chapter number where the target information is located, determine the starting page number of the search; based on the next chapter number of the aforementioned chapter number, determine the ending page number of the search, thereby forming a numerically assisted positioning range; S4, Multi-information Aggregation: Obtain the page number corresponding to the target information in the directory, and aggregate it with the numerical auxiliary positioning range to update the final search page number range; S5, Image Text Recognition: Image processing and optical character recognition are performed on the pages covered by the final search page range and the candidate page set to locate and extract the target information; S6, Information Location and Extraction: First, the text content is extracted using an OCR engine. Then, the semantic entity recognition module obtains the semantic entities in the image. Finally, the relationship extraction module obtains the correspondence between the semantic entities, thereby extracting the key information needed.

[0021] Preferably, step S1 specifically includes: Extracting titles from a PDF: Use regular expressions to restrict text attributes containing Chinese numerals or Arabic numerals, thereby determining whether the current text attribute belongs to the title content; Extracting tables from a PDF: The table extraction function based on pdfminer is used to extract the table of the current page into an array. For tables that span multiple pages, it is analyzed whether the first column of the current page table contains bidding keywords. If it does, it is treated as a new table; if it does not, the current table is merged with the last table of the previous page. Extracting images from a PDF: Use pdfminer to traverse each page element of the PDF document. If the current element's attribute is LTFigure, save it as an image. Identifying signatures and seals in PDFs: The page of the tender document should correspond to the legal representative's signature or the company seal.

[0022] Preferably, the identification of signatures and seals in a PDF specifically includes: First, use the pdf2image library to convert the tender document page into an image: For seal data: Use the seal detection OCR interface to determine whether a company seal exists on the current page; For the company's legal representative's signature: First, use OCR technology and regular expressions to locate the left and right boundaries of the signature keywords; after saving the fixed-boundary rectangle on the page, use an algorithm to calculate the proportion of black pixels in the fixed rectangle. The specific steps are as follows: A threshold is set. If the proportion of black pixels in a fixed rectangular frame exceeds the threshold, it is determined that a legal person's signature exists in the frame. First, the original image is converted to a grayscale image. Then, the grayscale image is debinarized to calculate the proportion of black pixels to the total number of pixels in order to evaluate the pixel ratio.

[0023] Preferably, the interval merging algorithm in step S2 includes: Let the interval array be The merged interval array is ; Sort all initial intervals by their starting page number ; Iterate through the sorted intervals sequentially. If the current interval overlaps with a merged interval, extend the end page number of the merged interval; otherwise, store the merged interval in the result set and start a new merge with the current interval. After the traversal is complete, the final merged interval is added to the candidate page number set.

[0024] Preferably, step S3, which involves determining the start and end page numbers of the search based on the chapter number, specifically includes: Establish a mapping dictionary between Chinese numeric serial numbers and their successor serial numbers; Based on the chapter number, locate the first occurrence of the text information in the parsed text information to determine the starting page number; Based on the mapping dictionary, the next chapter number corresponding to the chapter number is calculated, and the position of the first occurrence of the next chapter number is located in the parsed text information to determine the end page number.

[0025] Preferably, the rule for calculating the next chapter number includes: If the last digit of the current chapter number is "nine", then the next chapter number is the first digit of the current chapter number plus "ten". If the last digit of the current chapter number is "ten", then the next chapter number will be the current chapter number plus "one". If the last digit of the current chapter number is any other number, then the next chapter number is the remaining part of the current chapter number after removing the last digit, plus the successor value of the last digit in the mapping dictionary.

[0026] Preferably, the aggregation operation in step S4 is as follows: taking the maximum value between the directory page number and the starting page number of the numerical auxiliary positioning interval as the starting page number of the final search interval; and taking the ending page number of the numerical auxiliary positioning interval as the ending page number of the final search interval.

[0027] Preferably, the image processing in step S5 includes: separating the color channels of the image, extracting the red channel and performing thresholding to remove stamp interference, and then performing optical character recognition.

[0028] Preferably, a computer device includes a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the numerically assisted multi-information multi-round PDF document paragraph search and location method.

[0029] Preferably, a computer-readable storage medium stores computer instructions that cause a computer to execute the numerically assisted, multi-information, multi-round PDF document paragraph search and location method.

[0030] Example 2: This embodiment provides a practical process for applying this method, and the steps are as follows: Step 1: PDF document content parsing. Use the pdfminer library to extract text (including title, body text, etc.), table, and image attributes from the entire PDF document, and save the extracted information locally.

[0031] Next, the titles in the PDF are extracted. Regular expressions are used to qualify text attributes containing Chinese or Arabic numerals to determine whether the current text attribute belongs to the title content.

[0032] Next, the tables in the PDF are extracted. The table extraction function based on pdfminer is used to extract the table on the current page into an array. For tables spanning multiple pages, it is analyzed whether the first column of the current page table contains bidding keywords. If it does, it is treated as a new table; otherwise, the current table is merged with the last table on the previous page.

[0033] Next, the images in the PDF are extracted. Using pdfminer, each page element of the PDF document is traversed, and if the current element's attribute is LTFigure, it is saved as an image.

[0034] Next, we will identify the signatures and seals in the PDF. The tender document's bid letter page needs to be signed by the company's legal representative or stamped with the company seal. We will use the pdf2image library to convert the bid letter page into an image.

[0035] Next, for the seal data, seal detection and recognition technology is used to determine whether a company seal exists on the current page and to identify the seal content.

[0036] Next, for the company's legal representative's signature, OCR technology and regular expressions are used to locate the left and right boundaries of the signature keywords (e.g., "Legal Representative:" is the left boundary, and "Legal Representative CA Electronic Seal" is the right boundary). After saving the fixed-boundary rectangle image on the page, the algorithm is used to calculate the proportion of black pixels in the fixed rectangle image. The specific operations are as follows: 1) Set a threshold. If the proportion of black pixels in the fixed rectangle image exceeds the threshold, it is determined that there is a legal representative's signature in the image; 2) First, convert the original image to a grayscale image, and then perform debinarization on the grayscale image to calculate the proportion of black pixels to the total number of pixels in order to evaluate the pixel proportion.

[0037] Step 2: Based on the candidate page number set of interval merging, since some material content is located in different chapter intervals, a preliminary candidate page number set E is constructed for initial screening. Taking business license information as an example, in different tender documents, its location may be in different chapters such as qualification review materials, other materials, attachments, and photocopies. Therefore, regular expressions are first used to perform initial screening on the title information and body text information obtained after parsing the PDF. If it contains possible keywords (qualification review materials, other materials, attachments, photocopies, etc.), its page number and the next level page number are extracted to construct an interval [s, e]. After this step is completed, all the currently extracted intervals are merged, and the page numbers of the merged intervals are added to the candidate page number set E. The interval merging algorithm is as follows: Let the interval array be The merged interval array is .

[0038] First, sort the array by its first and last elements: ; Then, initialize the boundaries: l = -1, r = -1; Iterate through the sorted interval array and get the current values ​​s and e. if l == -1 and r == -1; Initialize l = s, r = e; ; Update r = e; ; Add the current l and r to ; Update l = s, r = e; Step 3: Numerical-assisted positioning.

[0039] The search starts by determining the page number based on the chapter number, and then ends by using the page number from the next chapter. For example, if the required information is located in a fixed paragraph, such as "IX. Qualification Review Materials", the search will automatically locate the starting page number in the title and body text of the PDF after parsing the page number "IX". Similarly, the search will automatically locate the ending page number in the title and body text of the next chapter after parsing the page number "X".

[0040] Next, identify the keywords at the next chapter level that need to be analyzed in the tender documents. Certain qualifications and requirements are often found in specific paragraphs within the tender document. For example, information such as business licenses and qualification certificates may be found in sections like "Qualification Review Materials," "Other Materials," "Attachments," and "Photocopies." Therefore, when locating information such as business licenses and qualification certificates, you can analyze the page range of these sections within the "Qualification Review Materials," "Other Materials," "Attachments," and "Photocopies" sections to narrow down your search.

[0041] Next, a numerical-assisted algorithm is used to automatically determine the search keywords for the next chapter. For chapters one through N (where N is an uppercase number), a dictionary Dict is created to store the information for the next chapter: {'one': 'two', ...}. For chapters that require a carry-over, a check is performed on the last digit.

[0042] If the last digit of the current value current_index is 'nine', then the search keyword for the next chapter is Dict[current_index[0]] + 'ten'; If the last digit of the current value current_index is 'ten', then the search keyword for the next chapter is current_index + 'one'; If the last element of the current_index is any of the other key values ​​in Dict, then the search keyword for the next chapter is current_index[:-1] + Dict[current_index[-1]].

[0043] For example, if the current chapter number is six, then the last character of the search keyword for the next chapter is seven; if the current chapter number is twenty-three, then the last character of the search keyword for the next chapter is four. Step 4: Multi-information aggregation. While using numerical-assisted localization, parse the page number corresponding to the target information in the directory, denoted as page_x0. Assume the page number range for numerical-assisted localization is [page_x1, page_y1]. Update the search range to [x, y] using the parsed page_x0 from the directory, where x = max(page_x0, page_x1), y = page_y1. For example, parsing the page number 97 corresponding to the supplier's business license information in the directory, the supplier's business license's page number range for numerical-assisted localization is [93, 98]. Therefore, the location range of the supplier's business license is [97, 98].

[0044] Step 5: Image text recognition. Traverse the search start page number, search end page number, and candidate page number set E after narrowing down the search scope, and use the pdf2image library to obtain the image of the current page.

[0045] The process involves separating the channels of the current image, extracting the red channel, removing the red stamp using a threshold, and then using OCR or PaddlePaddle-based table OCR technology to recognize the information on the current page.

[0046] Next, we will use PaddlePaddle's table OCR for recognition.

[0047] The overall architecture of the table OCR uses SLANet-LCNetV2, which improves accuracy by 0.8% and inference speed by 3 times compared to the PP-Structure table recognition model. The optimization approach is as follows: Improved reasoning process, stopping at EOS, resulting in a 3x speed increase; Upgrade Backbone to LCNetV2; Row and column feature enhancement module; The resolution was increased from 488 to 512. Three-stage training strategy; Next, PP-Structure is used to recognize the table in the image.

[0048] Next, image correction technology is used to determine the orientation of the image, determine the overall orientation of the image, and straighten it. Next, layout information analysis is performed. For the layout analysis task, the image is divided into different regions such as text, tables, and images by the layout analysis model. Then, these regions are identified separately: the table region is sent to the table recognition module for structured recognition; the text region is sent to the OCR engine for text recognition; and the layout restoration module is used to restore it into a Word or PDF document with the same layout as the original image.

[0049] Step 6: Information Location and Extraction. First, an OCR engine is used to extract the text content. Then, the semantic entity recognition module obtains the semantic entities in the image. Finally, the relationship extraction module obtains the correspondence between semantic entities, thereby extracting the required key information. For example, for a supplier's business license, the Prompt technology in the large model is used to extract the company name, validity period, business scope, unified credit code, etc., from the business license attachments. For qualification certificates, the Prompt technology in the large model is used to extract the name, validity period, certificate number, seal information, etc., from the qualification certificate attachments.

Claims

1. A numerically assisted, multi-information, multi-round PDF document paragraph search and location method, characterized in that, Includes the following steps: S1, PDF document content analysis: Use the pdfminer library to extract text, table, and image attributes from a full PDF document and save the extracted information locally. S2, Construct a candidate page number set based on interval merging: Based on the parsed text information, chapter titles containing predefined keywords are identified, and their corresponding start and end page numbers are extracted to form an initial interval; an interval merging algorithm is performed on the initial interval to obtain the merged page number interval and add it to the candidate page number set; S3, Numerical Assisted Positioning: Based on the chapter number where the target information is located, determine the starting page number of the search; based on the next chapter number of the aforementioned chapter number, determine the ending page number of the search, thereby forming a numerically assisted positioning range; S4, Multi-information Aggregation: Obtain the page number corresponding to the target information in the directory, and aggregate it with the numerical auxiliary positioning range to update the final search page number range; S5, Image Text Recognition: Image processing and optical character recognition are performed on the pages covered by the final search page range and the candidate page set to locate and extract the target information; S6, Information Location and Extraction: First, the text content is extracted using an OCR engine. Then, the semantic entity recognition module obtains the semantic entities in the image. Finally, the relationship extraction module obtains the correspondence between the semantic entities, thereby extracting the key information needed.

2. The method for searching and locating paragraphs in a PDF document based on numerical assistance and multiple rounds as described in claim 1, characterized in that, Step S1 specifically includes: Extracting titles from a PDF: Use regular expressions to restrict text attributes containing Chinese numerals or Arabic numerals, thereby determining whether the current text attribute belongs to the title content; Extracting tables from a PDF: The table extraction function based on pdfminer is used to extract the table of the current page into an array. For tables that span multiple pages, it is analyzed whether the first column of the current page table contains bidding keywords. If it does, it is treated as a new table; if it does not, the current table is merged with the last table of the previous page. Extracting images from a PDF: Use pdfminer to traverse each page element of the PDF document. If the current element's attribute is LTFigure, save it as an image. Identifying signatures and seals in PDFs: The page of the tender document should correspond to the legal representative's signature or the company seal.

3. The method for searching and locating paragraphs in a PDF document based on numerical assistance and multiple rounds, as described in claim 2, is characterized in that... The identification of signatures and seals in PDFs specifically includes: First, use the pdf2image library to convert the tender document page into an image: For seal data: Use the seal detection OCR interface to determine whether a company seal exists on the current page; For the company's legal representative's signature: First, use OCR technology and regular expressions to locate the left and right boundaries of the signature keywords; after saving the fixed-boundary rectangle on the page, use an algorithm to calculate the proportion of black pixels in the fixed rectangle. The specific steps are as follows: A threshold is set. If the proportion of black pixels in a fixed rectangular frame exceeds the threshold, it is determined that a legal person's signature exists in the frame. First, the original image is converted to a grayscale image. Then, the grayscale image is debinarized to calculate the proportion of black pixels to the total number of pixels in order to evaluate the pixel ratio.

4. The method for searching and locating paragraphs in a PDF document based on numerical assistance and multiple rounds as described in claim 1, characterized in that, The interval merging algorithm described in step S2 includes: Let the interval array be The merged interval array is ; Sort all initial intervals by their starting page number ; Iterate through the sorted intervals sequentially. If the current interval overlaps with a merged interval, extend the end page number of the merged interval; otherwise, store the merged interval in the result set and start a new merge with the current interval. After the traversal is complete, the final merged interval is added to the candidate page number set.

5. The method for searching and locating paragraphs in a PDF document based on numerical assistance and multiple rounds as described in claim 1, characterized in that, Step S3, which involves determining the start and end page numbers of the search based on the chapter number, specifically includes: Establish a mapping dictionary between Chinese numeric serial numbers and their successor serial numbers; Based on the chapter number, locate the first occurrence of the text information in the parsed text information to determine the starting page number; Based on the mapping dictionary, the next chapter number corresponding to the chapter number is calculated, and the position of the first occurrence of the next chapter number is located in the parsed text information to determine the end page number.

6. The method for searching and locating paragraphs in a PDF document based on numerical assistance and multiple rounds as described in claim 5, characterized in that, The rules for calculating the next chapter number include: If the last digit of the current chapter number is "nine", then the next chapter number is the first digit of the current chapter number plus "ten". If the last digit of the current chapter number is "ten", then the next chapter number will be the current chapter number plus "one". If the last digit of the current chapter number is any other number, then the next chapter number is the remaining part of the current chapter number after removing the last digit, plus the successor value of the last digit in the mapping dictionary.

7. The method for searching and locating paragraphs in a PDF document based on numerical assistance and multiple rounds as described in claim 1, characterized in that, The aggregation operation in step S4 is as follows: take the maximum value between the directory page number and the starting page number of the numerical auxiliary positioning interval, and use it as the starting page number of the final search interval; use the ending page number of the numerical auxiliary positioning interval as the ending page number of the final search interval.

8. The method for searching and locating paragraphs in a PDF document based on numerical assistance and multiple rounds as described in claim 1, characterized in that, The image processing in step S5 includes: separating the color channels of the image, extracting the red channel and removing stamp interference through thresholding, and then performing optical character recognition.

9. A computer device, characterized in that, It includes a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the numerically assisted, multi-information, multi-round PDF document paragraph search and location method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute any one of claims 1 to 8, a numerically assisted, multi-information, multi-round PDF document paragraph search and location method.

Images