A method and system for intelligent analysis of complex format PDF documents and index deviation analysis supporting private domain closed loop deployment
By employing directory-driven local image processing and sliding window modeling, combined with a private domain closed-loop mechanism, the problems of parsing complex PDF documents and analyzing indicator deviations are solved, achieving efficient and secure document parsing and analysis report generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHENJIANG NO 1 PEOPLES HOSPITAL
- Filing Date
- 2026-03-31
- Publication Date
- 2026-06-26
Smart Images
Figure CN122287602A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent document processing and data analysis, specifically to a method and system for intelligent parsing and indicator deviation analysis of complex PDF documents that supports closed-loop deployment in private domains. Background Technology
[0002] In applications such as quality management, professional quality control management, and industry statistical analysis, relevant regulatory agencies or data publishing organizations typically publish reports containing statistical indicators and analytical information in complex PDF formats on a regular basis.
[0003] These types of documents generally have the following characteristics: They typically have many pages and a complex structure, possibly including a table of contents, sections for multiple professional fields, and spreadsheets or tables spanning multiple pages; they usually cover multiple professional fields, each corresponding to multiple key indicators. For example, in a medical quality management scenario, documents may be divided according to different clinical specialties, with each specialty containing multiple key quality indicators; the documents usually do not cover all indicators in the business data platform, but rather select some key indicators from indicator systems across multiple professional fields and provide corresponding benchmark values, such as provincial averages, regional averages, or preset benchmark values; the user organization needs to extract the professional category, indicator name, and corresponding benchmark values from these PDF documents and compare them with the corresponding indicator data in their own business data platform to identify indicator deviations and dynamically generate comprehensive analytical reports.
[0004] Existing solutions typically handle such documents in the following ways: First, directly extract the full text from the PDF or perform OCR; second, send the entire document into a large model for processing at once; third, recognize each page separately and then stitch them together.
[0005] The above methods have the following shortcomings: Traditional full-text extraction or OCR is poorly adapted to complex layouts, tables, mixed text and images, and cross-page content. Therefore, there are several very clear problems: it is difficult to efficiently locate the table of contents, multiple professional field section pages, and indicator pages of complex PDF documents; it is difficult to reliably extract complex tables and cross-page content; feeding the entire document to a large model results in a large data exposure surface and high processing costs; in application scenarios with high requirements for data security and privacy protection, such as medical quality management, existing technologies generally lack a closed-loop processing mechanism for document parsing, data extraction, and indicator analysis that supports local deployment or private cloud deployment; and it is difficult to automatically map the benchmark values extracted from complex PDF documents with the corresponding indicator values in the business data platform and generate deviation result analysis reports.
[0006] Therefore, since complex PDF documents need to undergo controlled image processing, directory-driven positioning, sliding window modeling, and private domain closed-loop processing without using full-text OCR, it is necessary to propose new technical solutions to accurately extract key indicator benchmark values and then conduct deviation analysis between the benchmark values and the business data platform values in a natural and reasonable manner. Summary of the Invention
[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system for intelligent parsing and indicator deviation analysis of complex-formatted PDF documents that supports private domain closed-loop deployment. It adopts an overall approach of "directory-driven local image processing + sliding window cross-page modeling + minimum exposure inference under private domain closed loop + platform indicator mapping comparison," rather than performing full-text OCR on the entire PDF document or submitting the entire document at once. This achieves a complete technical process from parsing complex-formatted PDF documents to generating deviation analysis reports.
[0008] The technical solution of this invention to solve the above problems is: a method for intelligent parsing and indicator deviation analysis of complex PDF documents that supports closed-loop deployment in private domains, specifically including the following steps: Step S1, PDF report document access and verification: Access the PDF report document to be processed through file interface, batch import method or automatic access method based on file monitoring mechanism. After accessing the file, verify the file format, file readability and integrity. Step S2, Secure storage of PDF documents in private domain: Securely store and uniformly manage verified PDF documents in the private domain deployment environment; Step S3: Scan the candidate pages of the directory and generate images of the candidate pages of the directory; Step S4: Directory identification and extended scanning; Step S5: Generate structured directory entries, generating the structured directory entry results; Step S6: Professional Page Range Determination: Determine the physical page range corresponding to each professional field based on the directory entry results; Step S7: Professional Page Image Generation: Based on the determined professional page range, perform page image processing only on the physical page range corresponding to each professional field to generate a professional page image sequence; Step S8: Sliding window image segmentation: The professional page image sequence is segmented into multiple overlapping image windows for model parsing by sliding window segmentation according to the preset window length and step size; Step S9: Sensitive Area Desensitization Processing: Before the model is sent into the task window, sensitive areas in the image window are desensitized. Step S10: Construct a multimodal model input and call the local model or cloud-based private model; according to the preset deployment strategy, route the desensitized image window data to the local model service or cloud-based private model service for inference processing; Step S11: Extract structured results; convert the window image data into the model input format, combine it with preset task prompts to construct a multimodal model input, and send it to the multimodal model for inference processing to obtain structured parsing results; Step S12: Map and compare the structured indicator data with the corresponding indicator values in the business data platform; perform field mapping and difference calculation between the structured indicator data and the corresponding professional indicator values in the business data platform to form a structured comparison result. Step S13: Generate a difference analysis report.
[0009] In one implementation, in step S4, the candidate page images of the directory are subjected to directory recognition, and a preset recognition condition is determined based on the directory recognition confidence level. When the recognition result does not meet the preset confidence condition, the system expands the directory scanning range and re-scans, performing supplementary scanning on the extended page range of the PDF document until the preset maximum number of pages is reached or the early stopping condition is met. When the recognition result meets the requirements, a directory page image sequence is generated. A reasonable early stopping condition can be one of the following: several consecutive pages are not recognized as directory pages, the number of directory entries has reached a stable threshold, or the directory recognition confidence level continuously reaches a threshold.
[0010] In one embodiment, in step S5, the confirmed candidate page images of the directory are processed for directory recognition, and a multimodal recognition model or equivalent recognition component is called to output the directory recognition result. Subsequently, the system performs post-processing on the recognition result, including item merging, duplicate item deduplication, item sorting, and page number normalization, and finally generates a structured directory entry result.
[0011] In one implementation, in step S6, the physical page range corresponding to each major is located based on the major name and page number information contained in the catalog entries.
[0012] Furthermore, two methods can be adopted: The first is the logical page number mapping method: by analyzing the main text page numbers, footer numbers, or other page number identification information in the document, a unified mapping relationship between logical page numbers and physical page numbers is established; the second is the professional-specific positioning method: when a stable logical page number mapping relationship cannot be established, the system adopts the professional-specific positioning method, which specifically includes: 1) locating the corresponding page based on the professional name and candidate starting page in the directory entry; 2) performing title matching or text matching verification on the page to confirm the professional homepage; 4) combining the starting page information of the next professional or the partial scan results to determine the physical page range of the current professional.
[0013] A system for intelligent parsing and indicator deviation analysis of complex-format PDF documents supporting private domain closed-loop deployment, used to implement the above-described method, is characterized by comprising: The document receiving module is used to receive PDF documents to be parsed and perform format verification. The directory scanning module is used to perform fast scanning and extended scanning of directory candidate pages; The directory recognition module is used to generate directory entry results; The page range positioning module is used to deduce the physical page range corresponding to each major. The image conversion module is used to generate image sequences according to the professional page range; The window splitting module is used to generate sliding window tasks; The desensitization module is used to desensitize sensitive areas in window images; The model routing module is used to select either a local model service or a cloud-based private model service based on the deployment strategy. The structured extraction module is used to extract major names, indicator names, and provincial averages. The indicator mapping module is used to map the extraction results to the corresponding indicator values in the business data platform and to calculate the differences. The report generation module is used to output analysis reports; The audit module is used to record the entire process logs and access logs.
[0014] The present invention has the following beneficial effects: (1) By combining fast scanning of candidate pages with extended scanning, the amount of invalid processing in the directory location stage is reduced; (2) By using directory-driven professional page positioning, only the target page segment is subjected to image processing and model inference, thereby reducing the risk of exposing the entire long document; (3) Improve the contextual continuity of cross-page tables and cross-page descriptions through the sliding window mechanism; (4) Improve security and controllability in sensitive scenarios such as medical quality control and regulatory analysis by using private domain closed-loop deployment and sensitive area desensitization mechanism; (5) By automatically comparing the average data in the notification document with the actual indicator values in the business platform, the efficiency and consistency of report generation are improved. Attached Figure Description
[0015] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a flowchart of steps S3-S5; Figure 3 The flowchart for steps S6-S7 is as follows; Figure 4This is a flowchart of steps S8-S14; Figure 5 This is a structural diagram of a PDF notification document intelligent parsing and indicator difference analysis system that supports closed-loop deployment in private domains. Detailed Implementation
[0016] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0017] like Figure 1 As shown, a method for intelligent parsing and indicator deviation analysis of complex-format PDF documents that supports closed-loop deployment in private domains specifically includes the following steps: Step S1, PDF Report Document Access and Verification: Access the PDF report documents to be processed via file interface, batch import, or automatic access based on a file monitoring mechanism, and verify the file format of the accessed documents. After accessing the files, the system verifies the file format, readability, and integrity. If the detected document format does not meet preset requirements or the document is unreadable, the system rejects the access to the document and records relevant log information. For PDF documents that pass verification, the system generates a unique document ID for each uploaded document and records document metadata, including document name, upload time, source information, and file hash value.
[0018] Step S2, Secure Private Storage Processing of PDF Documents: Securely store and uniformly manage verified PDF documents in the private domain deployment environment. For example, save the original PDF file to the private domain storage system and establish document index information for subsequent document parsing processes.
[0019] Step S3: Table of Contents Candidate Page Scanning: Perform a fast scan of the table of contents candidate pages within a preset initial page range of the PDF document, and image the page numbers within this range to generate table of contents candidate page images. Preferably, the preset initial page range includes the first few pages of the PDF document, such as one or more pages from page 3 to page 5 of the physical pages. The table of contents page images are generated using preset image parameters, such as a resolution of 72 dpi, JPEG quality of 55, and grayscale mode, to improve the efficiency of subsequent table of contents recognition.
[0020] For example: First, perform directory identification by quickly scanning page range 3 / 4 / 5. If it fails, start incremental scanning from page 2. It is subject to the extended scan parameter toc_max_scan_pages (default 12, minimum increased to 5 at runtime) and the effective upper limit min(page_count, max(config_max, quick_end)).
[0021] Step S4: Directory Recognition and Extended Scanning: The candidate page images of the directory are subjected to directory recognition, and the system determines whether the preset recognition conditions are met based on the directory recognition confidence level. If the recognition result does not meet the preset confidence conditions, the system expands the directory scanning range and re-scans, performing supplementary scanning on the extended page range of the PDF document until the preset maximum number of pages is reached or the early stopping condition is met. When the recognition result meets the requirements, a sequence of directory page images is generated. A reasonable early stopping condition can be one of the following: several consecutive pages are not recognized as directory pages, the number of directory entries has reached a stable threshold, or the directory recognition confidence level continuously reaches the threshold.
[0022] Step S5: Structured Directory Entries Generation: The confirmed candidate page images are processed using directory recognition, employing a multimodal recognition model or equivalent component to output the directory recognition results. The system then performs post-processing on the recognition results, including entry merging, duplicate entry removal, entry sorting, and page number normalization, ultimately generating structured directory entries. These entries should at least include the subject name and corresponding page number, and may also provide the directory recognition confidence score, the physical page number, and the original recognized text.
[0023] Step S6: Professional Page Range Location: Determine the physical page range corresponding to each professional category based on the directory entry results. The system locates the physical page range corresponding to each professional category based on the professional name and page number information contained in the directory entry. To improve the accuracy of location, two explicit and hierarchical implementation methods can be adopted: The first logical page number mapping method: By analyzing the main text page numbers, footer numbers, or other page number identification information in the document, a unified mapping relationship between logical page numbers and physical page numbers is established; The second professional-by-professional location method: When a stable logical page number mapping relationship cannot be established, the system adopts a professional-by-professional location method, which specifically includes: 1) Locating the corresponding page based on the professional name and candidate starting page in the directory entry; 2) Performing title matching or text matching verification on the page to confirm the professional homepage; 4) Combining the starting page information of the next professional category or the partial scan results, determining the physical page range of the current professional category.
[0024] Step S7: Professional Page Image Generation: Based on the defined professional page range, only the physical page range corresponding to each professional page is image-processed to generate a professional page image sequence, without performing the same level of image processing on all pages of the entire document. Professional page images can be generated using preset image parameters, such as resolution: 100dpi, JPEG quality: 65, color mode, etc. The generated professional page image sequence is organized according to task identifier and page number order and stored in a private image directory for use in subsequent image segmentation, sensitive area processing, and multimodal model parsing steps.
[0025] Step S8: Sliding Window Image Segmentation: The professional page image sequence is segmented into multiple overlapping image window tasks for model parsing by sliding window segmentation according to a preset window length and step size. Since the system performs sliding window segmentation on the professional page image sequence according to the given window length and step size, and the text naturally and reasonably selects a window length of 2 pages and a step size of 1 page, for a 4-page image sequence of a certain professional subject, three window tasks—1-2, 2-3, and 3-4—can be obtained naturally. This also effectively balances controlling the scale of a single model input with preserving the continuous contextual information of cross-page tables and descriptions.
[0026] S9: Sensitive Area Desensitization: Before the model is loaded into the task window, sensitive areas in the image window are desensitized to reduce the risk of sensitive information leakage. Pre-defined sensitive areas include at least one or more items in the header, footer, organization name, internal number, barcode, QR code, and stamp areas, which are cropped, masked, replaced, or blurred.
[0027] Step S10: Construct a multimodal model input and call the local model or the cloud-based private model; According to a preset deployment strategy, the system routes the anonymized image window data to either a local model service or a cloud-based private model service for inference processing. This model routing mechanism allows for the selection of a suitable model service to execute inference tasks based on system deployment strategies or resource status. Furthermore, the mechanism dynamically selects model services based on computing resources, network conditions, or deployment strategies, thereby improving system processing efficiency.
[0028] Step S11: Extract structured results such as major name, indicator name, and provincial average; The window image data is converted into the model input format, and a multimodal model input is constructed by combining it with preset task prompts. The input is then sent to the multimodal model for inference processing to obtain structured parsing results. The window image byte data is Base64 encoded and combined with preset prompts before being input into the multimodal model to obtain model parsing results. The structured results returned by the model include at least: (1) professional name; (2) data reporting status: including the name of the institution that did not submit data and the corresponding time information; (3) monitoring and evaluation status, including the institution name, the evaluation level of this period (e.g., A, B, C, D) and the evaluation level of each time period within the period; (4) indicator statistical analysis information: including the indicator name, the current period average, median, maximum, minimum, the previous period average, the month-on-month change rate, the historical same period value and the year-on-year change rate; (5) supplementary information: the window number, page range, table title and confidence information. Through the above methods, the system can automatically convert the unstructured information in the image into standardized structured indicator data, providing a foundation for subsequent data processing and analysis.
[0029] Step S12: Map and compare the values with the corresponding indicator values in the business data platform; The structured indicator data is mapped to the corresponding professional indicator values in the business data platform, and differences are calculated to form structured comparison results. The difference calculation results may include one or more of the following: the institute's value, the provincial average, the difference, the deviation rate, whether the standard is met, and the difference level.
[0030] The field mapping can be implemented using at least one of the following methods: category name matching, indicator name matching, standardized field name matching, indicator code matching, and matching with a preset indicator mapping table. After completing the field mapping, the system performs difference calculations on the mapped indicator data and generates indicator comparison analysis results. The analysis results include at least one or more of the following information: the unit's indicator value, the reference average value, the indicator difference, the deviation ratio, the compliance judgment result, and the difference level. Through the above methods, automatic association between image parsing data and business system data can be achieved, and indicator deviation analysis can be completed.
[0031] Step S13: Generate a difference analysis report. Based on the difference calculation results for each indicator, the system performs categorized statistical analysis and comprehensive analysis of the indicator performance, generating an analysis report containing the following information: basic information and statistical results of the indicators, indicator differences and trends, indicator compliance status and level assessment, key indicator anomaly alerts, risk warnings, and improvement suggestions. The analysis report can be automatically generated according to a preset template and output in the form of a structured document or visual charts for users to further view, store, or share. Through the above methods, the system can achieve an automated processing flow from image data parsing and indicator comparison calculation to analysis report generation.
[0032] Step S14: Perform audit logging on the entire process; An analysis report is generated based on the structured comparison results, and audit logs are created for the processes of PDF reception, catalog scanning, image generation, window segmentation, model invocation, result access, and report export. The audit logs must include at least one of the following: upload user ID, document hash value, scan range, window page number range, model service ID, invocation time, result access records, and report export records. These logs are used to record the entire process and access logs.
[0033] like Figure 5 As shown, this invention also provides a PDF notification document intelligent parsing and indicator difference analysis system that supports private domain closed-loop deployment, including: The document receiving module is used to receive PDF documents to be parsed and perform format verification. The directory scanning module is used to perform fast scanning and extended scanning of directory candidate pages; The directory recognition module is used to generate directory entry results; The page range positioning module is used to deduce the physical page range corresponding to each major. The image conversion module is used to generate image sequences according to the professional page range; The window splitting module is used to generate sliding window tasks; The desensitization module is used to desensitize sensitive areas in window images; The model routing module is used to select either a local model service or a cloud-based private model service based on the deployment strategy. The structured extraction module is used to extract major names, indicator names, and provincial averages. The indicator mapping module is used to map the extraction results to the corresponding indicator values in the business data platform and to calculate the differences. The report generation module is used to output analysis reports; The audit module is used to record the entire process logs and access logs.
[0034] (1) Directory-driven local image-based page segment positioning mechanism: Instead of performing full-text OCR or full-text image-based mapping on the entire PDF, this invention first obtains the directory entries through rapid scanning and extended scanning of the directory candidate pages, and then reverses the professional physical page range based on this, so that subsequent image-based mapping and model processing are only performed on the target page segments; (2) Sliding window-based cross-page context modeling mechanism: This invention adopts a sliding window segmentation method with fixed window length and overlap step for cross-page tables and cross-page analysis content, so that the content related to the same indicator can remain continuous in adjacent windows, reducing the impact of cross-page information breakage on the analysis results; (3) Data Minimum Exposure Inference Mechanism under Private Domain Closed Loop: This invention incorporates the original PDF, images, intermediate results and report results into the private domain controlled environment for processing, and only sends the minimum window image required to complete the current task into the model. At the same time, it performs desensitization processing before inputting the data into the model to reduce the risk of sensitive information exposure. (4) Automatic mapping and comparison mechanism between the average value of the notification document indicators and the indicator values of the business platform: This invention not only completes PDF parsing, but also further performs standardized field mapping and difference calculation between the extracted provincial average value and the corresponding indicator value in the business data platform to form a structured output for business reports.
[0035] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A method for intelligent parsing and indicator deviation analysis of complex-format PDF documents supporting private domain closed-loop deployment, characterized in that: Specifically, the following steps are included: Step S1, PDF report document access and verification: Access the PDF report document to be processed through file interface, batch import method or automatic access method based on file monitoring mechanism. After accessing the file, verify the file format, file readability and integrity. Step S2, Secure storage of PDF documents in private domain: Securely store and uniformly manage verified PDF documents in the private domain deployment environment; Step S3: Scan the candidate pages of the directory and generate images of the candidate pages of the directory; Step S4: Directory identification and extended scanning; Step S5: Generate structured directory entries, generating the structured directory entry results; Step S6: Professional Page Range Determination: Determine the physical page range corresponding to each professional field based on the directory entry results; Step S7: Professional Page Image Generation: Based on the determined professional page range, perform page image processing only on the physical page range corresponding to each professional field to generate a professional page image sequence; Step S8: Sliding window image segmentation: The professional page image sequence is segmented into multiple overlapping image windows for model parsing by sliding window segmentation according to the preset window length and step size; Step S9: Construct a multimodal model input and call the local model or cloud-based private model; according to the preset deployment strategy, route the desensitized image window data to the local model service or cloud-based private model service for inference processing; Step S10: Extract the structured results; The window image data is converted into the model input format, and combined with preset task prompts to construct a multimodal model input, which is then sent to the multimodal model for inference processing to obtain structured parsing results. Step S11: Map and compare the data with the corresponding indicator values in the business data platform; The structured indicator data is mapped to the corresponding professional indicator values in the business data platform, and the differences are calculated to form a structured comparison result. Step S12: Generate a difference analysis report.
2. The method for intelligent parsing and indicator deviation analysis of complex-format PDF documents supporting private domain closed-loop deployment as described in claim 1, characterized in that: In step S4, the candidate page images of the directory are subjected to directory recognition, and it is determined whether the preset recognition conditions are met based on the directory recognition confidence level. When the recognition result does not meet the preset confidence conditions, the system expands the directory scanning range and re-scans, performing supplementary scanning on the extended page range of the PDF document until the preset maximum number of pages to be scanned is reached or the early stopping condition is met. When the recognition result meets the requirements, a directory page image sequence is generated. A reasonable early stopping condition can be one of the following: several consecutive pages are not recognized as directory pages, the number of directory entries has reached a stable threshold, or the directory recognition confidence level has continuously reached the threshold.
3. The method for intelligent parsing and indicator deviation analysis of complex-format PDF documents supporting private domain closed-loop deployment as described in claim 1, characterized in that: In step S5, the confirmed candidate page images of the directory are processed for directory recognition, and a multimodal recognition model or equivalent recognition component is called to output the directory recognition result. The system then performs post-processing on the recognition results, including merging entries, removing duplicate entries, sorting entries, and normalizing page numbers, ultimately generating structured directory entries.
4. The method for intelligent parsing and indicator deviation analysis of complex-format PDF documents supporting private domain closed-loop deployment as described in claim 1, characterized in that: In step S6, the physical page range corresponding to each major is located based on the major name and page number information contained in the catalog entries.
5. The method for intelligent parsing and indicator deviation analysis of complex-format PDF documents supporting private domain closed-loop deployment as described in claim 4, characterized in that: Two methods can be used: The first is logical page number mapping: by analyzing the main text page numbers, footer numbers, or other page number identification information in the document, a unified mapping relationship between logical page numbers and physical page numbers is established; the second is discipline-by-discipline positioning: when a stable logical page number mapping relationship cannot be established, the system adopts discipline-by-discipline positioning, which specifically includes: 1) locating the corresponding page based on the discipline name and candidate starting page in the directory entry; 2) performing title matching or text matching verification on the page to confirm the discipline's homepage; 4) combining the starting page information of the next discipline or the results of partial scanning to determine the physical page range of the current discipline.
6. The method for intelligent parsing and indicator deviation analysis of complex-format PDF documents supporting private domain closed-loop deployment as described in claim 1, characterized in that: After step S8, sensitive areas are desensitized: before the model is sent into the task window, sensitive areas in the image window are desensitized.
7. A system for intelligent parsing and indicator deviation analysis of complex-format PDF documents supporting private domain closed-loop deployment, used to implement the method described in any one of claims 1-6, characterized in that: include: The document receiving module is used to receive PDF documents to be parsed and perform format verification. The directory scanning module is used to perform fast scanning and extended scanning of directory candidate pages; The directory recognition module is used to generate directory entry results; The page range positioning module is used to deduce the physical page range corresponding to each major. The image conversion module is used to generate image sequences according to a professional page range; The window splitting module is used to generate sliding window tasks; The desensitization module is used to desensitize sensitive areas in window images; The model routing module is used to select either a local model service or a cloud-based private model service based on the deployment strategy. The structured extraction module is used to extract major names, indicator names, and provincial averages. The indicator mapping module is used to map the extraction results to the corresponding indicator values in the business data platform and to calculate the differences. The report generation module is used to output analysis reports.
8. The system as described in claim 7, characterized in that: Also includes: The audit module is used to record the entire process logs and access logs.